Complete Memory & Storage Tracker

The Memory
Supercycle

CONNECTING…
As of · Aug 17, 2026
Cycle · SHORTAGE
Anchored — reported/public, cited in each footnote Modeled — built from anchors (interpolated / triangulated) Projection — hatched / shaded, scenario not forecast Band — low/high scenario range ○ Archived — dated capture of a source page (§03·F) ● Live — refreshes in Claude; else dated snapshot ⚑ Conflict — sources disagree; flagged, not merged Every section's footnote states exactly what's anchored vs modeled. Not investment advice.

Artificial intelligence has turned the entire memory and storage stack — HBM, DRAM, NAND and hard disk — into the scarcest resource in computing. Every wafer routed to AI memory, and every platter routed to a hyperscaler, is denied to the device in your hands.

A standing dashboard of the whole memory landscape, from the fastest HBM cube to the coldest nearline drive. Pricing, market share, the company roster and the signal feed pull current figures live through Claude with web search; everything renders from a compiled baseline first.

$0B
Projected total memory market, 2026
TrendForce · DRAM revenue +144% YoY
0%
Peak DRAM contract price jump, Q1 2026 (QoQ)
TrendForce · record quarterly move
0%
Share of high-end memory consumed by data centers
IDC / Avnet · 2026 estimate
0
Year HBM and nearline hard drives are sold out through
Micron, SK Hynix, Samsung, WD, Seagate
Signals now · tap any chip to jump to its section
01The Memory Stack◷ verified Jul 2 2026
01

The Memory Stack

● live prices

Four tiers, fastest and priciest at top to cheapest-per-terabyte at bottom. AI demand is straining every one of them at once — the first time the whole stack has gone tight together.

↑ Faster · pricier · per-bitCheaper · denser · per-terabyte ↓
HBM
High-Bandwidth Memory
Stacked DRAM cubes (12–16 dies) wired with through-silicon vias.
AI role: the bandwidth that feeds every GPU. The bottleneck of the buildout.
SK Hynix · Samsung · Micron  (CXMT emerging)
~$200–500
per stack · by gen
Sold out '26
DRAM
DDR5 · LPDDR5X · Server
Main working memory — and the raw dies HBM is built from.
AI role: server DDR5 modules and the base for every HBM stack.
Samsung · SK Hynix · Micron  · CXMT · Nanya
+58–63%
contract QoQ · latest
Shortage
NAND / SSD
Flash · Enterprise & Client
Solid-state flash; 200+ layer 3D NAND in enterprise SSDs.
AI role: fast storage that feeds training/inference data to the GPUs.
Samsung · SK Hynix/Solidigm · Kioxia · Micron · SanDisk · YMTC
+70–75%
contract QoQ · latest
Tight · rising
HDD
Hard Disk · Nearline
Spinning platters; HAMR/MAMR drives at 30–44TB and climbing.
AI role: lowest cost-per-TB home for the flood of AI-generated cold data.
Seagate · Western Digital · Toshiba  (>95% of shipments)
~+50%
consumer · 5 months
Sold out '26
01·BMemory 101 — A Field Guide◷ structural
01·B

Memory 101 — A Field Guide

Every kind of memory and storage, what it physically is, how it works, and where it's used. The whole industry splits on one line: volatile memory that needs constant power to remember, and non-volatile storage that keeps data when the power is off. Speed and cost trade off against capacity all the way down.

Volatile Working memory

Holds data only while powered. It's fast and sits close to the processor, acting as the workspace a chip computes in. Cut the power and it forgets instantly. This is "memory" in the strict sense — SRAM and DRAM, including HBM.

Examples: CPU/GPU cache (SRAM) · system RAM (DDR5) · phone RAM (LPDDR5X) · GPU memory (HBM, GDDR)

Non-volatile Storage

Keeps data with the power off. It's slower but far cheaper per gigabyte and vastly higher capacity, so it stores everything persistently — the operating system, apps, files, models and archives. This is "storage."

Examples: SSD (NAND flash) · hard drives (HDD) · tape archives · firmware (NOR flash)
The Memory Hierarchy
Fast, costly and tiny at the top; slow, cheap and enormous at the bottom. Every device blends these tiers — the art is keeping hot data high and cold data low.
SRAM cache~1 ns
Registers · L1/L2/L3
On-die CPU/GPU cache, the fastest tier
HBM~80–150 ns · TB/s
HBM3E / HBM4
AI accelerator & HPC memory
DRAM~10–90 ns
DDR5 · LPDDR5X
System main memory — PCs, phones, servers
NAND / SSD~20–100 µs
Flash storage
Fast storage: OS, apps, datasets, eSSD
HDD~5–10 ms
Hard disk
Bulk & nearline cloud / cold storage
Tapeseconds+
LTO tape
Deep archive, backup, air-gapped data
↑ Faster · costlier per bit · lower capacitycheaper per bit · higher capacity ↓

Volatile — Working Memory

needs power · the chip's workspace
SRAM Static RAM
Volatile
Stores each bit in a six-transistor latch that holds its state as long as power is on — no refresh needed, so it's the fastest memory there is. The catch: big cells mean low density and high cost, so it's used sparingly and never sold as a commodity chip.
All usesCPU/GPU cache (L1/L2/L3), register files, on-chip buffers, network-router packet buffers, small embedded SRAM in microcontrollers and ASICs
Speed / role~1 ns · the closest, fastest tier to the compute cores
Built on-die by every chipmaker · not a standalone market
DRAM Dynamic RAM
Volatile
One transistor and one capacitor per bit. The capacitor leaks, so every cell must be refreshed thousands of times a second — "dynamic." That simplicity makes it dense and cheap, which is why it's the main memory in essentially every computing device. It's also the raw die that HBM is built from.
All usesPC/laptop/server main memory, smartphone RAM, GPU memory, game consoles, automotive, networking — and the base dies stacked into HBM
Speed / role~10–90 ns · the working memory the processor reads and writes constantly
Samsung · SK Hynix · Micron · CXMT · Nanya · Winbond

DRAM comes in packaging variants — same cells, different bus & form factor

DDR DDR4 · DDR5
The mainstream standard. Pluggable DIMMs on a 64-bit bus; server RDIMMs feed CPUs in huge capacities. DDR5 is current.
Use: desktops, laptops, servers
LPDDR LPDDR5X
Low-power, soldered-down DRAM optimized for battery life and density. Now also the system memory on AI CPUs (Grace/Vera).
Use: phones, thin laptops, autos, AI superchips
GDDR GDDR6 · GDDR7
Graphics DRAM on a very wide, very fast bus, soldered around a GPU. High bandwidth at lower cost/complexity than HBM.
Use: gaming & pro GPUs, consoles
HBM HBM3E · HBM4
DRAM dies stacked 8–16 high and wired with through-silicon vias for enormous parallel bandwidth, sitting beside the GPU on one package.
Use: AI accelerators, HPC, high-end networking

Non-Volatile — Storage

keeps data with the power off · cheap per TB
NAND Flash SSD · UFS · eMMC
Non-volatileAI
Traps electric charge in a cell to hold a bit without power. Cells are organized into pages and blocks (written in pages, erased in blocks), and modern NAND stacks 200–400+ layers vertically (3D NAND). Storing more bits per cell — SLC → MLC → TLC → QLC — trades endurance and speed for capacity and cost.
All usesSSDs (consumer & enterprise eSSD), phone/tablet storage (UFS), memory cards, USB drives, embedded storage, AI training-data & checkpoint storage
Speed / role~20–100 µs · fast persistent storage, the hot tier of the storage layer
Samsung · SK Hynix/Solidigm · Kioxia · Micron · SanDisk · YMTC
NOR Flash code storage
Non-volatile
A different flash arrangement that is byte-addressable and supports "execute-in-place" — a processor can run code directly from it. Far lower density than NAND, but fast, reliable random reads make it ideal for storing firmware that must boot instantly.
All usesFirmware / BIOS / boot code, microcontrollers, automotive ECUs, industrial & IoT devices, anything that executes code in place
Speed / roleFast reads, slow writes · small-capacity code & config store
Macronix · Winbond · Infineon · Microchip · GigaDevice
HDD Hard Disk Drive
Non-volatileAI
Magnetic platters spinning at 5,400–7,200 rpm with read/write heads floating nanometers above the surface. New recording methods (HAMR, MAMR) keep pushing areal density, now reaching 30–44 TB per drive. Mechanical, so slower — but unbeatable on cost per terabyte.
All usesNearline / cloud cold storage, data-center bulk capacity, backups & archives, surveillance, NAS — the home for AI's flood of generated data
Speed / role~5–10 ms seek · cheapest $/TB · the cold tier
Seagate · Western Digital · Toshiba (the entire industry)
Magnetic Tape LTO
Non-volatile
Data written sequentially to magnetic tape in removable cartridges. Offline and sequential-access, so it's slow to reach any one file — but the lowest cost per terabyte of all, with decades of shelf life and no power draw at rest.
All usesDeep archive, long-term backup, regulatory retention, "cold" cloud tiers, air-gapped / ransomware-resilient copies
Speed / roleSeconds to load · the coldest, cheapest, most durable tier
IBM · Fujifilm · Sony · HPE (LTO consortium)

Emerging & Specialized

new tiers, many built for AI
HBF High Bandwidth Flash
Non-volatileAI
NAND flash re-architected into an HBM-style stacked cube with massively parallel sub-arrays, sitting in the same footprint as HBM beside the GPU. Targets HBM-class bandwidth (~1.6 TB/s) with 8–16× the capacity, and it's non-volatile — but flash's limited write endurance aims it at read-heavy inference, not training.
All usesAI inference — keeping a full model's weights on-package, HBM capacity extension, edge AI where models are static
StatusSamples H2 2026, first products ~2027 · standardizing via the Open Compute Project
SanDisk + SK Hynix (OCP standard) · Samsung
MRAM Magnetoresistive RAM
Non-volatile
Stores bits as the magnetic orientation of a tiny tunnel junction (STT-MRAM). It's fast, non-volatile and effectively unlimited in write endurance — a rare combination — but density and cost keep it in specialized roles rather than mass storage.
All usesEmbedded non-volatile memory replacing NOR/eFlash in microcontrollers, last-level caches, automotive, aerospace & industrial, persistent buffers
Speed / roleSRAM-like speed with persistence · niche but growing in embedded
Everspin · Samsung · TSMC & GlobalFoundries (embedded)
ReRAM & PCM resistive / phase-change
Non-volatile
Store bits as a change in electrical resistance — by forming conductive filaments (ReRAM) or switching a material between crystalline and amorphous phases (PCM). Both promise a "storage-class memory" tier between DRAM and NAND. Intel/Micron's 3D XPoint (Optane) was the most famous PCM product — discontinued in 2022, but the idea persists.
All usesStorage-class memory, embedded NVM, IoT, and in-memory / neuromorphic compute where resistance states do analog math for AI
StatusMostly research & niche today · watched as a future analog-AI substrate
Weebit Nano · TSMC · Infineon · (legacy: Intel/Micron Optane)

How This Maps to Your Brain

a useful analogy — and where it breaks down

Why the comparison is tempting

The memory hierarchy above — tiny-and-instant at the top, vast-and-slow at the bottom — rhymes with how neuroscientists describe human memory: a fast, fragile "working" store for what you're thinking about right now, and slower, durable stores for what you've learned. The mapping below is a teaching aid, not a literal equivalence. It's genuinely useful for intuition, and genuinely misleading if taken too far — so the honest caveats follow right after.

SRAM cacheregisters · L1/L2/L3 · ~1 ns
The focus of attention attention
The handful of things held in conscious focus this instant — a phone number you're about to dial. Vanishingly small, effectively instant, and gone the moment attention shifts. Like cache: the closest, fastest, tiniest tier, sitting right next to the "compute."
DRAMworking memory · volatile · ~10–90 ns
Working memory prefrontal cortex
The mental "scratchpad" that holds the task you're actively working on — the sentence you're reading, the plan you're juggling. It's volatile: stop paying attention (lose "power") and it's wiped. The catch in the analogy: yours holds roughly 4–7 items, not gigabytes.
HBMbandwidth to the compute · TB/s
White-matter bandwidth connectivity
HBM exists to feed the processor fast enough — its value is bandwidth, not capacity. The brain's analog is the massive parallel wiring between regions: billions of axons moving signals at once. When that connectivity is the bottleneck, more raw storage doesn't help — exactly HBM's design premise.
NAND / SSDfast non-volatile · ~µs
Consolidated long-term memory hippocampus → cortex
Things you've learned well enough to recall reliably — your address, how to ride a bike. Non-volatile (survives sleep, survives "power off"), but slower to retrieve than working memory. Memory consolidation during sleep is loosely the brain "writing" working memory down to durable storage.
HDD / nearlinecheap bulk · ~ms
Remote & episodic memory distributed cortex
The vast store of things you can recall but rarely do — a childhood birthday, a face from years ago. High capacity, cheap to keep, slow to reach. Retrieval takes a noticeable beat, the way a spinning disk seeks before it reads.
Tape / archivecold · offline · seconds+
Deeply buried memories cue-dependent
Memories you can't reach on demand but that surface with the right cue — a smell that drops you into a specific moment decades ago. Effectively "offline" until something mounts the archive and loads it back.

⚠ Where the analogy breaks down — important

  • The brain has no clean tiers. Computers separate storage (the chips) from processing (the CPU). The brain does both in the same place — the synapses that store a memory are the same ones that compute with it. There is no "memory chip" distinct from the "processor."
  • Memory isn't a file you read back. Storage returns exactly what was written. Human recall is reconstructive — you rebuild a memory each time, and the act of recalling it can change it. A hard drive doesn't rewrite a file just because you opened it.
  • Capacity works nothing alike. A drive fills up bit by bit toward a fixed limit. The brain stores in distributed patterns across billions of synapses with no known fixed "GB" ceiling, and forgetting is an active, useful process — not "running out of space."
  • "Volatile vs non-volatile" is fuzzy in biology. Working memory fades in seconds, long-term memory lasts decades — but the line between them is gradual, chemical, and still debated, not a clean power-on/power-off switch.
  • So why bother with the analogy? Because the pressures are real and shared: both systems trade speed against capacity against cost/energy, both need fast local stores near the compute, and both are bottlenecked by bandwidth as much as raw storage. That's the genuine insight — and it's exactly why AI accelerators pile on HBM rather than just more DRAM.

Latencies and layer counts are representative orders of magnitude — exact figures vary by product, generation and process. "AI" tags mark tiers being reshaped directly by AI demand. The boundary between memory and storage is blurring: HBF, storage-class memory and CXL pooling all aim at the gap between fast-but-small DRAM and cheap-but-slow flash. The brain mapping is an educational analogy; neuroscience is still actively researching how human memory is encoded, consolidated and retrieved, and no current model treats the brain as a literal storage hierarchy.

02Why This Cycle Is Different◷ structural
02

Why This Cycle Is Different

Memory has always been brutally cyclical. The 2026 shortage breaks the pattern because it is demand-driven, not supply-driven — and it spans the entire stack at once. The 2021 crunch came from shuttered fabs; it healed when factories reopened.

This time the constraint is AI itself. Training and inference are bandwidth-bound, so GPU racks swallow HBM; the data they train on needs enterprise SSDs; the exhaust — checkpoints, logs, generated media — lands on nearline HDD. Makers pour capacity into the high-margin top of the stack, starving everything below.

Relief in volume isn't expected until late 2027–2028. Until then hyperscalers pre-pay to lock allocation. Hard-drive makers have sold out 2026 with agreements stretching to 2028; the three memory makers have added roughly $900 billion in combined market value since September 2025. Crucially, the makers are choosing not to flood the market: Samsung and SK hynix can fulfill only ~70% of DRAM orders yet are deliberately restraining capex to "minimize the risk of oversupply" — the discipline that could make this cycle outlast prior boom-busts (or, cynically, exactly what every supplier says before a glut).

The 3-to-1 Wafer Tax

1 bit of HBM shipped…
↓ costs ↓
3 bits of conventional DRAM not made

HBM's stacked construction is far more wafer-intensive than ordinary memory — one bit of HBM forgoes ~three bits of standard DRAM. HBM alone now absorbs an estimated 23% of all DRAM wafers, draining the commodity pool that feeds phones, PCs and SSDs.

02·BCycle Position & Risks◷ Jul 17 2026 · capex refresh
02·B

Cycle Position & Risks

The investor's section. Memory is the most violently cyclical corner of tech — so the questions that matter aren't "is AI big?" but where are we in the cycle, what sustains it, and what breaks the thesis? This pulls together the leading indicators: the bit supply-demand gap, capex (the supply pipeline), valuation in cycle context, and an explicit bear case with the signals that would flip each risk.

Thesis Scorecard — the signals to watch
We're in the boom phase — historically the most dangerous time to extrapolate

Memory cycles run trough → recovery → boom → peak → correction. Every prior cycle has reverted: in the 2017–18 DRAM supercycle, supply tightened, margins crossed 50%, analysts declared the cycle "structurally different" — then supply caught up and margins fell toward zero within two years.

This cycle has a more credible structural case (AI demand + the HBM wafer tax constraining supply). But the mechanism that ends memory cycles hasn't changed: high margins attract capacity, capacity arrives in 12–18 months, and the gap closes.

Consensus read: DRAM pricing in its "middle phase" · earnings peak projected Q4'26–Q2'27 · first new fabs arrive 2027. The bull/bear debate is entirely about how long the boom lasts, not whether it's real. Update (Micron FQ3, Jun 24 2026): record $41.5B revenue and a record $50B FQ4 guide confirm the boom is still accelerating, with HBM booked through 2027 into 2028 and new fabs yielding no meaningful output until fiscal 2028 (bull-supportive) — but management also guided an explicit "meaningful moderation in the rate of price increases" after FQ4 (the first crack the bear case has been waiting for). Update (TrendForce 3Q26 contract prices, Jul 3): that moderation is now in the data — conventional DRAM +13–18% QoQ (from ~+low-90s% in 2Q) and NAND +10–15% (from +70–75%), still rising but decelerating hard as consumer demand hits an affordability ceiling. Prices remain at record highs; the pace is what turned.
The Bit Gap — memory's single most important cyclical indicator

When annual bit-demand growth runs above bit-supply growth, prices rise and the cycle is bullish. When supply growth crosses above demand, the cycle turns. Right now demand leads on both DRAM and NAND — but the gap is the thing to watch, not the price.

DRAM bit growth

demand vs. supply · % YoY
Bit demand Bit supply
Demand has led supply since 2024; HBM's 3:1 wafer tax (1 bit HBM = 3 bits DDR5 foregone) keeps supply structurally capped.

NAND bit growth

demand vs. supply · % YoY
Bit demand Bit supply
NAND's gap is tighter than DRAM's; enterprise SSD demand is the swing factor. Easier to add supply than DRAM, so a faster turn risk.
The Supply Pipeline — combined DRAM capex

The bull case rested on supply being unable to respond — and that is now changing on the spend side, but not yet on the bit side. The response has begun: Micron raised FY2026 capex ~81% to ~$25B (having exited consumer memory to chase AI); SK hynix is spending well above its 2025 ₩27.5T (~$20B) and raised a record $26.5B ADR (Jul 10) earmarked for its Yongin and Cheongju fabs; Samsung restarted mothballed Pyeongtaek lines (Plant 4 Phase 2, converted foundry→memory; Line 5). But new cleanrooms take 2–3 years — SK hynix's Yongin Fab 1 targets 2027, Micron's Idaho 2027, Samsung's P5 ~2028 — so the extra bits land 2027–28, not 2026. The spend is up; the supply isn't here yet. That is exactly kill-switch #2 (§02·D): capex has stopped being the bear's missing piece, yet the 2026 shortage persists because you cannot ship DRAM from a fab that is still being equipped.

Combined memory capital expenditure, $B/year, stacked by maker (DRAM-led; the makers don't cleanly split DRAM-only capex, so this tracks total memory capex — the basis the series always used). 2026–27 are forecasts. Micron is the hard anchor (FY26 ≈$25B, +81% YoY, cited); SK hynix & Samsung are guidance-based estimates (“considerably higher” / restarted lines); 2027 is a projection. Cleanroom lead times are lengthening — so this spend mostly lifts 2027–28 supply, not 2026. Sources: Micron FQ guidance, SK hynix disclosures & F-1, Samsung, Seoul Economic Daily (Dec 2025), Omdia.
SK Hynix Samsung Micron

Market revenue (TAM)

$B/year · price × bits · where the dollars actually are
HBM DRAM (ex-HBM) NAND
HBM TAM: ~$35B (2025) → ~$100B (2028), pulled forward two years. By 2028 HBM alone ≈ the entire 2024 DRAM market.

Valuation in cycle context

Price / book · the metric value investors use for memory (earnings are too cyclical for P/E)
cycle rangepeak
Forward P/E looks "cheap" (single digits) precisely because earnings are at a cyclical peak — the classic memory trap. P/B shows each name vs. its own historical range; near the top of the band = priced for the boom to continue.
Catalyst Calendar — what moves the thesis next
DRAM Supply vs. Demand Balance — Quarterly · 2022Q1 → 2035Q4E

Modeled quarterly DRAM bit balance in billion gigabytes (GB): positive bars = supply surplus, negative = deficit. Real high-confidence anchors (TrendForce, Micron earnings) define the shape through 2026; everything beyond is an explicit modeled scenario, not a forecast — interest rates and AI demand 10 years out aren't really knowable. The whole point of plotting it is to make the assumptions inspectable.

Assumptions checked against the newest prints (Jul 6 2026): the deficit shape through 2026 still matches the data — but the rate story has a fresh anchor: TrendForce’s 3Q26 call has conventional DRAM contract prices decelerating to +13–18% QoQ (from ~+90s%) as consumer buyers hit an affordability ceiling (§03, kill-switch #1) — consistent with a deficit that is peaking, not widening. On the supply side, kill-switch #2 is ARMED for CY2028 (§02·D): 2026–27 capacity announcements land as the bear case’s 2028 flood if they all ship. Two structural offsets the model carries via the HBM wafer tax: Micron’s take-or-pay SCAs now lock ~20% of DRAM and ~30% of NAND through 2030 (§06·F·M), and HBM3E+HBM4 are booked through 2027 into 2028 (§10·A feed, Jun 24). Modeled bars beyond 2026 were not rewritten on one quarter’s print — the scenario spread (bull/base/bear) is the honest container for that uncertainty; watch the Jul 23 Samsung/SK capex commentary (§02·E) for the next assumption test.

Scenario
Surplus (anchor) Surplus (modeled) Deficit (anchor) Deficit (modeled) Quarter pinned to a hard public anchor

Anchors (high-confidence quarters): 2022Q1–Q4 surplus from TrendForce's "DRAM 2022 oversupply, bit supply +18.6% vs demand +17.1%"; 2023 trough from the documented ~30% memory revenue collapse; 2024 recovery from TrendForce; 2025 transition to deficit from rising contract prices; 2026 record deficit from Micron's December 2025 disclosure that "supply will remain substantially short of demand" and the industry can meet only half to two-thirds of demand. Modeled beyond 2026Q4 using: bit-demand CAGR (base 18%, bull 22%, bear 14%); supply growth lagged ~6 quarters behind capex (Micron Idaho online 2027; second fab 2028); and historical cycle behavior (memory always reverts — bear case = 2017-18 precedent applied). Quarters are interpolated within annual anchors using observed seasonality. This is a model, not a prediction. The further right you look, the wider the error band.

The Bear Case — risks & their trigger signals

Every thesis needs its disconfirming evidence. These are the real risks, ranked by severity, each paired with the specific signal that would tell you it's materializing.

Supply Discipline — the mechanism behind "this time is different"

The crux of the bull thesis isn't just demand — it's that the three makers are deliberately not flooding the market. They're fulfilling only ~70% of orders yet restraining capex to protect pricing, and the new fabs that would break the shortage don't come online until 2027–28. This is the structural reason prices may stay high longer than a normal cycle — and, read cynically, also the setup for an eventual glut. The same fact supports both the bull and bear case.

~70%
DRAM orders filled
Samsung & SK hynix can only meet ~70% of incoming DRAM demand (record-low fulfillment).
35% / 23%
Demand vs supply growth '26
TrendForce: bit demand +35% in 2026 against just +23% supply — the gap that sustains pricing.
~30%
Of sales to capex
SK hynix to invest ~30% of revenue in fabs in 2026 — aggressive, but "still won't resolve the shortage."
₩37.6T
SK hynix Q1'26 op profit
Record quarter (₩52.6T rev); Samsung chip op profit ₩53.7T, ~94% of group profit. No incentive to spoil it.
When relief actually arrives — new fab capacity online
2H 2026→27
SK hynix M15X
Cheongju; output begins 2H 2026, ramps to ~80K wpm through 2027. Yongin cluster behind it.
~2028
Samsung P5
Pyeongtaek; cleanroom secured "in advance" — but deliberately measured, not a flood.
H2 2028
Micron Japan
~$10B Hiroshima DRAM fab; won't ship until late 2028. Virginia (US) DRAM already starting.

Bit-growth, capex and TAM figures: TrendForce, Micron earnings, IDC, Counterpoint (2025–26 actuals; 2027–28 forecast). Supply-discipline figures: Samsung & SK hynix Q1 2026 earnings calls (~70% DRAM fulfillment, ~30%-of-sales capex, record profits), TrendForce (demand +35% vs supply +23% in 2026), DCD / Tom's Hardware (fab timing: SK M15X output 2H 2026→ramp 2027, Samsung P5 ~2028, Micron Japan H2-2028). The 2017–18 precedent: SK Hynix/Micron margins fell from 50%+ toward zero within ~2 years as supply caught up. Micron's own 10-K notes DRAM ASPs have swung between +40% and −40% year-over-year in the past five years. P/B ranges are approximate cycle bands. Cycle-phase placement is a judgment call, not a precise measurement. Everything here is analysis, not investment advice — and the bear case is as evidence-based as the bull case deliberately.

02·CWhat Inning Are We In?◷ Jun 24 2026
02·C

What Inning Are We In?

A 9-inning baseball read on the memory supercycle — an intuitive translation of the cycle-position analysis above. The 9 innings run from the 2023 trough (the bottom of the order) to the eventual peak and turn. Reviewing every signal in this dashboard — the bit-gap, prices, inventory, capex discipline, and the projected supply wave — here's the call.

6th
of 9 innings

We're in the middle of the 6th — full boom, but the late innings haven't started

Prices have already spiked hard (DRAM up ~9× off the 2023 trough), so the early innings are clearly behind us. But the shortage is still structural: demand outruns supply, inventories are at record lows, capex is still disciplined, and no meaningful new fab capacity arrives until 2027. The peak (earnings top, projected Q4'26–Q2'27) is the 7th-inning stretch — still 2–4 quarters away — and the supply wave that ends the game is an 8th/9th-inning event in 2027–28.

Reasonable analysts could place us anywhere from the 5th (weight the structural shortage) to the 7th (weight how far prices have already run). Central call: 6th.
● This inning (6th)

Full boom

2026 now: demand still > supply, inventories record-low, capex still restrained so supply can't yet respond. Prices climbing but the rate of ascent is what to watch.

⚾ Extra innings (the bull case)

If AI demand keeps compounding and HBM's 3:1 wafer tax keeps commodity DRAM structurally starved, this upcycle could run longer than a normal memory cycle. The new supply-discipline evidence cuts this way: Samsung/SK hynix are restraining expansion (only ~70% order fulfillment), meaningful new volume doesn't land until 2027–28 (SK M15X ramps 2027, Samsung P5 ~2028, Micron Japan H2-2028), and research firms now say "no scenario where prices correct in 2H 2027." The game stretches — the peak slips toward 2028.

⚑ Called early (the bear case)

A demand air-pocket — an AI capex pause, or a double-ordering unwind — could end the game abruptly, jumping straight to the 9th. Memory's history is full of games called on account of rain: in 2017–18, margins went from 50%+ to near zero in two years once the mood turned. The disciplined-supply story is itself the risk: every maker is also quietly adding capacity (Samsung P5, SK M15X/Yongin, Micron Japan/Virginia), and "minimize oversupply" is precisely what suppliers say right before a glut. Wildcard: helium supply (critical to fabs) is exposed to the Middle East conflict.

Jul 6 2026 re-read — the clock moved. Since this section’s last calibration, four late-inning markers printed: (1) the pricing rate peaked — 3Q26 contract guide decelerates to +13–18% DRAM from ~+90s% (§03, kill-switch #1 TRIPPED-early); (2) demand destruction went from theory to mechanism — Apple raised Mac prices $100–500 and confirmed iPhone hikes on memory costs (§08·C, KS#5’s canary singing); (3) sentiment hit classic late-cycle furniture — Street-high targets doubled to $2,200 in days, a marquee short (Burry) arrived, and Korea’s retail 2×-leveraged complex drew a regulator warning (§04·D, §04·D·S, §04·F); (4) Korea’s export price/kg printed its first MoM decline in nine months (§03·B). Against that, the innings-remaining evidence also firmed: HBM booked through 2027 into 2028, take-or-pay SCAs locking ~20% DRAM / ~30% NAND to 2030, and a modeled deficit that peaks rather than closes before 2027–28 (§01, §06·F·M). Net read: the price-cycle inning advanced one notch — the rate of change has peaked — while the volume/infrastructure game is still mid-innings. Those are two different scoreboards, and conflating them is how both bulls and bears get this cycle wrong. Next umpire calls: Jul 8 Samsung prelim, Jul 23 Korean earnings, late-Sep Micron FQ4 (§02·E).

This is an analogy, not a measurement — memory cycles don't have fixed innings, and the inning number is a judgment that translates the cycle-position analysis (§02·B) into an intuitive frame. It rests on the same signals shown throughout this dashboard: the bit-gap (demand still leads), prices (already spiked), inventory (record-low), capex (disciplined, supply can't respond until 2027), and the modeled supply/demand balance (crosses over ~2027). The honest range is the 5th–7th. The game can go to extra innings or be called early — and the further into the late innings you think we are, the more the risk register (§02·B) matters. Not investment advice.

02·DThesis Kill-Switches◷ Jul 2 2026
02·D

Thesis Kill-Switches

The bear case, concentrated: six falsification criteria that would break or bend the supercycle thesis, each with a status as of Jul 2 2026 and the specific evidence that flips it. Statuses are dated judgments, not predictions — the whole point is that they're checkable. One is already tripped.

1 · Pricing-power inflection

TRIPPED · EARLY
Micron itself guided a "meaningful moderation in the rate of price increases" beyond FQ4 (Jun 24), and its SCA price ceilings are set at CQ2-2026 levels — a contractual cap on further upside for ~20–30% of output. The second derivative has clearly turned: QoQ DRAM contract gains decelerate ~+90–95% (1Q) → ~+58–63% (2Q) → +13–18% (3Q); NAND ~+70–75% (2Q) → +10–15% (3Q). Corroboration stacking up: (a) Korea’s DRAM export price/kg posted its first MoM decline in nine months in Jun (−~$1K to $60K, §03·B); (b) the clincher — TrendForce’s 3Q26 contract-price call (Jul 3) is a hard, quantified deceleration: conventional DRAM +13–18% QoQ and NAND +10–15%, down sharply from 2Q26’s ~+58–63% DRAM / +70–75% NAND — itself down from 1Q26’s record ~+90–95% DRAM. Still rising — but the rate has broken, and TrendForce names the cause outright: consumer PCs and smartphones have hit their affordability limit. This is the moderation Micron guided, now visible in third-party contract data.
Full trip: first sequential ASP decline in any major segment, or FQ1-27 guidance below street.

2 · Supply response lands

ARMED · CY2028
Greenfield bits are dated and public: Micron's new fabs contribute from CY2028 (capex ~$27B FY26, rising FY27); SK hynix M15X ramps 2027; Samsung P5 ~2028 (§02·B). Until then, supply growth is mostly node migration — the structural underpinning of the thesis.
Trips when: industry bit-supply growth prints above demand growth — watch the first supplier to re-guide bit growth up while blended ASPs flatten.

3 · Packaging unlock (CoWoS)

WATCH · EASING
TSMC CoWoS: 75–80K wpm → 120–140K by end-2026, ~170K end-2027; OSATs add 40–60K (industry → ~200K). The supply-demand gap narrows from ~20% to ~10% by end-2026 (TrendForce). CoWoS-L/S remain fully booked, 52–78-week leads, NVIDIA ~60% allocation.
Two-sided: if packaging slots outrun HBM supply, memory becomes the sole binder (bullish); if both unlock together into softening token demand — accelerator glut, and memory follows.

4 · China supply ramp

WATCH
CXMT/YMTC positioning is tracked in §09 — the tripwire is HBM-class qualification at scale or China DRAM bit share pushing past the mid-teens, either of which would import the next downcycle early. Meanwhile the H200's approved export to China adds a demand vector that cuts the other way.
Trips when: a Chinese maker ships qualified HBM3-class stacks to a major accelerator, or export-grade capacity additions accelerate.

5 · Demand destruction

WATCH
Memory is now inflating end-device BOMs: Samsung is taking 2026 HBM contracts up high-teens/low-20s%, and commodity DRAM/NAND spot is up 60–85% QoQ. PCs, phones and consumer SSDs absorb this with lags — elasticity eventually answers (TrendForce has flagged early build-plan trims).
Trips when: major OEMs cut unit forecasts citing memory cost, or channel sell-through rolls while sell-in holds — the classic double-ordering tell.

6 · The valuation paradox

WATCH · THE CLOCK
Memory stocks historically top when they look cheapest: in mid-2018 MU peaked around ~$64 at roughly ~5× forward earnings — then fell >50% as the estimates, not the multiple, collapsed; 2021–22 rhymed [approx., from history]. Today MU trades ~9× record EPS (§04). Cheap ≠ safe when E is the cyclical variable. Clock inputs: Micron inventory ~120 days (FQ3), and from Jul 10 the SK hynix ADR gives U.S. money a same-exchange relative-value valve.
Trips when: consensus FY-forward EPS gets cut for the first time while the multiple expands — the historical signature of the top.
Status ladder — all six at a glance · dated judgment, Jul 2 2026
CALM
— none —
WATCH
#3 Packaging (easing) · #4 China ramp · #5 Demand destruction · #6 Valuation (the clock)
ARMED
#2 Supply response — dated CY2028 (greenfield bits not here yet)
TRIPPED · EARLY
#1 Pricing-power inflection — the rate broke (3Q26 decel), not the level

The spread is the point: five of six are not tripped, and the one that is tripped early on the second derivative, not the first. Categorical placement mirrors each card's chip above — a comparative re-plot, not new data.

Method. Statuses are dated editorial judgments (Jul 2 2026) applied to sourced readings: Micron FQ3 call and prepared remarks (moderation guidance, SCA ceilings, inventory days, price moves); §02·B fab timeline (Micron CY2028 greenfield, SK M15X 2027, Samsung P5 ~2028); TrendForce CoWoS reporting (Jun 2026) and TSMC expansion disclosures; Samsung 2026 HBM contract reporting; §09 for China positioning; 2018/2021 cycle references are approximate historical figures, labeled as such. This section exists because of the honesty contract: a thesis that can't name what would falsify it isn't a thesis. Not investment advice.

02·ECatalyst Calendar — Every Dated Event the Thesis Meets Next◷ 43 dated events · +Korea export cadence
02·E

Catalyst Calendar

Every dated event the thesis meets next, in one place — now including the full quarterly reporting schedule across the complex: memory makers, hyperscalers, accelerator vendors, WFE/foundry, and the edge names whose margins reveal cost pass-through. Each row says which sections it feeds and which kill-switch (§02·D) it tests, so this reads as a falsification schedule rather than a hype reel. Completed events carry what actually happened, not just that they occurred. Dates marked confirmed come from company IR announcements; ~ approx are windows inferred from reporting history and will shift. Aug 15 2026 — South Korea's export releases are now on the calendar as a recurring cadence, because they are the highest-frequency hard data the memory thesis meets: the Korea Customs Service publishes flash exports for the 1st–10th (around the 11th) and the 1st–20th (around the 21st), and MOTIE publishes the full month on the first business day of the next. The 10-day flash carries the DRAM ex-modules unit price in $/kg — the single fastest public read on memory pricing on this board, and the series §03·B is built from. Nine such releases are scheduled below.

Provenance. Rebuilt Jul 30 2026. [confirmed] dates are company-announced: AMD Aug 4, SanDisk FQ4 and Western Digital FQ4 Aug 5, Marvell Aug 20, NVIDIA FQ2 Aug 26, Broadcom Q3 Sep 8, Qualcomm FQ4 Nov 11. [~ approximate] dates are inferred from each issuer's reporting history and are explicitly marked as such — Seagate, Applied Materials, Oracle FQ1, Micron FQ4 (~Sep 29 per its recent late-September cadence), ASML/TSMC Q3, and the October Q3 cluster for Alphabet, Microsoft, Amazon, SK hynix and Samsung. Completed rows carry outcomes as reported and cross-referenced to the sections that consumed them, including the two that reshaped this board in the last fortnight: SK hynix's record-but-missed Q2 (Jul 29) which triggered the KOSPI's first back-to-back circuit breakers (§03·C), and Amazon's Jul 30 capex raise to $220B attributed to memory prices (§06·F·D). Correction carried forward: the SK hynix Nasdaq listing is recorded here at its priced $26.5B, not the ~$28–29B an earlier build of this calendar showed (see §04·F and the Correction Log). Honesty note: a calendar is a schedule, not a forecast — every row states what the event would corroborate or falsify, and none predicts an outcome. Scheduled dates move; confirm against issuer IR before relying on any of them. Korea release dates [rule-derived, marked ~ approx]: the nine Korean rows are not read off a published KCS calendar. They are computed from the stated cadence — KCS flash after the 10th and after the 20th, MOTIE on the first business day of the following month — then shifted forward past weekends and Korean public holidays. The weekend rule is not uniform, and this board got it wrong once already. MOTIE demonstrably publishes on the 1st even at weekends — the completed Aug 1 2026 row below was a Saturday — so the Sep 1, Oct 1 and Nov 1 dates are left on the 1st, including the Sunday. KCS weekend behaviour is not evidenced either way in anything this board has seen, so the one KCS date that lands on a Sunday (Oct 11) is shifted to Oct 12 as an assumption, and is the least reliable date in the set. Treat every Korean date here as ±1 business day; the cadence is reliable, the exact posting time is not, and Chuseok (Sep 24–26 2026) sits close enough to the Sep 21 flash to move it. Two Korean rows are completed and carry outcomes: the Aug 1 MOTIE July print (semiconductors $41.01B, +178.8%) and the Aug 11 KCS Aug 1–10 flash (semiconductors $9.95B, +155.4%, a record August high at 46.8% of all exports). Staleness fixed in this pass: the calendar's internal "today" was pinned to Jul 30, so every countdown read 16 days short and six already-past events still displayed as upcoming; it now reads Aug 15. Five of those six are earnings prints whose results this board has not yet captured; a sixth, Tesla's Jul 22 Q2, was already sitting completed with an empty outcome field and was caught by the same sweep. All six are marked completed and explicitly flagged as outcome-pending in the table rather than given invented results. Sources: company investor-relations calendars and earnings announcements, Investing.com, MarketBeat, TipRanks, Nasdaq, Wall Street Horizon, Korea Customs Service, MOTIE, Yonhap, Korea Herald, Seoul Economic Daily. Not investment advice.

03The Price Spiral◷ live feed
03

The Price Spiral

● live feed

A decade of contract prices in one chart — five years of history, today, and five years of consensus forecasts. The 2021 peak gave way to the brutal 2022–23 crash (DRAM and NAND each lost ~30–40%), a brief 2025 soft patch, then the AI ignition that became the supercycle. Forward of Q1 2026 (dashed), the lines plot the path most analysts now see: another ~10% climb into 2H26, a plateau through 2027, then a gradual softening as new fab capacity comes online in 2028–2030.

Contract Price Index — Q4 2025 = 100 · 2021 → 2030
Latest · 3Q26 guide (TrendForce Jul 3): DRAM +13–18% · NAND +10–15% QoQ — decelerating from 2Q26 · HDD ~+50% over 5 months (TrendForce, Gartner, channel data)
Source conflict — resolved Jul 14 2026 (flagged Jul 6, per the correction convention): the compiled 1Q/2Q26 bars encode +93%/+61% (DRAM) and +58%/+72% (NAND) QoQ, while the Jul-3 TrendForce-session record elsewhere on this board (§02·D, §10·A) describes 2Q26 as ~+low-90s% DRAM / +70–75% NAND. Re-verified against the primary source: TrendForce's 2Q26 forecast put conventional DRAM at +58–63% QoQ (headline +63%) and NAND up to +75% — matching the compiled +61%/+72% bars, on a conventional-DRAM basis (not blended). The stray "~+low-90s% DRAM" recorded elsewhere for 2Q26 was actually 1Q26's record (+90–95%) mis-attributed to the wrong quarter; corrected at §02·D and logged (Tom's Hardware / Yahoo Finance via TrendForce, Mar 31 & May 2026). The 3Q26 deceleration guide (+13–18% / +10–15%) is unaffected either way, and the in-app live feed supersedes when available.
DRAM NAND Flash HDD (consumer, est.) Projected Historical points are indexed approximations; the dashed forecast follows consensus from TrendForce, Gartner & Micron's TAM model. Forecasts are inherently uncertain.

HBM is priced per stack — each generation a steep premium (market estimates, 2026):

HBM3 · per stack
~$200
HBM3E · per stack
~$300
HBM4 · per stack (est.)
~$500
03·BKorea DRAM Export Price◷ live feed · Jul flash · 4 series Jun–Jul
03·B

Korea DRAM Export Price

● LIVE

South Korea exports roughly 70% of the world's DRAM (Samsung + SK Hynix), so its monthly customs export-price data is one of the cleanest real-time reads on the memory market — published faster than contract-price surveys. This tracks the average DRAM export unit price, monthly, over the past three years. It attempts to refresh live; opened as a plain file it shows the compiled series (through the latest reported month). Aug 1 update — the June wobble was noise, not a turn: June printed the first month-on-month decline in nine months and partially armed kill-switch #1; July reversed it hard, with the DRAM ex-modules unit price at $96,520/kg in the first ten days (+540% YoY) and semiconductor exports of $41.01B (+178.8%) — a second consecutive $40B+ month. The deceleration signal did not follow through.

Latest reported · Jul 2026 · +17% MoM
June's one-month dip did not hold: DRAM (ex-modules) export price ran to $96,520/kg in Jul 1–10, +540% YoY; July semi exports $41.01B (+178.8%), a second straight $40B+ month, on total exports of $98.89B (+62.8%) — the second-highest month ever behind June's $102.25B record
DRAM Export Unit Price Index — monthly · 3-year history
Indexed to Jan 2025 = 100. Orange dots are months anchored to a hard customs/MOTIE/NABO disclosure; the line between is interpolated from documented quarterly and YoY moves. The vertical spike from late 2025 is the AI supercycle hitting export prices. Basis note, now explicit: Korea publishes DRAM unit price on two bases and they differ by ~40% — excluding modules (raw chips) and including modules. June was $60,000/kg incl-modules but ~$82,260/kg ex-modules; the July flash figure of $96,520/kg is the ex-modules basis. This series is indexed, so it is the move that matters — but the July point is derived from the ex-modules month-on-month change (~+17%) applied to the index, and is therefore a derived anchor from a 10-day flash, not a full-month print. It will be restated when the full-month unit-price detail publishes.
DRAM export price index Month with a hard public anchor Jan 2025 = 100
The four export unit-price series — and why the number you quote matters CITED $/kg · PRIOR MONTH DERIVED
Korea Customs does not publish one memory price — it publishes four, and they disagree by more than 2×. The index above tracks the headline; this breaks it into the components that actually move it. Solid bar = the latest cited reading, faint bar behind = the prior month back-derived from the stated MoM change. The single most useful fact here: DRAM excluding modules and DRAM including modules differ by ~64% — modules are heavy per bit, so including them drags the average down. Most reporting doesn't say which basis it's quoting.
What the spread between these four is telling you. Read together they decompose the squeeze rather than just measuring it. HBM carries the highest price per kilogram (~$96K) — unsurprising, it is the most value-dense memory made — but it is rising the slowest, at +119% YoY. Meanwhile plain DRAM chips are up ~540–576% and NAND/SSD is the fastest sequential mover at +28% in a month. That ordering is the whole thesis in one line: HBM was already repriced; what is repricing now is everything else — commodity DRAM first, storage second — because HBM's wafer appetite (§05·C·G) is starving the conventional lines. The mirror image shows up in memory modules at just ~$30K/kg and falling 13.9% MoM: the assembled, heavier, lower-margin end of the market is the one part going the other way.
A second vendor set, kept separate
A widely-circulated set labelled "May 2026" but published May 12 — which, given Korea's release calendar, is almost certainly the April full-month print rather than May. Its DRAM figure ($89,498) does not chain with the June −6%/+6% sequence above, which is the tell. It is shown here rather than merged, because the two cannot both be May on the same basis.

Hard anchors (real customs/government data): Jan 2026 DRAM export price/kg = $28,057, +133.6% YoY, ≈+60% MoM, monthly export value a record ~$8.67B (Hankyung Aicel / Korea Customs); Mar 2026 semiconductor exports $32.83B, +151.4% YoY — the first month ever above $30B (MOTIE); Q1 2026 DRAM exports $35.79B (+249.1% YoY) and NAND $5.39B (+377.5%) on the revised MTI codes, while export tonnage fell ~12% YoY — 2.7× the value on less weight, the HBM-mix signature (MOTIE, KITA, Korea Herald, AJU); May 2026 DRAM exports $18.6B, +369.8% YoY; semis $37.16B (+169.4%), 42.3% of all exports (KITA, TradingEconomics); Jun 2026 DRAM export price/kg ≈$60,000, down ~$1,000 MoM — the first monthly decline in nine months, with semiconductor exports $44.82B (43.8% share) as total Korean exports crossed $100B for the first time (Seoul Economic Daily / MOTIE, Jul 1). Revision note: the Feb–May 2026 interpolated points were re-fit upward when the May/Jun $/kg prints arrived — the earlier path (built before those prints) understated the spring surge by ~40%; 2025 quarterly DRAM fixed price $1.35 → $2.12 → $5.30 → $8.13 (NABO, Korea Customs & Trade Development Institute); Dec 2024 chip exports +31.5% on HBM/DDR5 strength despite general-memory price softness (MOTIE); 2023 documented downcycle — general-purpose memory prices fell sharply (the ~30% chip-revenue collapse). Months between anchors are interpolated to fit the documented quarterly averages and reported MoM/YoY moves. This is an indexed reconstruction calibrated to public Korean trade data, not a verbatim monthly customs series (Korea publishes export value and weight monthly; a clean per-bit price index requires combining them). Real-time updates depend on running inside Claude; the baseline is current to the latest reported month. Sources: Korea Customs Service, MOTIE, Bank of Korea, NABO, KED Global, Hankyung Aicel. Not investment advice.

03·ECost per Gigabyte — 25 Years of DDR, and the First Real Reversal◷ modeled + cited
03·E

DDR Cost per Gigabyte — 2001→2026

The longest lens on the board. For a quarter century, DRAM had one economic constant: $/GB fell relentlessly — from roughly ~$180/GB in 2001 (DDR1) to a glut-trough near ~$1.60/GB in 2023, a ~110× decline, punctuated only by cyclical bumps (2017–18’s supercycle among them) that always mean-reverted within ~2 years. 2024→2026 is different in kind: ~6× off the trough to ~$10/GB — consumer DDR5 back above 2014 levels, a decade of deflation erased — and for the first time the cause isn’t a fab accident or a demand blip but a structural reallocation of wafers to AI memory. Whether this is the biggest-ever bump on the old curve, or the end of the curve, is kill-switch #1’s question in long-run form.

Approximate street/spot price per GB for mainstream DDR modules, log scale (each gridline 10×). Solid = industry-series history [modeled/approx]; the 2024–26 tail is anchored to cited moves. Generation markers show DDR1→5 transitions. Hover-free static render for durability.

Series construction & honesty labels. 2001–2023 [MODELED/approx]: capacity-weighted street prices for mainstream DDR modules, following the industry-standard long-run series convention (McCallum / Our World in Data lineage): ~$180 (2001, DDR1) → ~$90 (2005, DDR2) → ~$12 (2010, DDR3) → ~$4.50 (2015, DDR4) → ~$8 at the 2018 supercycle peak → ~$3.50 (2019) → ~$1.60 at the 2023 glut trough (DDR4/early DDR5). Annual points between the named anchors are interpolated on the same series convention; point values are approximations of a noisy retail distribution, good to ±30% — the shape is the data, not any single point. 2024–2026 [cited]: the tail is anchored to reported moves — DRAM and NAND prices up more than 300% since 2023, with TechInsights projecting continued increases into 2027; DRAM/NAND contract prices jumped 90–95% QoQ in Q1 2026; Gartner projects a 130% year-on-year DRAM price rise for 2026, and even legacy DDR4 — no longer used in phones — is spiking, a signal of how distorted the market has become. Mid-2026 mainstream DDR5 lands ~$9–12/GB at street level (~$300+ for a 32 GB kit that was ~$90 in 2024). Why it matters here: every prior bump on this curve mean-reverted as supply caught up — the bet embedded in memory-maker valuations (§04, kill-switch #6) is that this one doesn’t, because the marginal wafer now has a structurally better-paying customer (§05·C·G, §06·D·W). The falsification path runs through kill-switch #1 (§02·D): the 3Q26 contract-price deceleration (§03) is the first data point of the answer. Modeled series + cited tail — not investment advice. See also: §08·G, where the same cost curve running backwards shows up as PC price rises, and §03·F for the spot and contract tape driving it today.

03·FSpot & Contract DRAM — TrendForce's Board, Capture by Capture◷ dated weekly prints + live
03·F

Spot & Contract DRAM — TrendForce's Board, Mar 2025 → Today

The spot market is the supercycle's leading indicator — thin, twitchy, and honest about scarcity before contracts admit it. This rebuilds TrendForce's spot board item by item, print by print, from 17 archived captures of the page (Mar 2025 → Jul 2026), TrendForce's weekly "Memory Spot Price Update" posts, and today's live read. Every dot is a dated publication — nothing between them is invented. The series now starts before the fire: mainstream DDR4 8Gb at $1.50 in Mar 2025 → $42.45 today, a 28× repricing. The news in this refresh is not the level — it is the deceleration. TrendForce's last four weekly prints on the mainstream chip read +3.27% → +1.39% → +0.17%, and on today's board that item printed 0.00%. Four of the seven items on the live board moved exactly zero today. Prices are at record highs and have stopped going up. Three further things a single line hides. (1) DDR4 outran DDR5 by 3×: DDR4 16Gb went $3.02 → $86.74 (28.7×) while DDR5 16Gb went $5.14 → $51.77 (10.1×) — the EOL squeeze on the "obsolete" node beat the leading-edge part decisively. (2) The quality tier split in spring 2026: branded chips kept climbing while the speculative eTT tier peaked in March and rolled over — DDR4 8Gb eTT −39.0% off its high and now below the level it entered the board at in Nov 2025, DDR4 16Gb eTT −17.1%. Both were flat-to-down again this fortnight while branded chips rose. (3) DDR3, the oldest node here, is up 17× ($0.81 → $14.02) as buyers shortage-shifted down the stack. Below the spot chart, the same archive method is applied to the contract tab — 15 captures, 105 points — which matters because contract is where the volume actually trades and is what §03's price ladder models. The new Jun 30 contract capture closes calendar 2Q26 and retires a caveat this section carried for three weeks: the quarter came in at +61.5% on DDR4 8Gb chips, inside TrendForce's own +58–63% headline, which this board could previously only call "consistent with, not confirming."

All seven spot items — every dot a dated publication, log scale · click any item to toggle
Every dot is a dated publication — ○ hollow = archived page capture (the Session Average as it stood that day), ● solid = TrendForce weekly post, ◎ ringed = today's live read, dashed ring = a superseded secondary anchor. Hover any dot for item, date, value and source type. Dashed segments = no print between those dates. Toggle any item to rescale both axes. Two extra series are available but off by default (†): DDR4 16Gb (1Gx16) and DDR4 8Gb (512Mx16) — organizations TrendForce retired from the board between the Nov 11 and Nov 26 2025 captures, when the three eTT tiers replaced them. Their lines therefore stop in Nov 2025 by construction, not by market closure, and the eTT series begin there for the same reason.
Contract prices — the other half of the board, 15 captures, Jan 2025 → Jun 2026
The same archive method applied to TrendForce's DRAM Contract Price tab — 105 transcribed points, every one a dated capture (○). Contract is where the volume trades; spot (above) only leads it. Two things jump out. Chips outran modules ~2×: DDR4 16Gb 2Gx8 contract went $2.55 → $42.00 (16.5×) while DDR4 16GB SO-DIMM modules went $32 → $227 (7.1×) — module makers absorbed part of the chip move before passing it on. And the step is visible: the Jan-2026 print jumped SO-DIMM modules +83% to +96% in a single update — the moment contract pricing caught up to the spot market's autumn ignition. And in the newest capture the two tiers separated: on Jun 30 both DDR4 SO-DIMM modules printed exactly 0.00% while every chip item rose +5.0% to +13.6% — module pricing paused for the month, chip pricing did not. The public page still shows the 2H-Jun session; TrendForce's July contract update exists (stamped Jul 31 2026) but is Gold+ only, so this tape is structurally about a month behind the spot tape above — which is part of why spot leads it.
Today's full spot board — live page read, Aug 10 2026 18:10 GMT+8
Read the spot tape with both hands
(1) Spot is thin and chaotic, and getting thinner: today's DDR4 8Gb session ranged $20.60–$74.00 around a $42.45 average, and branded chips trade at 10.1× the eTT tier ($42.45 vs $4.21) — a bimodal, low-liquidity market where "the price" is a wide distribution, not a number. TrendForce's own language for the last three weeks is consistent: quotes are up but "a consensus has yet to form in the higher price ranges, keeping overall trading volume limited." A tape that stops moving because nobody is transacting is not the same as a tape that stops moving because supply loosened, and this chart cannot distinguish the two. (2) Spot ≠ contract: the §03 ladder is quarterly contract pricing (most volume); spot is the marginal trade. It led the whole cycle — spot was up ~5× (Apr→Oct 2025) before 1Q26 contracts printed +90–95%. So the flattening is the thing to watch: if spot leads, a mainstream chip printing 0.00% while TrendForce still guides server contract +13–18% QoQ for 3Q26 is the early tell that the contract deceleration continues past that. If it merely reflects an August lull in a thin market, it means nothing. Two more weekly prints will separate those. (3) The spring stall says be careful: Mar 24 $34.00 → May 5 $32.40 (−4.7%) while contracts surged — buyers briefly refused the marginal price, and then the tape resumed and made new highs. This board called that a stall, not a top, and was right. The same restraint applies now.

Provenance & method — and a sourcing note. Three source types, all dated, none interpolated. (1) Archived page captures [17 dates, 111 transcribed points]: the Session Average exactly as the page displayed it, from Wayback snapshots of trendforce.com/price/dram/dram_spot supplied as screenshots and transcribed here \u2014 Last-Update stamps 2025: 03-20 18:10, 04-29 14:40, 05-23 18:10, 06-13 18:10, 08-11 18:10, 09-05 18:10, 10-02 18:10, 11-11 18:10, 11-26 11:00, 12-22 14:40; 2026: 01-22 11:00, 02-13 18:10, 03-13 18:10, 03-25 18:10, 04-17 18:10, 06-22 18:10, 07-17 18:10 (all GMT+8) — the Jul 17 read was this section’s “live” point until this build and has been re-marked as an archived capture, which is what it now is. Board schema change [dated]: captures through 2025-11-11 list DDR4 16Gb (1Gx16) and DDR4 8Gb (512Mx16); by 2025-11-26 TrendForce had replaced both with the three eTT tiers. Those two retired items are kept as opt-in series (\u2020) rather than dropped, and no eTT value is back-filled before its first capture. (web.archive.org is blocked at this build environment's fetch layer, so these were read by the user's browser and transcribed rather than fetched \u2014 the values are the archive's, the transcription is ours.) (2) TrendForce weekly "[Insights] Memory Spot Price Update" posts \u2014 which quote only the mainstream DDR4 1Gx8 figure numerically, hence that series alone is weekly-dense. (3) Live page read (Aug 10 2026 18:10 GMT+8) — DDR5 16Gb $51.767 (+0.32%), DDR5 16Gb eTT $23.64 (0.00%), DDR4 16Gb $86.744 (+0.55%), DDR4 16Gb eTT $11.40 (0.00%), DDR4 8Gb $42.45 (0.00%), DDR4 8Gb eTT $4.208 (0.00%), DDR3 4Gb $14.024 (+0.34%). Cross-check [passed]: where the two independent streams overlap they agree \u2014 archive 2026-03-25 $34.00 vs weekly Mar 24 $34.00; archive 2026-06-22 $35.90 vs weekly Jun 24 $35.90; archive 2025-12-22 $21.142 vs weekly Dec 23 $21.75 (one day apart in a rising tape). The archive captures also close the old Dec\u2013Mar print gap that this chart previously drew dashed. Cited weekly prints (mainstream DDR4 1Gx8 3200 average): 2025 — Apr 15 $1.673, Apr 22 $1.720 (+2.81%); Oct 15 $7.219, Oct 22 $7.931 (+9.86%); Oct 29 $9.523, Nov 5 $10.629 (+11.61%); Nov 12 $11.857, Nov 18 $12.757; Nov 25 $14.914, Dec 2 $16.514; Dec 9 $17.064, Dec 16 $19.86, Dec 23 $21.75 (+9.52%). 2026 — Mar 24 $34.00, Mar 31 $33.96; May 5 $32.40, May 26 $33.60; Jun 2 $34.80, Jun 9 $35.90, Jun 24 $35.90, Jun 30 $36.00, Jul 1 $36.10, Jul 7 $37.14 (+2.88%); Jul 15 $39.80; Jul 17 $40.50 (page capture, +1.50% on the day); Jul 21 $41.10 (+3.27%), Jul 28 $42.08 (+1.39%), Aug 4 $42.11 (+0.17%); Aug 10 $42.45 (live page session average, 0.00% on the day). Source-consistency note — TrendForce restates its own prior week, and the gap is bigger than the move: the Jul 29 post gives the previous week as $41.50 (Jul 22) where the Jul 22 post itself printed $41.10 (Jul 21); the Aug 5 post gives the previous week as $42.04 (Jul 29) where the Jul 29 post printed $42.08 (Jul 28). Survey dates differ by a day and the restated figures differ by 1.0% and 0.1% respectively. This chart plots each post's own headline "this week" figure and does not plot the restatements. The practical consequence is that a single weekly print here should be read as good to roughly ±1% — which is smaller than the moves through 2025 and larger than the +0.17% move that this update's headline rests on. That is stated rather than smoothed: the deceleration is unambiguous across three prints; the final week's near-zero is inside the source's own restatement noise. Point counts per item are on the legend buttons: DDR4 8Gb carries 46 (archive + every weekly post), DDR4 16Gb / DDR5 16Gb / DDR3 carry 18 each (Mar 2025 onward), and the three eTT tiers carry 10 each (from their Nov 2025 introduction). TrendForce's own numeric per-item history remains Gold+ members-only, so this is the publicly reconstructable series, not the full daily tape. Each legend button states its point count. Quality-spread note: eTT (\u201ceffectively tested\u201d \u2014 untested/lower-grade die) trades at a steep discount and is the speculative tier; its March peak and subsequent rollover, against branded chips still rising, is the clearest divergence on this chart and is drawn from the captures, not modeled. Context prints: DDR5 spot +30% in one week (Nov 5 2025) and +307% since September by Nov 19 2025 (TrendForce headline); the Jan 2025 $1.63 figure is a secondary-source anchor (Silicon Analysts) and is now superseded: the 2025-03-20 capture shows DDR4 8Gb at $1.50, below it, so $1.63 was not the trough. It is kept as a dashed-ring point for provenance, relabelled, and logged. Gaps between prints (esp. Jan–Mar 2026) are publication-fetch gaps, not market closures — bridged dashed, never invented. Spot quotes are DRAMeXchange/TrendForce session averages; the live-board updater (weekday 7am) watches the same page. Contract-price series [15 captures \u00d7 7 items = 105 points]: same method, TrendForce's DRAM Contract Price tab, Last-Update stamps 2025: 01-24, 02-27, 04-30, 06-30, 07-31, 08-29, 09-30, 10-31, 11-28, 12-31; 2026: 01-30, 02-26, 03-31, 05-29, 06-30 (GMT+8). Contract prints are monthly/semi-monthly sessions, so the line is a step function by nature, not an interpolation. Cross-check against \u00a703's contract ladder [independent confirmation]: measuring quarter-end to quarter-end on these captures, 1Q26 (Dec 31 \u2192 Mar 31) gives DDR4 16GB SO-DIMM +95.4%, DDR5 8GB SO-DIMM +89.9%, DDR4 8GB SO-DIMM +82.8% \u2014 which brackets the +90\u201395% TrendForce headline this board cites in \u00a703, and confirms that headline tracks module contract pricing. The chip items moved less over the same window (DDR4 16Gb 2Gx8 +40.5%), so item choice, not error, explains the spread \u2014 recorded here rather than merged. 2Q26 is now complete, and it confirms the figure this board has been citing. Measuring Mar 31 → Jun 30 on these captures: DDR4 4Gb +73.3%, DDR3 4Gb +66.7%, DDR4 8Gb +61.5%, DDR5 8GB SO-DIMM +53.3%, DDR4 16Gb +42.4%, DDR4 8GB SO-DIMM +40.0%, DDR4 16GB SO-DIMM +33.5%. The chip items bracket TrendForce's +58–63% 2Q26 headline; the module items sit below it, the same item-choice spread this section documented for 1Q26. The prior build could only call this "consistent with, not confirming" — it now confirms. Independent arithmetic check on the new capture [passed]: the Jun 30 page publishes its own month-on-month change for each item, and all seven reconcile exactly against this archive's May 29 values — SO-DIMM 16GB 0.00%, SO-DIMM 8GB 0.00%, DDR5 SO-DIMM +2.68%, DDR4 16Gb +5.00%, DDR4 8Gb +5.00%, DDR4 4Gb +8.33%, DDR3 4Gb +13.64%. That is seven independent confirmations that the earlier transcription was right. 3Q26 guidance [cited, Jul 9 2026]: server DRAM contract +13–18% QoQ, decelerating hard from 2Q26 — because several U.S. CSPs have signed multi-year LTAs that bar suppliers from raising their prices, so from 3Q26 the increases come from non-LTA customers and from incremental volume sold outside the LTAs. TrendForce also flags 2027 RDIMM bit supply growing only 15–20% YoY, behind server CPU shipment growth, and a server DRAM shortage already anticipated for 2027; CSPs are shifting RDIMM configurations down from 96/128 GB to 32/64 GB modules to contain cost. Spot-to-contract premium, same-published-date: on Mar 31 2026 DDR4 8Gb spot was 2.61× its contract price; by Jun 30 that had narrowed to 1.71× — contract catching up to spot, which is what the end of a lead looks like. What this refresh does not show, and it matters: the sector's equity peak was Jun 22 2026 (§04·K), after which the leveraged semiconductor complex fell 63% — while over the same weeks this spot tape went $35.90 → $42.45, up 18%. The physical market kept tightening while the equity market broke. Spot did not lead the equity drawdown; it did not confirm it either. Whatever repriced those shares (§03·C's Korean margin unwind, §04·K's leverage mechanics, or a valuation de-rate) is not visible in DRAM prices, and this chart is evidence against reading the equity move as the market front-running a memory price collapse — at least so far. Sources: TrendForce weekly spot update, Aug 5 2026, TrendForce 3Q26 server DRAM contract forecast, Jul 9 2026, TrendForce News weekly spot updates (dated URLs), trendforce.com/price/dram/dram_spot and /dram_contract, Silicon Analysts. Not investment advice.

03·GNAND Flash — Spot & Contract, the Swing Factor◷ cited weekly prints
03·G

NAND Flash — the Swing Factor

The other half of the memory market, and the looser half: NAND supply is easier to add than DRAM, so its cycle turns faster and its gap is tighter (§08). This tracks TrendForce's benchmark 512Gb TLC wafer spot price — every dot a dated weekly "Memory Spot Price Update" print — plus the NAND contract ladder. From $2.76 (Apr 2025) to $19.86 (Jun 2026), a ~7× repricing that trails DRAM's but rhymes with it. Enterprise-SSD restocking is the demand swing.

512Gb TLC NAND wafer — spot price, weekly prints, log scale
Every dot is a dated TrendForce weekly print. This is a starter series (3 cited prints so far) — a full per-item NAND capture archive parallel to §03·F awaits page captures; the dots are the data, the line guides the eye.
NAND contract price — QoQ step, by quarter

Provenance & method. Spot [cited]: TrendForce weekly "[Insights] Memory Spot Price Update" — 512Gb TLC wafer session average: $2.764 (Apr 23 2025), $7.421 (Nov 19 2025, +14.97%), $19.862 (Jun 30 2026, −1.03%). Contract [cited]: NAND contract QoQ ≈+58% (1Q26)+72% (2Q26)+10–15% guide (3Q26), decelerating in step with DRAM (TrendForce, §03·§02·D). Honesty note: this is deliberately lean — three public spot prints and the contract ladder — versus §03·F's 202-point DRAM archive, because NAND per-item captures haven't been supplied yet; paste them (as with the DRAM board) and this becomes a full multi-item tape. NAND's looser supply is exactly why it's the cycle's swing factor: a faster NAND turn is a leading tell for the broader memory cycle (kill-switches §02·D). Sources: TrendForce News weekly updates. Not investment advice.

03·CLeverage in the Korean Market◷ Aug 15 2026 · leverage relocates
03·C

Leverage in the Korean Market

◷ snapshot

The memory rally has a leverage story, and it's concentrated in Korea — home of Samsung and SK Hynix, and one of the most retail-leveraged equity markets on earth (the local term for debt-fueled trading is bittu). Margin debt hit a 20-year high in 2026, retail piled into the two chip giants, and on May 27, 2026, Korea launched its first-ever single-stock leveraged ETFs — 2× products on Samsung and SK Hynix that drew billions in days. This is the most direct measure of how much risk retail is putting on the memory trade.

Aug 15 2026 — the leverage did not leave. It relocated. And this section's Aug 1 read was too strong.
Two weeks ago this section concluded that "the appetite has now broken too." The domestic half of that was right. The word "appetite" was wrong. What broke was one instrument, and it broke because it was regulated, not because anyone stopped wanting leverage. The domestic channel closed almost completely. Turnover in the single-stock leveraged and inverse funds fell from ₩12.4485T on Jul 30 to ₩3.1518T on Jul 31 — a 74.7% collapse in one session, the day the ₩30M all-cash deposit took effect — then to ₩1.3872T (Aug 3) and below ₩1T for the first time since listing. By Aug 14 the four KODEX and TIGER flagship funds averaged ₩735.6B a day, down 91.0% from ₩8.16T a month earlier. Their share of all KOSPI trading went from a July average of 33% to 4.8%. On that evidence the intervention worked exactly as designed. Then look where the money went. Korean retail sold $664.39M of the Direxion Daily Semiconductor Bull fund on Aug 3 and $1.639B cumulatively through Aug 11 — and then bought $662.85M back over Aug 12–13 alone, about thirteen times their next-largest purchase. Their holding in that single fund is $6.559B, now their fourth-largest US position behind only Tesla, Nvidia and Alphabet, up from tenth at the end of April. The replacement instrument carries more leverage than the one that was restricted, not less — 3× versus 2× — traded in a jurisdiction Korea's FSC does not regulate. The only thing that changed is that the leverage is now on a semiconductor index rather than on two named stocks, which diversifies the single-name risk and does nothing about the reflexivity. And the domestic borrowing is rebuilding. Margin loans fell roughly ₩10T from the ₩38.02T peak through the forced-liquidation wave — and have now risen seven consecutive sessions to ₩30.9262T (Aug 13), with investor deposits back above ₩100T. The market recovered alongside it: the KOSPI closed 6,977.94 on Aug 14, +2.42%, a fifth straight gain on $2.1B of foreign buying; Morgan Stanley moved Korea to overweight on Aug 2, arguing the leverage unwind was past halfway. The KOSDAQ tripped buy-side sidecars three sessions running — the mirror image of July. The honest conclusion is narrower than either "it's over" or "it's broken": a regulator can close a specific venue very fast, and closing it does not reduce the total appetite for leverage. It moves it somewhere with less oversight. The instrument this section was built to track has been largely switched off; the behaviour it was built to measure has not.
Where the leverage went — the domestic channel closed, a 3× US fund opened
Panel A is domestic single-stock leveraged and inverse ETF turnover per day. Panel B is Korean retail net buying of a US 3× semiconductor fund over the same fortnight. They are different instruments in different currencies under different regulators, and they are deliberately not netted against each other — the point is the timing, not a balance. Two bars are drawn dashed because they are not straight cited prints: the second Aug 4 bar in Panel A is a conflicting figure for the same day from the same report, and the Aug 14 bar is a different fund basis (four funds, not sixteen). In Panel B the Aug 4–11 bar is derived by subtraction from two cited cumulative figures — the arithmetic closes to the cent, but nobody published that number. Hover any bar.
Three numbers in the new reporting that this board cannot make agree
(1) The drawdown. The Aug 14 report describes the KOSPI as having fallen "as much as 38.6% from its annual high". This board's own dated anchors — a 9,385.59 peak and a 5,593.56 close on Jul 30 — give −40.40%, and the intraday break below 5,300 gives more. A −38.6% drawdown from 9,385.59 implies a trough of 5,762.8, which is above every close this section has recorded for that week. Neither figure is adjusted to fit the other. (2) The rebound. The same report says the index "rebounded 5.8% this month". From this board's last recorded close of 5,593.56 to the cited Aug 14 close of 6,977.94 is +24.75%. For "+5.8%" to be month-to-date, August would have to have opened near 6,595 — which would require a +17.9% single session on Jul 31, an all-time record by a wide margin that no source reports. The likeliest reading is that the cited percentage is measured over a shorter window than a full month, but nothing published says so, so the Aug 14 close is used as the anchor and the percentage is reproduced rather than relied on. (3) The peak date. That report dates the annual high to June 22; this board carries June 19, and a third translation earlier gave "July 19", a future date. Three dates for one high. June 19 is retained because it is the one corroborated across the most sources, and the disagreement is logged rather than hidden. Also noted for source quality: the Aug 6 report underlying Panel A prints two different turnover figures for Aug 4 (₩1.2556T and ₩919.8B) and labels two different days "Aug 4" in consecutive sentences — both figures are drawn rather than one silently chosen.
Jul 13–30 2026 — the unwind, the intervention, and then the worst two days in KOSPI history
This section warned that Korea’s retail leverage was the most fragile plumbing under the memory trade. That risk has now materialized, and the mechanism ran exactly as described. The KOSPI peaked at 9,385.59 on Jun 19 and has fallen −27.5% since; Jul 13 closed −8.95%, tripping a sidecar and a Level-1 circuit breaker. VKOSPI hit 96.94 — highest since the global financial crisis. June alone set a record three circuit breakers and eleven sidecars. Forced liquidations went from a ₩91.7B May record to ₩425.8B in the first ten days of July. And the single-stock 2× ETFs are structurally implicated: to hold their leverage ratio, managers must buy as the stock rises and sell as it falls — a short-gamma reflexivity that Bloomberg Intelligence estimates forced ~$6.0B of mechanical selling of Samsung and SK hynix on the Jun 23 crash day alone, ~14% of their combined trading value. The funds amplified the melt-up; they are now amplifying the melt-down. Jul 16 — the intervention: new single-stock leveraged listings halted, deposit tripled to ₩30M all-cash, ads banned, 20-share minimum order (details in §03·D). Jul 17–23 — a brief stabilization (margin debt off its record, foreigners buying, KOSPI +4.40% to 7,096.89 on Jul 23) — which the following week erased. Jul 28–29: the worst two sessions in the index's history. The KOSPI fell 16.17% (1,092.51 points) across two days and tripped sell-side sidecars and circuit breakers on two consecutive days — a first, ever, crashing below 5,300 intraday. The trigger was SK hynix's Q2 print (Jul 29): all-time record results that still missed consensus, with an HBM outlook and shareholder-return plan below expectations (§04·H). What turned a miss into a crash is this section's own mechanism — retail panic selling colliding with the mechanical rebalancing of the single-stock leveraged funds. ~120,000 more investors took margin calls; 32,000–46,000 accounts were wiped out entirely. Regulators convened an emergency F4 meeting and economic chiefs apologised before parliament. Jul 30 — Samsung's record quarter could not lift it: operating profit ₩89.49T, +1,814% YoY, a clear beat — the stock rose 0.7%, but the KOSPI whipsawed from +5.5% intraday to close −1.23% at 5,593.56, SK hynix −5.64%. For July as a whole the index is now down roughly a third. The regulatory ratchet tightened again: the ₩30M all-cash deposit was accelerated to take effect Jul 31, and authorities added a cap limiting any individual to 20% of their portfolio in single-stock leveraged ETFs, mandatory simulated trading, and a new legal basis for emergency stabilisation measures (modelled on Hong Kong). The lesson this section has been making for two months, now printed twice: leverage does not exit quietly — it is liquidated, and the funds that amplified the rally amplify the fall. Aug 1 — the appetite has now broken too. For nine weeks the defining feature of this episode was that assets kept rising while prices collapsed — retail averaging down with leverage. That stopped: KODEX SK hynix 2×'s assets fell from ₩4.21T to ₩2.52T within days of the crash (−40%), its NAV is −76.9% on the month, and the ₩30M all-cash deposit took effect Jul 31–Aug 5 with some brokers halting new transactions outright. The cumulative toll: ₩34T (~$22.9B) wiped from retail margin accounts in a month, 320,000–460,000 accounts liquidated, and a public apology from the finance minister. Retail bought a net ₩14T of these funds since May 27; a large part of that is now gone.
Is the Korean Market Overheated? — 12-month trends

Six metrics that together gauge whether Korea's retail-driven, leverage-fueled rally is overheating, each over the trailing ~12 months. The composite score blends them into one read; the small charts show how each got there. Note the March spike in every stress metric — that was the Iran-crisis crash (KOSPI −12% in a day, its worst ever) and the forced-selling cascade that followed.

88
/ 100 heat
OVERHEATED

The signals are flashing overheated — but the earnings are real

Margin debt, volatility, and speculative ETF flows are all at or near records, and regulators have stepped in to curb leverage. Yet the rally rests on genuine record chip profits — so this reads as a stretched bull market with high unwind risk, not (yet) a pure bubble. The composite is a constructed index, not an official measure.

calmnormalelevatedoverheated

Dashed line on each chart = the metric's historically "normal" level. Heat flag reflects the latest reading vs that normal band. The composite score weights all six equally and is our construction for illustration, not a published index. X-axis months (Aug 2025 \u2192 Jul 2026) place each point in the trailing year; per the section provenance, the monthly series are approximations calibrated to the dated anchors named on each card \u2014 mid-series month alignment is approximate, while the dated anchors themselves (peaks, records, latest prints) are exact and cited.

2× Leveraged Single-Stock ETF AUM — the three memory makers

Assets under management in the 2× daily-leverage ETFs tied to each memory maker, drawn as monthly lines since each fund's launch. Note the radically different ages: Micron's MUU has traded since Oct 2024; the Samsung and SK Hynix products — Korea's first-ever single-stock leverage ETFs — launched May 27, 2026 and have now lived a full boom–bust round trip in nine weeks: category-record inflows into the Jun 25 peak, then an unwind that has kept deepening. Bloomberg’s Jul 14 verdict was “Leveraged Chip Bets Backfire in Korea” at −45% since debut; by Jul 28–29 — SK hynix’s record-but-missed Q2 and the KOSPI’s first back-to-back circuit breakers (§03·C) — KODEX SK hynix 2× fell 28.4% in a single session to ₩9,930 and all fourteen funds now trade below half their ₩20,000 listing price, with NAV down 61.6% in a month. On Jul 16 the regulators moved (listings halted, ₩30M all-cash deposit, ad ban, 20-share minimum), and on Jul 29–30 added a 20%-of-portfolio cap and mandatory simulated trading. The tell in the data: KODEX SK hynix’s assets rose to ₩4.21T even as its NAV halved — retail kept buying the whole way down.

Monthly observation window · since each launch → Jul 2026 · Korean flagships peaked Jun 25; unwind now cited through Jul 16 · US second wave live Jul 13–14
CSOP SK 2× (HK 7709)MUU · Micron 2×SNXX · SanDisk 2× (US)CSOP Samsung 2× (HK)KODEX+TIGER SK (KRX)KODEX+TIGER Samsung (KRX)● larger outlined dot = official/cited anchor · small dot = bridge estimate
Y-axis: USD billions. Cited anchors (the values sourced elsewhere on this board) are drawn as outlined dots; small dots are bridge estimates connecting them so the trend reads — use anchors when precision matters. SK hynix is covered via CSOP HK 7709 (and, since Jul 13–14, the US SKHY wave); Samsung via CSOP Samsung 2× (HK — one cited print, $1.65B; its flat line means no later print, not unchanged). Correction (logged): this note previously called SNXX a Samsung fund — SNXX is Tradr's 2× SanDisk (SNDK), the fourth memory maker on the chart. Lifespans differ wildly by design: MUU since Oct 2024, SNXX since Jan 2026, the CSOP SK fund only since May 27 2026 (→ $16.8B, world’s largest single-stock leveraged ETF, in six weeks). Jul 13 update — the boom and the bust are now both on the chart [cited]: the two Korean flagship complexes (KODEX + TIGER on each maker, listed May 27) are added. Their debut net assets are cited (₩2.0T SK / ₩1.75T Samsung); the Jun and Jul points are bridge estimates splitting the cited combined flagship totals — peak ₩15.41T on Jun 25 → ₩12.31T by Jul 7, a ₩3.1T drawdown in 12 sessions including ₩2.69T of pure valuation loss — by per-name arithmetic from the same FnGuide report — cumulative inflows minus valuation losses (SK ₩9.12T−₩1.54T ≈ ₩7.6T / Samsung ₩5.88T−₩1.15T ≈ ₩4.7T ≈ 62/38; this supersedes the earlier 59/41 Jun-3 split). Won→USD at ~₩1,495/$. Jul 14 [cited]: Bloomberg — the biggest single fund, KODEX SK hynix 2×, holds ~$3.4B and is −45% since its May 27 debut (“Leveraged Chip Bets Backfire in Korea”) — a per-fund print consistent with the pair estimate drawn here. The 14 leveraged chip products averaged −26.8% that month vs −4.8% for ETFs overall (§03·C). Jul 29–30 [cited] — month-end readings, and the deepest leg yet: KODEX SK hynix 2× reports net assets of ₩4.21T with its price at ₩9,930 after a −28.43% single session (Jul 28) and NAV −61.63% over the month; KODEX Samsung 2× at ₩1.5083T ($1.09B) and ₩9,415 after −26.90%. All fourteen funds are now below half the ₩20,000 listing price (Seoul Economic Daily, Jul 29) — they had all broken below listing by Jul 8. CSOP SK 2× (HK 7709) reads HK$45.17B ≈ $5.79B (Investing.com/Cbonds, Jul 30) with the price at HK$25.36 vs a HK$32.70 prior close — −36% from the $9.0B Jul 16 print and −66% from the $16.8B June peak. The KRX pair values are derived, not printed: only the KODEX legs are publicly quoted, so TIGER is scaled at its historical ~61% of KODEX (₩1.370T vs ₩2.227T on Jun 3) — giving ~₩6.8T (SK) and ~₩2.4T (Samsung), converted at ₩1,480/$. Those two bars are therefore estimates built on a cited anchor, and are the least precise points on this chart. Turnover is finally fading — “leverage trading loses steam” (Jul 27) — after a month in which these products were >25% of all Korean ETF volume. Jul 16 [cited] — the prior prints: CSOP SK 2× reads ~HK$71B ≈ $9.0B (issuer/quote pages, Jul 16; ~$13B on Jul 1 per Bloomberg) — −46% from the $16.8B Jun 23 peak. The previous dashed modeled tail (~$14.2B) is replaced by this cited point; the model understated the unwind by ~$5B — exactly why modeled segments are drawn dashed and replaced the moment a print exists. SNXX reads $3.22B (quote pages, from ~$3.5B mid-Jun). MUU's $5.9B remains a Jul 3 pre-selloff print: its NAV fell 26.65% in the Jul 13 session and still no post-selloff AUM print has been published — so the line is held flat rather than extrapolated, and it is certainly overstated today. SNXX likewise holds its $3.22B Jul 16 print (its shares changed hands near $8.82 on Jul 30, but share price is not AUM); CSOP Samsung 2× holds its single cited $1.65B. Three flat tails, three missing prints — not three unchanged funds. Jul 13–14 — the US second wave [cited]: SK hynix’s ADR (SKHY) priced 177.9M shares at $149 to raise $26.5B — the largest US IPO ever by a foreign company, past Alibaba’s $21.8B (2014), oversubscribed sevenfold, and debuted +13% above the offering price on Jul 10. One business day later the leverage arrived: Leverage Shares’ SKHX (2× long) and SKHZ (1× short — the only non-leveraged short) plus ProShares’ SKHU (2×) on the 13th; GraniteShares’ SKUU (2×) and SKDD (2× inverse) and a Kurv 2× fund on the 14th; Direxion’s SKHL filed — with Reuters reporting at least ten managers have applied. Fees are 0.75%, ~40% below the category average, and options on SKHX/SKHZ are expected within days; the same issuer ran this play at SpaceX’s IPO, where the long/short pair set the record for the biggest first week in ETF history at over $4B of volume in four days. None of these are drawn: they are days old and have no AUM print — and none is invented (same rule as RAM in §04·F). Correction: this board earlier described the offering as “~$29B downsized to ~$28B”; the priced figure is $26.5B ($149 × 177.9M ADS). Sources still differ (“near $28B” appears in the issuer’s own release) — the priced number is used here. Source note: a user-supplied reference chart shows MUU/SNXX ending ~$7.9B/$6.4B by Jun 2026 vs the ~$5.9B/$3.5B cited anchors drawn here — basis/date difference flagged, not merged (§04·F).

Jul 12 refresh — the boom met the decay math [cited]. Net assets: the 16 funds held a combined ₩6.75T ($4.89B) by Jun 3, led by KODEX SK hynix at ₩2.23T; the SK-hynix leverage complex then nearly doubled from ₩4.84T (Jun 10) to ₩9.15T (Jun 19), much of it NAV appreciation feeding ever-larger rebalancing trades. Then the turn: the two flagship managers’ Samsung+SK funds peaked at ₩15.41T on Jun 25 and shed ~₩3.1T in 12 sessions to ₩12.31T by Jul 7, with ₩2.69T of valuation lossesan average −26.8% month for the 14 leveraged products vs −4.8% for ETFs overall, and on the Jun 23 −9.99% KOSPI day Bloomberg Intelligence estimates managers mechanically sold ~$6.0B of the two stocks to rebalance — roughly 14% of their combined trading value, sell orders piled onto a falling market. Retail kept averaging down anyway: ₩1.58T net into the two SK funds and ₩772B into the two Samsung funds Jul 1–8, even as all fourteen fell below the ₩20,000 listing price, while institutions were heavy net sellers. Scale check: June volume in the 14 funds hit ₩212T — 26.6% of all Korean ETF trading. The US second wave is live: Leverage Shares lists SKHX (2× long) and SKHZ (1× short) on Jul 13, and Direxion has filed SKHL to trade shortly after the SKHY ADR. Source conflict flagged: Seoul Economic reports the CSOP SK hynix HK fund became the world’s largest single-stock leveraged ETF upon crossing $5.38B on May 7 — which collides with this board’s §04·F anchor of "$16.8B, world’s largest (Jun 23)". Possible series conflation (CSOP HK vs the Korean complex aggregate); the $16.8B provenance is scheduled for primary-source re-verification and is not silently rewritten here. Not investment advice.

The composite overheating score is a constructed index (six metrics, equally weighted, normalized to each metric's historical range) — an illustrative synthesis, not an official or standardized measure. Underlying anchors (real): margin loan balance hit a record ₩36.6T on May 15 when the KOSPI touched 8,000 (Korea Financial Investment Association); the VKOSPI averaged 50.3 in 2026 vs a historical 18.8, hitting 65.5 on May 21 (normal ≈20, >30 heightened, >40 extreme); forced selling reached ₩91.7B on May 18, the largest since July 2023; the KOSPI swung from a 5,288 record (Feb 3) through a −12% single-day crash in March (its worst ever, ~₩625B wiped) to 7,815 (May 21, +8.42%); foreign investors sold a record ₩21.14T in Feb 2026 and ~$13.2B in one May week; Korean retail poured ~$40B into US leveraged ETFs in 2025 ($7B in December alone), then ₩3.87T into the new domestic 2× chip ETFs in three days. Regulators (FSS/FSC) and brokerages (KB, Mirae, Toss, KakaoPay) imposed mandatory training and margin curbs. Bloomberg called the KOSPI the world's most volatile major index; Citi flagged "exuberance" and took profits. Jul 13 2026 — the unwind [cited]: margin loans peaked at a record ₩38.02T (May 29), having jumped ₩954B in a single day and stood at ₩37.33T ($24.0B) on Jun 30, +36% YTD but now 78% concentrated in KOSPI names (88% by borrowed-share value) while KOSDAQ margin shrank 21% and KOSDAQ turnover halved. Volatility broke records: VKOSPI closed 96.94 on Jun 29, the highest since the global financial crisis; June alone set a record three circuit breakers, with eleven sidecars since the single-stock ETFs listed; turnover in those funds exceeds 12× the ETF-market average. The reflexivity is explicit — to hold target leverage, managers and LPs must buy spot/futures as the underlying rises and offload as it falls, a “short gamma” structure that intensifies rallies and deepens declines. Then the break: the KOSPI fell 27.47% from its 9,385.59 peak (Jun 19) and closed −8.95% on Jul 13, triggering the sell-side sidecar and a Level-1 circuit breaker; it had already broken below 7,000 on Jul 12, intraday −6%, for the first time in about two months. Forced selling followed: ₩425.8B forcibly liquidated Jul 1–10, including ₩142.2B on Jul 9 — 10.2% of outstanding margin receivables, after the year’s high of 10.5% on Jun 9 (₩169.8B); because collateral shortfalls lag forced execution, more liquidation supply is still to come — the Jul 13 crash is not yet in the data. Valuation now sits at a high-5x forward P/E, below both the global-financial-crisis low and the post-Iran-war trough. Two source conflicts, flagged not resolved: (a) July forced-liquidation totals differ by outlet — ₩425.8B for Jul 1–10 vs ₩344.2B cumulative for July as of a Jul 13 report (different bases: unpaid-balance consignment trading vs all margin); (b) one translation dates the 9,385.59 peak to “July 19,” a future date — June 19 is used here. Not a vindication lap: this section flagged the fragility, but the composite score is still an illustrative construct, the 12-month series are monthly approximations calibrated to these anchors, and “overheating” never predicted timing — it described a structure. The structure did what structures do. ETF AUM (real): MUU ~$5.4B (Direxion, launched Oct 2024); KODEX/TIGER Samsung & SK hynix 2× Leverage launched May 27, 2026 — debut net assets ≈ ₩1.75T (Samsung) and ₩2.0T (SK hynix). Won→USD ~₩1,495/$. Sources: KOFIA, Korea Exchange, Bloomberg, CNBC, UPI, Seoul Economic Daily, Digital Today, ETF.com, Direxion. Leveraged ETFs are short-term trading tools, not buy-and-hold investments — and this is not investment advice.

03·DSK hynix 2× Leveraged ETF — Full Evaluation◷ snapshot
03·D

SK hynix 2× Leveraged ETF — Full Evaluation

◷ snapshot

A deep-dive on the single product that best embodies the leverage story above: the KODEX SK hynix Single-Stock Leverage ETF — the largest of Korea's first-ever single-stock 2× leveraged funds (launched May 27, 2026). It seeks twice the daily return of SK hynix's share price. Retail money flooded in, AUM led the entire category within a week — and then SK hynix fell ~14% over two sessions and the fund dropped ~26%, dragging essentially every holder underwater. It's a near-perfect case study of the overheating dynamics this dashboard tracks: huge leveraged retail inflows into a single AI-memory name at the top, then a violent unwind amplified by 2× exposure and negative compounding. Seven weeks in, the case study is complete: −45% since debut and >60% off its June peak, with ~$3.4B still inside (Bloomberg, Jul 14) — and on Jul 16 Korea’s regulators intervened: new single-stock leveraged listings halted, deposit tripled to ₩30M (now all-cash), advertising banned, minimum order raised to 20 shares.

KRX · SAMSUNG ASSET MGMT
KODEX SK hynix Single-Stock Leverage
Objective2× daily return of SK hynix
ListedMay 27, 2026 (KRX)
StructureSingle-stock, futures-based 2×
Mgmt fee0.29% (vs 0.7–1.0% HK)
Max daily swingup to ±60%
Peer productsTIGER, 1Q, ACE, RISE, SOL
Evaluation · High Risk

A textbook leverage blow-off, completed: category-leading AUM → −45% since debut → regulators halt new listings

KODEX won the asset race (₩2.23T by Jun 3; ~$3.4B on Jul 14 — still the category's largest) on Samsung Asset Management's distribution muscle. And the product did exactly what regulators warned, twice: the −26% two-session unwind (Jun 4–5), then the full collapse — −45% since debut, >60% off its June peak by Jul 14 — far deeper than SK hynix's own decline over the span: negative compounding realized, not hypothesized. On Jul 16 the FSC/FSS answered: new listings halted, ₩30M all-cash deposit, ad ban, 20-share minimum order. Useful as a short-term trading tool; structurally hostile to holding — now effectively official policy.

AUM & Price Since Launch — daily

Net assets (bars, left axis) and fund price (line, right axis) since the May 27 listing. Note the timeframe: this product is ~7 weeks old (listed May 27), so the full span here is May 27 → Jul 14 — a since-inception tracker, not a 12-month history (a 12-month series is impossible — it didn't exist before May 27). The solid line and solid bars are the launch-fortnight daily series (calibrated to dense reporting through Jun 12); after that the daily prints stop, so the dashed segment is a sparse-anchor extension — outlined dots are dated anchors (₩19,635 close Jul 8 [cited]; ≈₩11.0k on Jul 14, derived from Bloomberg's −45%-since-listing; AUM ≈₩5.1T = $3.4B [Bloomberg]) and small open dots are bridge estimates shaped by the documented tape. It ends at the last dated anchor: Jul 15–16 have no fund print yet (the ADR fell both days, so the direction is known — the level is not, and none is invented). The story the two panels tell together: the price collapsed 45% while net assets more than doubled — retail averaging down with leverage, all the way in.

KODEX SK hynix Single-Stock Leverage: net assets (₩ trillions, bars) and price (₩, line), May 27 → Jul 14 on a real date axis. Solid = launch-fortnight dailies (calibrated to prints; ₩2.04T cumulative net retail buying through Jun 5); dashed line + hatched bars = sparse-anchor extension (outlined dots cited/derived, small dots bridge). The dashed gray rule is the ₩20,000 listing price — crossed for good on Jul 8.
Net assets (AUM, left) — hatched = derived, no print Fund price (right) — solid daily · dashed sparse-anchor ◎ outlined dot = cited / derived-from-cited anchor · small dot = bridge estimate
What happened after Jun 12 — the series above ends where the prints do
The daily panel stops at its last dated print. The arc since, from cited sources: the SK-hynix leverage complex nearly doubled from ₩4.84T (Jun 10) to ₩9.15T (Jun 19) — much of it NAV appreciation feeding ever-bigger rebalancing trades — as the KOSPI peaked at 9,385.59 on Jun 19. The two flagship managers’ Samsung+SK funds topped out at ₩15.41T on Jun 25, then shed ₩3.1T in twelve sessions to ₩12.31T by Jul 7, including ₩2.69T of pure valuation loss. On the Jun 23 — when the KOSPI fell 9.99% — managers had to sell mechanically to hold their leverage ratio: Bloomberg Intelligence estimates ~$6.0B of Samsung and SK hynix sold in a single session, ~14% of their combined trading value. VKOSPI closed 96.94 on Jun 29, the highest since the global financial crisis; Jul 13 closed −8.95% and tripped a Level-1 circuit breaker. This fund did exactly what a 2× daily-reset product does in a violent tape: it amplified the melt-up, then amplified the melt-down, and the negative-compounding tax did the rest — every one of the fourteen now trades below its ₩20,000 listing price, and retail bought ₩1.58T more of the SK pair during Jul 1–8 anyway. Meanwhile the venue multiplied: SK hynix listed on Nasdaq as SKHY on Jul 10, six US leveraged/inverse SKHY funds went live Jul 13–14 (SKHX/SKHZ, SKHU, SKUU/SKDD, Kurv; Direxion’s SKHL filed) — and the ADR immediately traded like the Korean funds: −9% Mon · +27.3% Tue · −9% Wed · −7.6% Thu pre-market (Jul 13–16, Benzinga). Then the ending, Jul 14–16 [cited]: Bloomberg’s verdict — “Leveraged Chip Bets Backfire in Korea” — with the biggest fund at ~$3.4B, −45% since its May 27 debut and >60% off its June peak; and on Jul 16 the regulators ended the experiment’s expansion: new single-stock leveraged listings halted, minimum deposit tripled to ₩30M and now all-cash (previously 70% coverable with securities), advertising banned, minimum order raised 1→20 shares — after 24 emergency trading halts in nine weeks. Scale of what they were containing: the 16 funds went ₩4.4T (May 27) → >₩15T in a month, daily turnover past ₩18T, 92% retail (FSS Gov. Lee Chan-jin) — products the government itself fast-tracked in ~4 months to pull retail money home from US equities and relieve the won. The Korean experiment was exported to New York the same week Seoul shut it down at home (§03·C, §04·F).

Real anchors: Korea's first single-stock 2× leveraged ETFs listed May 27, 2026; 16 single-stock leverage/inverse products debuted that day with combined first-week net assets ≈ ₩6.75T ($4.89B) and ₩48.7T cumulative trading value. KODEX SK hynix Single-Stock Leverage led on assets at ₩2.227T ($1.61B) — ~₩850B ahead of TIGER SK hynix (₩1.370T, $993M); KODEX debut-day net assets were ~₩1.154T, fee 0.29%. Debut-week return vs NAV ≈ +29% (KODEX SK hynix +29.13%; SOL +29.21% led). Cumulative net retail buying of KODEX SK hynix reached ₩2.0425T (May 27–Jun 5), seven straight sessions of net buying; TIGER SK hynix ₩1.97–2.08T. Then a two-day rout: June 4–5 the SK hynix 2× products fell ~20% then more, ~26% over two days (KODEX −26.16% from ₩29,055→₩21,455; TIGER −26.38%; 1Q futures −27.90%), on SK hynix ~−14%, "all holders since launch now in losses." Regulators (FSC/FSS) flagged negative-compounding and rebalancing risk and deem these short-term-speculation only; concentration in Samsung+SK hynix (a large share of the KOSPI) is itself seen as amplifying index volatility. AUM figures move daily with both flows and price; the since-launch daily series here is calibrated to these reported prints. Chart extension (added Jul 16): the post–Jun 12 dashed segment is anchored to ₩19,635 (Jul 8 close, first below listing — KRX via BigGo), ≈₩22,280 (Jul 7, back-derived from Jul 8's cited −11.86% day), ≈₩11.0k (Jul 13–14, derived from Bloomberg's −45%-since-listing), and AUM points ≈₩3.7T (Jun 19: SK-complex ₩9.15T × KODEX's ~40% complex share), ≈₩4.7T (Jul 7: SK-pair ₩7.58T × KODEX's ~62% pair share) and ≈₩5.1T (Jul 14: $3.4B, Bloomberg, at ~₩1,495/$); bridge points between anchors follow the documented tape (Jun 19 KOSPI peak, Jun 23 −9.99% day, Jul 13 crash) and are drawn as estimates, not data. Then the whipsaw continued: June 8 the KOSPI plunged 8%+ intraday (the 16 single-stock leveraged/inverse ETFs lost 10.3% of NAV in a session) before June 12 ripped ~10% higher to 8,430 — a 2× product round-trips ~±25% on such days, which is the negative-compounding grinder in action. Won→USD ≈ ₩1,383/$ at report. Jul 14–16 anchors [cited]: KODEX SK hynix ~$3.4B AUM, −45% since debut, >60% off June peak (Bloomberg, Jul 14); brokerage margin balance ₩36.3T vs ₩20.9T a year ago (Benzinga, Jul 16 — a shade under KOFIA’s ₩37.33T Jun 30 print; different date/basis, both shown); EWY −21.7% in four weeks — worst since Mar 2020 — after +330% from its Mar-2025 low (Benzinga); Jul 16 FSC/FSS package: listings halt · ₩30M all-cash deposit · ad ban · 20-share minimum order. Sources: Korea Exchange, Seoul Economic Daily, Bloomingbit, BigGo, KED Global, TradingKey, Bloomberg, Yahoo Finance, Benzinga, Korea JoongAng Daily. Leveraged single-stock ETFs are high-risk, short-term instruments subject to severe loss and negative compounding — this is not investment advice.

04The Companies◷ Jun 24-27 2026
04

The Companies

● live index

The investable universe, anchored to the Roundhill Memory ETF (DRAM) — the first pure-play memory fund (launched April 2, 2026), holding only companies that draw 50%+ of revenue from memory. Its roster is the spine of this directory.

Memory Index · CBOE: DRAM
$52.56
▲ +79% since inception (Apr 2, 2026)
Roundhill Memory ETF · active · 9–16 holdings
delayed / approximate · holdings subject to change
Holdings — approx. weight (incl. swap exposure)
SK Hynix 000660.KS
~24%
Samsung 005930.KS
~24%
Micron MU
~24%
Kioxia 285A.T
~6%
SanDisk SNDK
~5%
Seagate STX
~5%
Western Digital WDC
~5%

The integrated giants — the only three firms that span HBM, DRAM and NAND, and together ~73% of the DRAM ETF:

SK Hynix 000660.KS
The Incumbent · Seoul
HBMDRAMNAND
HBM leader (~53–62%) and Nvidia's primary supplier. Owns Solidigm in enterprise SSD. Largest HBM4 allocation; secured its entire 2026 capacity.
HBM share~53–62%
NAND share#2 · ~22%
Samsung 005930.KS
The Comeback · Suwon
HBMDRAMNAND
The only true full-stack memory maker. "Back" on HBM4 with Nvidia qualification; #1 in NAND. Betting on a 1c-nm HBM4 process.
HBM share~30–35%
NAND share#1 · ~28%
Micron MU
The Specialist · Idaho
HBMDRAMNAND
"More than sold out." Exited consumer memory to chase AI data centers. In volume HBM4 production (36GB 12-Hi) for Vera Rubin as of GTC 2026.
HBM share~11–21%
Marginsrecord · >50%

The rest of the roster — NAND specialists, the hard-drive oligopoly, and China's challengers:

Kioxia285A.T
NAND
#3 NAND (~16%). Ex-Toshiba Memory; IPO'd Tokyo Dec 2024. Co-runs the world's largest flash fabs with SanDisk; BiCS10 300+ layer NAND in 2026.
DRAM ETF weight · ~6%
SanDiskSNDK
NAND
Spun off from Western Digital (Feb 2025) as a NAND pure-play; +559% in 2025, joined the S&P 500. Enterprise-SSD focus via the Kioxia JV.
DRAM ETF weight · ~5%
SeagateSTX
HDD
HAMR/Mozaic density leader, shipping 44TB drives, eyeing 100TB-class. Co-developing NVMe HDDs with Nvidia. Nearline sold out for 2026.
DRAM ETF weight · ~5%
Western DigitalWDC
HDD
HDD volume leader after the SanDisk split. ~89% of revenue now from cloud. UltraSMR + HAMR roadmap to 100TB by 2029. Sold out 2026.
DRAM ETF weight · ~5%
Toshiba6502.T
HDD
The third hard-drive maker. Conservative MAMR/FC-MAMR scaling (28–34TB) with HAMR test vehicles in 2026–27. Channel checks show similar backlog.
Not in DRAM ETF
CXMTprivate · IPO filed
DRAMHBM*
China's DRAM champion. ~$8B 2025 revenue (+130%), ~12% of global DRAM wafer capacity. HBM3 samples to Huawei; competitive HBM is a 2027–28 story.
Not in DRAM ETF · Shanghai STAR listing pending
YMTCprivate
NANDDRAM*
China's NAND leader (~13% shipments, nearing Micron). New Wuhan fab ramping H2 2026 could make it #3 NAND globally; pushing into DRAM/HBM.
Not in DRAM ETF · IPO planned H2 2026
Nanya2408.TW
DRAM
Taiwan's DRAM maker — specialty & commodity DDR. A smaller player riding the same pricing wave; not an HBM force.
Not in DRAM ETF

By the numbers — latest-quarter net income, valuation and growth. The memory makers trade at strikingly low multiples for their growth (single-digit forward P/Es, PEG ratios near zero) — the market still pricing them as cyclicals even as AI demand drives record profits. ● live

Company Latest Q net income Fwd P/E Fwd PEG Annual rev growth
Net income is most-recent reported quarter (GAAP unless noted). Forward P/E and PEG from consensus estimates; PEG colored green (<0.5) · amber (0.5–1) · grey (>1). Korean/Japanese figures converted to USD at recent rates. Approximate, late May 2026 — not investment advice.
04·BAnalyst Upgrades & Downgrades◷ live feed · Aug 15 2026 baseline
04·B

Analyst Upgrades & Downgrades

● live · since Jun 2025

The most recent Wall Street rating actions on the memory & storage names — upgrades, downgrades, initiations and price-target moves since June 2025, newest first. (This was labelled "trailing twelve months"; the two oldest entries are now ~14 months old, so the window is stated from the data instead of assumed.) This list attempts to refresh live; opened as a plain file it shows the compiled baseline (refreshed Aug 15 2026, post-SK-hynix-Q2 and post-Samsung-Q2). Read the block below before the list — the important development is not any single action but the collapse of agreement between them.

Jul 29 – Aug 11 2026 — the consensus came apart, and almost none of it was a downgrade
Six weeks ago this section described "the densest target-raise wave of the cycle." That wave has reversed — but reading it as a turn to bearishness would be wrong in a specific and checkable way. The cuts are real and fast. On SK hynix, Mirae Asset took its target from ₩4.2M to ₩2.8M and Shinhan from ₩4.2M to ₩2.7M — cuts of a third — with NH, Daishin, Samsung Securities and Kiwoom all following. The revision cadence is itself the signal: Kiwoom cut SK hynix from ₩2.6M to ₩2.2M on Jul 30 and again to ₩2.1M on Aug 11, and cut Samsung twice in five weeks. The stated reasons are consistent across firms — memory profits peaking this year rather than 2028, LTA-driven capacity arriving in 2027, high prices making handset makers cautious, weaker PC and laptop demand, and Chinese competition, with CXMT into PC and server and YMTC into mobile and client SSD. And on the same days, Korea Investment & Securities raised SK hynix from ₩3.8M to ₩4.7M — up 23.7% — and KB Securities held Samsung at ₩600K with Buy, forecasting a fourth consecutive record quarter at ₩112T operating profit, DRAM margins of 83%, a shortage lasting at least three years, and a shareholder-return programme of ₩100–200T a year. Shinhan cut its 2026–27 profit forecast and Korea Investment raised its 2026–27 profit forecast, from the same quarter's disclosures. Two things make this more than a disagreement about direction. First, these are price-target cuts, not downgrades — the reporting is explicit that most firms maintained Buy, and this section's own filter will show almost nothing in the ▼ Downgrades bucket. Second, every surviving target sits above the share price. SK hynix traded at ₩1.425M on Aug 11; the most bearish target on the board still implies +47%, and the most bullish implies +230%. A "bearish" analyst here is one who thinks the stock only doubles. The cleanest way to read it: the sell side has not turned against memory. It has stopped agreeing about commodity memory while remaining uniformly bullish on HBM — Kiwoom cut its target while simultaneously forecasting Samsung's 2027 HBM shipments +109% with blended ASP +81%. The dispersion is the finding; the direction is not.
SK hynix price targets — where seven brokers moved, and how far apart they ended
Each line runs from the old target (hollow ring) to the new one (filled dot), sorted low to high. Six cut. One raised. The shaded band at the left is everything below the current share price — no target lands in it. The right-hand column is each target's implied return from spot, which is the number that makes this chart worth drawing: after a wave of cuts described in the press as brokerages turning cautious, the lowest surviving target still implies +47%. Hover any row for the firm's stated reasoning.
Samsung Electronics — the same split, narrower
Four cuts and one firm holding at ₩600K with Buy. The high is 1.71× the low, against 2.24× on SK hynix — the disagreement is real on both names but roughly twice as wide on the one with more HBM leverage.
Where the sources disagree with themselves, and with each other
(1) A headline that understates its own reporting. The Aug 11 piece is headlined "SK hynix Targets Diverge by 2 Million Won" and states a range of ₩2.7M to ₩4.7M. But the same article reports Kiwoom's ₩2.1M four paragraphs earlier. Using the article's own figures the range is ₩2.1M–₩4.7M — a ₩2.6M spread, not ₩2.0M. This board plots the ₩2.6M spread and shows the ₩2.1M point, because excluding it requires a reason no source gives. (2) The peak price. One outlet puts Samsung's June intraday high at ₩374,000, another at ₩374,500; §03·C carries ₩374,500 and that is used here for consistency. A ₩500 difference changes nothing material and is noted only so the two are not silently reconciled. (3) The US consensus, which three trackers cannot agree on. Aggregators publishing on the same day give Micron consensus targets of roughly $1,502, $1,260 and $1,261 — a spread of about 19% in what is nominally the same average of the same analysts. None of those three figures is used anywhere on this board, and no dated US rating action has been added in this pass, because the aggregator snippets available carry no action dates and mix stale targets with current ones. The Korean actions above are dated to the day and sourced to named analysts; the US consensus figures are not, and are printed here only as an illustration that "consensus" is itself a contested number.
Filter

"PT" = 12-month price target. Ratings reflect each firm's own scale (Buy/Overweight/Outperform ≈ bullish; Hold/Neutral/Equal-Weight ≈ neutral; Sell/Underweight ≈ bearish). Korean & Japanese targets shown in local currency. This is a sampling of notable actions, not every rating; figures are approximate and time-stamped to the action date. Context: Micron's FQ3 print (Jun 24 — record $41.5B revenue, $25.11 EPS, ~$50B next-quarter guide, $22B in customer commitments) set off the densest target-raise wave of the cycle, with the Street high running from ~$1,100 to $2,200 (Melius) in days and SanDisk's high hitting $3,000 (Bernstein). Note the split character: several are price-target raises on maintained ratings rather than upgrades, and Goldman pointedly lifted its target to $1,100 while keeping Neutral — stronger fundamentals, but wary the rally already prices them in. On the bear side, Michael Burry (Scion) disclosed a Micron short and Morningstar flagged 20–30% downside risk (see §04·D·S). Not investment advice. Aug 17 2026 — the estimate cuts recorded here have a direct valuation consequence [§04·L]. Kiwoom's Aug 11 reductions (SK hynix 2028 −22.2%, Samsung 2028 −26.9%) landed six days before SK hynix printed a 6.88× forward multiple. Applying them raises the multiple to 8.8–9.4× with no share-price move — a +29% to +37% re-rating from the denominator alone.

04·CMicron Options Activity◷ live feed · EWY to Jul 31 close
04·C

Micron Options Activity

● LIVE

Micron (MU) is the most liquid pure-play way to trade the memory thesis, and its options market is where positioning and fear show up first. This reads the key sentiment gauges — put/call ratios, implied volatility, the expected move, and unusual activity. Options data is intensely time-sensitive (chains reprice every second), so this is a dated snapshot that rechecks for the latest on load; opened as a plain file it shows the last compiled values. Every metric is explained — and options are a high-risk instrument, not a recommendation to trade.

$1,130 −~10% Jul 1–2
MU · snapshot Jul 15 2026
Post-FQ3 whipsaw: $2,000+ Street targets vs the Burry short; 30-day IV ~108% around the $600–1000 strikes.

Implied Expected Move — what the options are pricing

The ± range the market expects over the next ~30 days, derived from at-the-money implied volatility (≈ IV × √(days/365) × price).
Korea ETF (EWY) Options — 12-Month Positioning ◷ snapshot

Zooming out from Micron to the whole Korean trade: the iShares MSCI South Korea ETF (EWY) is the most liquid US-listed way to play (or hedge) Korea's memory-led rally. Tracking its price against total call vs. put open interest reveals positioning — and the tell is clear: put open interest has overtaken call open interest even as the ETF ran to records, i.e. investors increasingly paying for downside protection. Then the unwind those hedges were paying for arrived, and kept going: EWY fell ~14% in early June (≈$204 → ≈$176) and then kept sliding to $144.21 by Jul 29–30 — roughly −35% from its $220.88 52-week high, four straight down days into Korea's back-to-back circuit breakers (§03·C). The positioning signal was vindicated: put open interest crossed above call OI in Dec 2025, months before the peak, and the ETF has since round-tripped a year of gains. Aug 1 update — a violent bounce, into even heavier hedging. EWY closed $157.10 on Jul 31, up from the $144.21 low around the Jul 28–29 circuit breakers (§03·C) but still ~29% below its $220.89 high, with a single-session range of $156.42–$166.04 — the tape is not calm, it is merely off the floor. And the option market is more defensive than before, not less: put volume of 112,716 against call volume of 31,663 — a put/call volume ratio of 3.56, far above the ~2.2 open-interest ratio, with implied volatility at 84.6% in the 96th percentile of its own history. Traders bought the bounce and hedged it harder. The concentration is why it moved so hard — SK hynix and Samsung alone are ~52% of the fund, and both roughly halved from their June highs (SK hynix ₩2.99M → ₩1.32M; Samsung ₩374K → ₩207K).

EWY last price (right axis) vs. total call and put open interest in contracts (left axis), trailing ~12 months. Open interest = all outstanding option contracts, a gauge of positioning. Hover for values.
EWY last price (right) Total call open interest (left) Total put open interest (left)

EWY options data is a dated snapshot, not live exchange data. Anchors: EWY ≈ $178.45 at the June 11 close (≈$175 on June 7), down from a ~$204 prior close and a 52-week high near $218 (a ~14% multi-day pullback); 52-week range ≈ $66–$218; P/E ≈ 10; AUM ≈ $20.6B. The positioning signal — total put open interest exceeding total call open interest while the ETF sat near record highs — mirrors the Bloomberg "US-Listed Korea ETF Options Show Caution" reading (data as of 6/4/26): investors favoring puts over calls is a hedging/cautious posture, consistent with the overheating signals in the Korea leverage panel above. The 12-month series here are calibrated to that picture (OI rising broadly with the rally, puts inflecting above calls in 2026, price spiking to ~$218 then dropping to ~$176); exact daily OI prints vary by source and by the second. Open interest reflects positioning, not direction certainty — high put OI can be outright bearish bets or protective hedges on long stock. Jul 30 2026 update [cited]: EWY closed $144.21 (Jul 29, −4.81% on the day, a fourth consecutive decline; Jul 30 range $142.64–$154.06), against a 52-week range of $70.36–$220.88 — about −35% from the high and now below where the section's original $176 anchor sat. Drivers are single-name and mechanical, not diversified: SK hynix + Samsung are ~52% of the portfolio; SK hynix fell from its ₩2,987,000 Jun 25 record to ₩1,322,000 (Jul 30, −5.64%) and Samsung from ~₩374,000 to ₩207,000, through the Jul 28–29 back-to-back circuit breakers (§03·C). Options remain expensive and hedge-heavy: at-the-money implied volatility was ~76% with ~896K contracts outstanding on the late-May chain (OptionsAnalysisSuite) — sell-side has moved to downgrades (Seeking Alpha, "more downside is likely"). Basis conflict, flagged not merged: that ~896K total-contracts print sits below this series' ~1.16M call+put for the same month — different bases (single-snapshot chain vs all-expiry aggregate) and differing vendor definitions; both are shown rather than reconciled. The Jul '26 point: price is cited; the call/put OI values continue this series' calibrated construction (rising OI with puts extending their lead) and are modeled, not a vendor print — the honest read is the shape (puts above calls, both rising), not the exact contract counts. Aug 1 2026 update [cited]: EWY closed $157.10 on Jul 31 (prior close $161.21; session range $156.42–$166.04), a sharp recovery from the ~$144.21 print around the Jul 28–29 back-to-back circuit breakers but still ~29% below the $220.89 52-week high (52-week low $70.36). Positioning got more defensive through the bounce, not less: put volume 112,716 vs call volume 31,663 — a put/call volume ratio of 3.56 — while at-the-money implied volatility reads 84.59%, in the 96th percentile of EWY's own distribution (AlphaQuery, Jun 30 basis). Two ratios, two meanings — do not conflate them: the chart plots put/call open interest (~2.2×, a stock of positions built over months); the 3.56 figure is volume (a single session's flow, far more reactive). A volume ratio well above the OI ratio says fresh hedging is being layered on top of existing protection, which is exactly what a bounce nobody trusts looks like. The July point is the month-end close ($157), replacing the mid-month $144 reading used in the prior build. Sources: Bloomberg (reference chart), Investing.com, Barchart, OptionCharts, OptionsAnalysisSuite, AlphaQuery, stockinvest.us, Seeking Alpha, Invezz, public.com. Options involve substantial risk of loss and are not suitable for all investors. Nothing here is investment advice.

Options metrics are dated snapshots, not live exchange data — every value reprices continuously during market hours and is stale within minutes. Anchors (refreshed Jul 6 2026, vintages disclosed per figure — spot itself is printing inconsistently across delayed feeds amid the post-FQ3 whipsaw, ~$940–$1,130): session put/call volume1.00 on ~720K contracts (≈360K puts vs 361K calls in a single selloff session, early Jul) — up from ~0.91 in mid-June, a visible hedging bid arriving with the Jul 1–2 ~10% drawdown; put/call open-interest1.32 (Aug-21 monthly chain, May-vintage OI); 30-day mean IV ≈ 108% (Jun 23) versus ~70% pre-earnings in January — a full regime shift, with Rule-16 implying ~5.6% daily moves on the Aug chain. Chain geometry tells its own story: max pain ≈ $660, peak call OI at the $1,000 strike, peak put OI at $300 — strikes set before the June melt-up, so the chain’s center of gravity sits far below spot; dealer positioning simply hasn’t rolled up with a stock that ran ~380% off its $201 Nov-2025 low. Ratios and IV vary by expiration and by the second; the figures here illustrate the method and current regime, not a tradeable quote. Sources: Barchart, Fintel, Macroaxis, Stock Options Channel, AlphaQuery, CBOE-style conventions. Options involve substantial risk of loss and are not suitable for all investors. Nothing here is investment advice.

04·DAnalyst Price Targets◷ Jul 24 2026
04·D

Analyst Price Targets

◷ snapshot

Where Wall Street and Seoul see these six memory names heading over the next 12 months. Each card shows the current price, the consensus and the high/low target range, the implied upside, and the specific recent calls behind it. A live caveat up front: targets are revised constantly in a fast-moving cycle, and for stocks that have multiplied this year the aggregators disagree wildly — so the most recent named-firm targets matter far more than blended "consensus" figures that lag. Korean targets are in won; everything is a dated snapshot.

Price targets are analyst opinions with a ~12-month horizon, revised constantly — a dated snapshot (refreshed Jul 24 2026), not live quotes. The story since the Jul 6 refresh: a sharp mid-July memory/chip selloff pulled most of these stocks down even as analysts kept raising targets — so the implied upside widened rather than closed. Two catalysts are days away and are not yet in these figures: SK hynix reports Q2 on Jul 29 and Samsung on Jul 30 (Samsung's preliminary Q2: revenue ~₩171T, ~52% operating margin). Refresh note: an earlier build expected SK hynix on Jul 23; the confirmed date is Jul 29. The standing honesty point holds — aggregator "consensus" is badly inconsistent for names that have multiplied: SK hynix targets alone span ₩1.2M to ₩5.3M across 37 analysts. Anchors (current price · notable recent calls): Micron ~$988 (Jul 24 — about 22% off its July high on the chip rout, with a new Anthropic memory-supply agreement in the mix); consensus $1,492 (45 analysts, Strong Buy), up from ~$1,338; Street high $2,200 (Melius / Ben Reitzes), Cantor $2,000, UBS $1,625 (from $535), Phillip $1,870; low $385 (Citigroup). Michael Burry's disclosed short and Morningstar's 20–30% give-back warning remain the standing bear case (§04·D·S). SK hynix ₩1,322,000 (Jul 30, −5.64%; corrected — an earlier build wrongly carried ₩2.9M, which was the Jun 25 all-time high, not the live price). Q2 (Jul 29) set records — revenue ₩79.3T, operating profit ₩60.5T at a 76% margin — but missed consensus and the stock fell 9.6%, helping trigger the Jul 28–29 crash (§03·C); Nomura raised to ₩4.7M (from ₩4.0M), consensus ₩3.41M (37 analysts, now Strong Buy), Korea Investment ₩3.8M; full range ₩1.2M–5.3M. Samsung ₩207,000 (Jul 30) — Q2 operating profit ₩89.49T, +1,814% YoY, a clear beat that lifted the stock only 0.7% on the day; KB Securities ₩600K (Jul 23, Buy), consensus ~₩493K, range ₩210K–850K; full Q2 results Jul 30. SanDisk ~$1,610 (Jul 23 — down ~33% into a bear market); BofA $2,500 (Jul 1) and Goldman $2,200 (Jul 5, from $1,200, on structural NAND tightness into 2027), consensus $2,188 (23 analysts, Buy); FQ4 due Aug 5. Seagate ~$913 (Jul 24); Melius $1,600, Cantor $1,300 (from $1,000), Citi $1,240; consensus $1,009 (22 analysts, Strong Buy); Q3 beat — EPS $4.10, revenue $3.1B (+44% YoY). Western Digital ~$524 (Jul 24 — down ~36% in 30 days from $746, a sector-wide rout rather than a fundamental crack; CY2026 HDD output sold out with contracts into 2028); Melius $1,050, Cantor $900 (from $660), consensus $897 (up from $685); Buy. A stock can trade well below a rising consensus (as most of these now do) when the tape sells off faster than analysts revise — the gap is information, not a promise; equally, MU spent much of June above consensus, which meant targets were mid-revision. Korean won converts at roughly ₩1,480/$. Sources: Benzinga, TipRanks, S&P Global / stockanalysis.com, MarketBeat, MarketScreener, TheStreet, Seoul Economic Daily, Nomura, KB Securities, Korea Investment & Securities, Mirae Asset, BofA, Goldman Sachs, Cantor Fitzgerald, Melius, Citi. Not investment advice. See also: §04·J tracks whether that consensus has actually moved, and §05·B·V is the same disagreement problem applied to HBM vendor share.

04·D·SMicron Short Interest & Sentiment — the Bull/Bear Tug-of-War◷ SI anchored · sentiment cited
04·D·S

Micron — Short Interest & Sentiment

A revealing internal contradiction in the memory trade: even as Micron surged ~790% in 52 weeks, short interest climbed steadily through 2026 to multi-year highs (~41.6M shares, 3.68% of shares out by late June). Meanwhile sell-side sentiment is overwhelmingly bullish (consensus Strong Buy, targets escalating toward $2,000) — yet insiders are net sellers — and as of Jul 2, Michael Burry (Scion) has disclosed a short position in Micron, amid a sharp ~10% two-day chip selloff. That's a genuine three-way tension: momentum bulls (targets now to $2,000) vs short sellers betting "too far, too fast," with management quietly trimming. This section maps all three. Honesty note: short-interest figures are anchored to NASDAQ-sourced reporting; the historical trajectory is interpolated between dated readings.

Short interest vs price — 2023 → 2026
Left axis: short interest (M shares, red) and % of float. Right axis: MU price (gray, log). ● = an anchored reading. The story the chart tells: shorts fell during the 2023–24 doldrums, then rose through 2025–26 into the parabolic move — bears pressing the valuation bet as the stock ran. Per the reference image's circled spike, SI hit its highest in years right at the top.
Sell-side sentiment — analyst ratings (41 analysts)

Consensus Strong Buy · avg PT ~$628 (stock trades well above it — targets mid-revision)

Escalating Street-high price targets (2026)

Provenance & the honesty line. Short interest [anchored — NASDAQ-sourced]: ~41.59M shares / 3.68% of shares outstanding (stockanalysis, Jun 27 2026); a reporting period rose 35.24M→37.55M = 3.34% of float, ~1 day to cover on ~55M avg daily volume (Benzinga); ~37.3M / 3.32% of float in late May, after +15.9% then +2.6% jumps in April, "near the highest levels seen in years… rising steadily through 2026" (Barchart/Yahoo, May 23 2026). Days-to-cover is low (~1) because MU is extremely liquid, so a classic squeeze is less likely than the raw share count suggests. Sentiment [cited]: consensus Strong Buy — of 41 analysts, 31 Strong Buy / 5 Moderate Buy / 5 Hold; average PT ~$628 (stock already above it); Street-high targets escalated through 2026: $1,100 → Wedbush $1,400 → KeyBanc $1,600 (from $600) → J.P. Morgan $1,600 → NZ/Bolton $1,650 → Susquehanna $2,000 (Jul 1, from $1,750) / Cantor $2,000 (Jun 29) / Phillip $1,870; consensus ~$965→~$1,486 (Barchart, TipRanks, Simply Wall St, TheStreet). Burry short [cited, Jul 2]: Scion Asset Management disclosed a bearish position (TipRanks/CNN Markets), coinciding with a ~10% two-session MU/SNDK drawdown on a broad chip rotation — the highest-profile bear yet against the name. Insider counterweight [cited]: insiders sold ~$112M more than they bought over 12 months; CEO Sanjay Mehrotra's direct holding fell from ~1.28M shares (Dec 2024) toward ~0.95M (Simply Wall St). Fundamentals backdrop (updated — FQ3 reported Jun 24 2026): a blowout — revenue $41.46B (up from $23.9B in FQ2 and $9.3B a year ago, ~4.5× YoY), non-GAAP EPS $25.11 on ~84.9% gross margin, all above the high end of guidance; the stock jumped ~14.6% on the print. DRAM was $31.3B (76%) and NAND $9.9B; data-center revenue hit $25B in the quarter (enterprise SSD $5B); operating cash flow was $25.4B and adjusted free cash flow a record $18.3B; the board raised the dividend 30%. FQ4 guidance is a record $50.0B revenue, ~86% gross margin, and $31.00 EPS. HBM3E and HBM4 are fully booked through calendar 2027 with demand into 2028. The bear case is now valuation / memory cyclicality / the explicitly guided "meaningful moderation in the rate of price increases" after FQ4 — not current results (which keep setting records). [FQ2 reference: $23.9B rev / $12.20 EPS.] Modeled [interpolated]: the short-interest and sentiment trajectories between dated readings — bi-monthly SI prints exist but a continuous public time series doesn't, so the curve shows shape through anchors, not every print. The price line is schematic to match the reference image's arc (the dashboard's live MU price lives in §04). On the reference image (Earnings Whispers): it shows a Sentiment line + Short Interest band; the circled right-edge spike to ~40M aligns with the ~41.6M NASDAQ figure. Sources: stockanalysis.com, Benzinga, Barchart/Yahoo Finance, TipRanks, Simply Wall St, NASDAQ short-interest data. Not investment advice.

04·EETF Asset-Gathering Race — DRAM vs the Fastest Launches Ever◷ milestones anchored · curves modeled
04·E

ETF Asset-Gathering Race

The memory super-cycle now has a pure-play vehicle: the Roundhill Memory ETF (DRAM), launched Apr 2 2026, holding Samsung / SK hynix / Micron (≥50%-of-revenue-from-memory rule). Its asset-gathering pace has been so extreme it belongs in the same conversation as the fastest ETF launches in history — the spot-crypto giants. This chart races AUM by days since launch, so funds that launched years apart can be compared on equal footing. Aug 1 update — DRAM just printed its first drawdown, and the fund it was racing has collapsed. DRAM peaked at $25.08B on Jul 27 (day 116) and fell to $21.96B by Jul 30−12.4% in three sessions, including a −8.89% single day on Jul 28, the day Korea printed its first-ever back-to-back circuit breakers (§03·C). DRAM holds Samsung and SK hynix; this is the same event measured through a US wrapper. Meanwhile IBIT has more than halved — from ~$99B in Oct 2025 to $46.5B on Jul 31 2026 (−53%) as Bitcoin fell from above $93K entering 2026 to a 21-month low of $57,800. So the race closed dramatically, for the wrong reason: DRAM at $22B is now 47% the size of IBIT while being 7.7× younger. The honest caveat still stands: the milestone points (●) are anchored to cited reporting; the connecting curves are modeled interpolations — daily AUM histories aren't fully public, so read the shape and ranking, not every point.

AUM vs days since launch — linear scale
Y-axis is AUM on a linear scale ($0→$100B); x-axis is calendar days since launch (the basis the cited milestones use). Read the trade-off deliberately: a linear axis shows absolute size honestly — IBIT's fall from ~$99B and the fact that DRAM is still a much smaller fund are both properly proportioned, where a log axis flattered every early ramp and made a $1B fund look comparable to a $50B one. The cost is that the first weeks are compressed into the baseline: DRAM's $0.05B→$1B→$6.5B climb, and every fund's race to the $10B line, are hard to separate down there. The record-book table below is the precise version of that early sprint. ● = a cited public milestone; lines between them are modeled. The record-book table below shows each fund's headline "days to $10B" as widely reported. DRAM (amber) is raced against the five fastest-to-$10B ETFs ever (Balchunas/Bloomberg) plus IBIT and ETHA explicitly. IBIT's full path is shown — to ~$100B by Oct 2025, then its first-ever monthly outflow in Nov 2025.
The ETF Omen stopped being a caveat and started being a data point
Earlier builds of this section carried a soft warning that niche thematic funds tend to launch near cycle peaks. That warning now has evidence attached, and it is the derivative products rather than DRAM itself. Within twelve weeks of DRAM's debut the leveraged wrappers arrived: RAM — the Roundhill T-REX 2X Long DRAM Daily Target ETF — launched Jun 24 2026 and traded roughly $385M of notional on day one, the largest first-day volume of any US-listed leveraged or inverse ETF ever recorded, beating a prior category record near $282M. A 2× inverse version (RAMZ) followed. Expense ratio on RAM is 1.50%, against 0.65% for DRAM itself. This board has already watched that exact sequence play out to its conclusion. §03·C and §04·F document Korea's single-stock leveraged ETFs: launched May 27 2026, category-record inflows, ₩14T of net retail buying, assets that kept rising while NAV halved — and then ~₩34T of retail margin equity destroyed with 320,000–460,000 accounts liquidated, all inside nine weeks. The parallel is not a prediction. RAM is a US-listed daily-reset product with different mechanics, a different investor base and a regulator that has not had to intervene. But record-setting appetite for 2× exposure to a theme, arriving after the underlying has already tripled, is the same signature — and §08·F now shows retail search interest in memory going from zero to a five-year high over the same months.
The record book — days from launch to $10B AUM

Provenance & the honesty line. Anchored milestones [cited]: DRAM (Roundhill Memory ETF, CBOE:DRAM, launched Apr 2 2026) — $1B in 10 trading days, $6.5B in ~36 days, $10B in 43 days, ~$20B by mid-June 2026 (~$17B early June); ~79% return since inception (cryptobriefing, US News, ETF.com, 24/7 Wall St, stockanalysis, Roundhill). IBIT (iShares Bitcoin, launched Jan 11 2024) — $10B in 34 days, ~$50B by ~11 months (Dec 2024), $70B in 341 days (5× faster than GLD's 1,691), then a climb to ~$99–100B by early Oct 2025 (aided by a ~160% BTC rally to ~$125K) before its first-ever monthly outflow in Nov 2025 (−$2.3B) pulled it back toward ~$82–88B (ETF.com, iShares 2025 trends, Benzinga/Bloomberg, cryptonews). IBIT's mid-curve points ($18–21B through 2024, the Oct-2025 peak, the Nov dip) are modeled through those cited monthly anchors. FBTC $10B in 54 days; ETHA (iShares Ethereum, launched ~Jul 2024) $10B in 251 days with a $5B→$10B sprint in 10 days; JEPQ (JPMorgan Nasdaq Equity Premium Income) $10B in 444 days (Eric Balchunas / Bloomberg "five fastest to $10B"; Cointelegraph, AInvest). Modeled [interpolated]: every point between the cited milestones — full daily AUM curves aren't public, so the lines are smooth interpolations through anchored points, not realized daily data. A genuine source conflict, surfaced not hidden: ETF.com called DRAM "the fastest ETF in history to reach $10B" (43 days), yet IBIT reached $10B in 34 days — so DRAM is not faster to $10B by that count. The most defensible reading: IBIT holds the outright $10B speed record (34d); DRAM is the fastest non-crypto / thematic equity ETF and among the fastest overall — the "fastest ever" phrasing likely reflects a category or re-measured basis. Apples-to-oranges caveat: crypto ETFs' AUM is inflated by underlying price rallies (IBIT rode BTC +160%), not just inflows; DRAM's holdings also rallied hard. AUM growth ≠ pure inflows. Aug 1 2026 update [cited]: DRAM — $23.34B (Jul 22), $25.08B (Jul 27, the peak), $21.96B with 23 holdings (Jul 30); the share price closed $47.77 on Jul 28, −8.89% on the day, recovering to $51.20 by Jul 31; standardized since-inception return through Jun 30 was +156.30% NAV / +161.51% market, and Roundhill cited a 179.84% total return when DRAM passed $20B. Top exposures as of Jun 30: Micron, Samsung, SK hynix, SanDisk, Kioxia. Note the provider spread: AUM prints for the same fund differed by ~$3B across sources inside one week — partly real (the fund did move that much) and partly methodology, so treat any single AUM figure as approximate. IBIT$46.52B (Jul 31 2026), down from $53.4B at end-Q1 on $3.3B of Q2 outflows, with a 10-consecutive-day July outflow streak totalling 35,980 BTC (~$2.24B); Bitcoin entered 2026 above $93,000 and printed a 21-month low of $57,800 before trading near $62,500. RAM / RAMZ [cited]: the Roundhill T-REX 2X Long DRAM Daily Target ETF launched Jun 24 2026 with ~$385M of day-one notional against a prior category record near $282M, gross expense ratio 1.50%; a 2× inverse (RAMZ) followed. What this section cannot tell you, and will not guess: whether DRAM's $3B decline from the peak was redemptions or simply its holdings falling. DRAM dropped 8.89% on Jul 28 alone, so most of the AUM move is mechanically price — daily creation/redemption data was not available at build time and none has been invented. That distinction is the entire difference between "investors left" and "the stocks fell," and it is unresolved here. The "ETF Omen": niche thematic funds have a historical tendency to launch near cycle peaks — and 2×/inverse wrappers arriving twelve weeks after the underlying is the sharper version of that signal. Sources: ETF.com, Bloomberg (Balchunas), cryptobriefing, US News, 24/7 Wall St, Cointelegraph, AInvest, stockanalysis, financecharts, TipRanks, Morningstar, Motley Fool, iShares fact sheet, cryptotimes, spotedcrypto, The Market Periodical, PRNewswire, Businesswire, REX Shares, Roundhill. Not investment advice.

04·FLeveraged ETF Assets — MUU · SNXX · ETHU · BITX◷ quote/filing AUM · curves modeled
04·F

Leveraged ETF Assets Over Time

The riskiest corner of the flows story just became one of the biggest: the global 2×-leveraged memory complex. What began as a pair of US single-stock funds is now a three-market phenomenon — the US pair (MUU, SNXX), Hong Kong’s CSOP pair on SK hynix and Samsung, Korea’s sixteen single-stock leveraged/inverse ETFs launched May 27, and now basket leverage (RAM, 2× the DRAM ETF, launched Jun 24). Global single-stock L&I assets doubled from <$30B to >$60B in six weeks (Goldman), the CSOP SK hynix 2× is the world’s largest single-stock leveraged ETF — $16.8B at the Jun 23 peak, ~$9B by Jul 16 after the unwind, still #1 — and the Korean products drive up to 60% of daily turnover in the underlyings — until Jul 16, when Korea halted new single-stock leveraged listings and tripled deposits to ₩30M all-cash (§03·D). On Jun 23 the tail wagged the dog: an FSS warning cascaded into Samsung −12.3% / SK −12.5% (worst day since 2008) and Micron −13%. Two of the long-window funds below are direct leveraged bets on names in this dashboard — MUU (2× Micron) and SNXX (2× SanDisk) — so their asset swings are a high-beta read on memory-trade conviction. The other two, BITX (2× Bitcoin) and ETHU (2× Ether), are the crypto comparison. This tracks AUM / net assets over a Jun 2023 → Jun 2026 window (BITX's launch onward), with monthly data points anchored to fund quotes/filings. Honesty note: current AUM values are from fund quotes/filings; the monthly path is anchored to dated observations with interpolation between — and leveraged-ETF NAV swings on both flows and the amplified price of the underlying.

⚠ These are daily-reset 2× leveraged products — built for short-term trading, not holding. Volatility drag means returns over more than a day can diverge sharply from 2× the underlying, and a >50% adverse single-day move in the underlying can wipe out the fund. AUM here reflects both investor flows and the leveraged mark-to-market of volatile underlyings, so a rising line can mean inflows, a rallying underlying, or both.
World's largest single-stock lev ETF
$16.8B→$3.3B
CSOP SK hynix 2× (7709): Jun 23 peak → Jul 16 read; ~$13B Jul 1 (Bloomberg) — still #1
Korea single-stock lev turnover
−94%
in one month from May 27 (92% retail) → ₩12.31T by Jul 7 · Jul 16: new listings halted
Global single-stock L&I
$30B→$60B+
doubled in six weeks to May 27 (Goldman)
Share of underlying turnover
up to 60%
Korean lev ETFs vs Samsung/SK volume (SA, Jun 30)
Tail wags dog — twice
−12% · −8.95%
Jun 23: FSS warning → Samsung/SK worst day since 2008, MU −13% · Jul 13: KOSPI circuit breaker
Aug 3 update — every fund on this page is now past its peak, and the Korean complex has been regulated into silence
The memory pair, off the top but not broken: MUU $5.90B (Jul 3) → $4.60B (Jul 30, −22%) — and note how it got there: MUU rose +37.4% on Jul 30 alone (from $19.63 to $26.97) as Micron ripped, so the AUM decline is redemptions fighting a violent price rally, not a collapse. Its one-year total return is still +2,782%. SNXX $3.90B → $2.02B (Jul 29, −48%), on a day it fell 8.7% with 147M shares traded against 146M outstanding — the entire float changed hands in a session. The Hong Kong flagship keeps unwinding: CSOP SK hynix 2× (7709) has gone $16.8B (Jun 23) → $9.0B (Jul 16) → $5.79B (Jul 30) → $3.30B (Aug 2) — an 80% destruction of the world's largest single-stock leveraged ETF in ten weeks. Korea is the real story, and it is now a regulatory one. The ₩30M all-cash deposit took effect Jul 31–Aug 5, and the complex went quiet almost overnight: turnover in the sixteen Samsung/SK leveraged and inverse funds fell to ₩1.2388T on Aug 3, −58.6% from ₩2.9907T on Jul 31, and −94% from the ₩12.45T peak. KODEX SK hynix leverage is more than 80% below its Jun 23 high; the Samsung equivalent ~75% below its Jun 3 high. The finance minister has publicly apologised (§03·C). The instrument did not fail — it worked exactly as designed, in both directions, and the regulator removed the buyers. And the crypto comparison finally inverted. BITX $787M and ETHU $738M — ETHU's one-year total return is −85.7%, its NAV $15.60 against a 52-week high of $188.73. For two years this section used crypto leverage as the cautionary comparison for memory leverage. Both are now down together, which is a less comfortable observation than either one alone: it points at leverage appetite as the common factor rather than anything specific to memory.
The memory leverage complex — AUM by week, Apr 3 → Aug 3 2026
Log scale, $1B→$70B. ● = a dated print (fund/press/Goldman); solid segments are log-interpolated weekly between prints; dashed segments are modeled or absent-print — the July tails are now cited prints, not models — 7709 $13B (Jul 1, Bloomberg) → ~$9B (Jul 16); MUU $5.72B (Jun 28) → $5.9B (Jul 3, its last print before the Jul 13 −26.65% NAV session — the line ends at the print, no post-selloff AUM is invented); SNXX $3.22B (Jul 16); Korean flagship basis ₩15.41T (Jun 25) → ₩12.31T (Jul 7). The flat global-L&I line after May 29 still means no later print, not "unchanged." Lines begin where the products or their first prints begin — that staircase of late entries is the story: most of this complex did not exist ten weeks ago. ▲/▼ mark the two defining events; RAM's Jun 24 launch is ticked (no AUM print yet — none is invented).
Read this carefully: the US funds have official issuer pages for structure, NAV, expenses and launch dates, but exact weekly AUM history is not consistently exposed in free public feeds, and Korean single-stock-linked ETF assets require KRX/issuer feeds for a fully audited series. Outlined dots are dated official prints; small dots on solid segments are interpolated weeks; dashed means modeled or absent-print — not an official weekly point. RAM (Roundhill T-REX 2× Long DRAM, launched Jun 24) still has no AUM print and none is invented. Source note: a user-supplied reference chart shows an MUU path near $2.2→$8.4B plus MUD and KORU series — outside this board’s cited set; flagged per the conflict convention, not merged. Jul 17 update: the modeled post–Jun 23 dashes are gone — replaced by dated prints (7709 $13B Jul 1 / ~$9B Jul 16; MUU $5.72B Jun 28 / $5.9B Jul 3; SNXX $3.22B Jul 16; Korean flagships ₩15.41T Jun 25 / ₩12.31T Jul 7) — and the Korea-16 series switches to the FnGuide flagship-manager basis (Samsung AM + Mirae, the bulk of the 16) from Jun 25. Correction (logged): this chart's legend previously mislabeled SNXX as a "Samsung 2×" fund — SNXX is Tradr's 2× SanDisk (SNDK); the table below always had it right.

Aug 3 2026 additions to all three tabs [cited]. SINGLE-STOCK tab: MUU $4.60B (Jul 30 close, stockanalysis/S&P Global — same page shows +37.39% on the day, prev close $19.63, 86.6M shares traded); SNXX $2.02B (Jul 29, 147.4M shares traded vs 145.6M outstanding); CSOP SK 2× $5.79B Jul 30 and $3.30B Aug 2 (HK$25.75B ÷ 7.8). KOREAN tab: no aggregate AUM print has been published since the ₩12.31T of Jul 7, so the Korea 16 line is held flat and dashed from that date rather than extrapolated — it is certainly overstated today, since KODEX SK hynix leverage alone is down >80% from its Jun 23 peak and the Samsung equivalent ~75% from Jun 3. What is cited for the period is turnover, not assets: ₩1.2388T on Aug 3, −58.6% from ₩2.9907T on Jul 31 and −94% from the ₩12.45T peak (Seoul Economic Daily, Bloomingbit). Turnover and AUM are different quantities and are not mixed on this chart. ALL + GLOBAL tab: Goldman's single-stock L&I aggregate still has no print after the >$60B of May 27, so it too stays dashed and flat — an absent print, not a flat market. One unresolved conflict, surfaced rather than smoothed: CSOP 7709's AUM fell 43% between Jul 30 and Aug 2 (HK$45.17B → HK$25.75B) while its quoted price rose from HK$25.36 to HK$42.52 over the same days. Those two facts can only both be true if there were enormous redemptions into a violent rally, or if the fund executed a share consolidation. This board could not confirm which from available sources, so the AUM figure is plotted (that is what the chart measures) and the price discrepancy is flagged here rather than reconciled by guessing. If it was a consolidation, the price series is not comparable across that date. Weekly-chart provenance. Anchors: global single-stock L&I <$30B at the start of April → >$60B by May 27 (Goldman Sachs / Chris Lucas via BigGo); CSOP SK hynix 2× ~$5.38B on May 7 (overtaking the 2× Tesla fund as world's largest, Seoul Economic Daily) → ~$16.8B by Jun 23 (ETF.com) — re-verified Jul 14 2026: Bloomberg still ranks 7709 the world's largest single-stock leveraged ETF at ~$13B on Jul 1, the pullback from Jun 23 tracking SK hynix's own drop (incl. −17% on a KIS profit-forecast cut), confirming one growing series rather than two conflated funds; CSOP Samsung 2× ~$1.65B mid-May (TradingKey — a single dated print, hence the flat dashed line); Korea's 16 funds $3B day one (May 27, ETF.com) → >$9B by Jun 23 (ETF.com, KuCoin; 92% retail); MUU ~$1.1B late Jan → $5.72B Jun 28 (ETF Database) → $5.9B Jul 3 peak; SNXX ~$2.7B late Feb → ~$3.9B mid-June → $3.22B Jul 16 (quote pages). July prints [added Jul 17]: 7709 ~$13B Jul 1 (Bloomberg via AI Weekly) → ~HK$71B ≈ $9.0B Jul 16 (issuer/quote pages); Korean flagship managers ₩15.41T Jun 25 → ₩12.31T Jul 7 (FnGuide via BigGo); KODEX SK hynix alone ~$3.4B Jul 14 (Bloomberg); Korea's 16 funds ₩4.4T May 27 → >₩15T in a month (Benzinga, Jul 16); Jul 16 FSC/FSS: new single-stock leveraged listings halted, ₩30M all-cash deposits. Event flags: May 26–27 — Micron and SK hynix cross $1T market cap within ten hours (Samsung crossed May 6) and Korea's 16 funds launch; Jun 22–23 — FSS governor's public criticism ("did little more than enrich securities firms"), then Samsung −12.31% / SK −12.47% / KOSPI −9.99% / MU −13%. Everything between anchors is interpolation; everything after Jun 23 without a print is dashed. Not investment advice.

The global complex — who's who
VehicleMarketExposureStatus · scale
MUU · Direxion 2× MUUS2× Micron$5.9B Jul 3 peak · −26.65% NAV Jul 13, no post-selloff print — the original memory-lev trade
SNXX · Tradr 2× SNDKUS2× SanDisk$3.22B (Jul 16 read) — fastest single-stock launch on record (Jan '26)
RAM · Roundhill T-REX 2× DRAMUS2× the memory basket (MU+Samsung+SK+SNDK)Launched Jun 24 — today's only US leveraged route to Samsung/SK
7709 · CSOP SK hynix 2×Hong Kong2× SK hynix$16.8B Jun 23 peak → ~$9B Jul 16 — still the world's largest single-stock lev ETF
7747 · CSOP Samsung 2×Hong Kong2× Samsung Electronics~$1.65B (mid-May print)
Korea 16 · KODEX, TIGER, et al.Korea2× long & −2× inverse, Samsung & SK₩4.4T→>₩15T in a month · ₩12.31T Jul 7 · all 14 below ₩20k listing · Jul 16: new listings banned, ₩30M all-cash deposit
T-Rex 2× SK hynix (REX)US2× the SK hynix ADRSuperseded by the wave: ADR listed Jul 10, then six US SKHY L&I funds went live Jul 13–14 (SKHX/SKHZ, SKHU, SKUU/SKDD, Kurv); Direxion's SKHL filed
Leverage Shares 2× MemoryUS2× the DRAM ETF (RAM competitor)SEC registration filed May '26
AUM / net assets — Jun 2023 → Jul 2026 (monthly)
Log scale ($10M→$10B) so funds spanning three orders of magnitude are all legible. ● = a cited observation (launch, milestone, or dated quote); monthly points between are interpolated. The divergence that defined this chart for two years has now closed: MUU (2× Micron) went parabolic to $5.9B on the memory super-cycle and has since given back to $4.60B; SNXX (2× SanDisk) rocketed from a Jan-2026 launch to $3.9B and is back to $2.02B; and the crypto pair BITX ($787M) and ETHU ($738M) round-tripped earlier and further. All four are now falling together — which says more about the price of leverage than about any of the four underlyings.
The four funds — what they are

Provenance & the honesty line. MUU — Direxion Daily MU Bull 2X (2× Micron), launched Oct 10 2024, 1.06% ER; ~$100M→~$400M by Nov 2025 (+$148M net flows in a month; ~378% YTD return), crossed $1.1B late Jan 2026 (one of ~7 single-stock ETFs over $1B), then went parabolic on Micron's rally to ~$5.9B by mid-June 2026 (ETF Database 1-yr net AUM change ≈ +$5.86B; NAV ~$972/share) — per Direxion/ETF Trends, ETF.com, ETF Database, AAII, SEC. SNXX — Tradr (AXS) 2× Long SNDK (2× SanDisk), launched Jan 26–27 2026, 1.49% ER; hit $650M in 24 days (fastest single-stock-ETF launch by avg AUM/day on record), ~$2.7B by late Feb, ~$3.9B by mid-June 2026 — per Tradr/PRNewswire, stockanalysis, Robinhood, Kraken. BITX — Volatility Shares 2× Bitcoin (first US leveraged crypto ETF), launched Jun 27 2023, ~2.38% ER; grew through the BTC bull market to a multi-billion peak (~$3B+ in late 2024/early 2025), then round-tripped — a ~76% one-year total-return loss and a peak-to-now AUM decline of ~⅔, to ~$1.0B by mid-2026 (1-yr net AUM change ≈ −$2.05B; NAV ~$12 vs 52-wk high ~$69) — per Yahoo, stockanalysis, ETF Database, Kraken. ETHU — Volatility Shares 2× Ether, launched Jun 4 2024, 2.67% ER; ran up to ~$1.4B in early 2025 then suffered a ~74% YTD drawdown to ~$730–790M by mid-2026 (NAV ~$14 vs 52-wk high ~$189) — per Yahoo, MutualFunds, Dividend.com. Monthly points [interpolated]: the per-month values between cited observations — full daily AUM histories aren't public, and leveraged-fund AUM moves on flows + the 2× mark of a volatile underlying, so monthly points show plausible shape through anchored dates, not realized daily data. A note on the source image: this section was requested from a chart whose dollar labels (e.g. MUU $19.2M) didn't match current filings — I rebuilt from current quote/filing AUM instead of tracing it. AUM ≠ inflows for leveraged funds even more than for spot ones: a doubling underlying doubles the leveraged exposure mechanically. Sources: ETF Trends/Direxion, Tradr/PRNewswire, Volatility Shares, Yahoo Finance, Robinhood, Kraken, stockanalysis, AAII, ETF Database, SEC filings. Not investment advice — and these instruments are especially unsuitable for non-traders.

04·IGross Margin by DRAM-ETF Holding — Quarterly since 2023◷ company filings · Aug 6 2026
04·I

Gross Margin by DRAM-ETF Holding — Quarterly since 2023

§04·E tracks how fast the Roundhill Memory ETF (DRAM) gathered assets. This tracks whether the companies inside it are actually making money on what they sell. Gross margin is the cleanest single read on a memory cycle — it strips out opex and one-offs and shows the raw spread between what a bit costs to make and what it sells for, which in this industry swings further than almost any other sector. Micron went from a −9.1% full-year gross margin in FY2023 to 84.6% in the quarter just reported — a ~94-point swing in three years. But the headline finding is the dispersion, not the level: in the same calendar quarter, the seven charted holdings span 52.7% to 84.6% — a 34.6-point spread inside one ETF — and the fund's single largest position has the lowest margin of the group for a reason that has nothing to do with memory.

Read this before comparing any two lines — they are not measured the same way
This chart puts seven companies on one axis. They do not report on a common basis, and pretending otherwise is the main way this comparison gets abused. (1) Samsung's number is whole-company. It includes smartphones, displays and foundry alongside memory, so its ~50% understates its memory economics enormously — in the same quarter its Device Solutions division ran a 52.2% operating margin, and Samsung does not publish divisional gross margin at all. Its line is the least comparable on the chart and it is 24% of the fund. (2) GAAP and non-GAAP are mixed. Micron and SK hynix are shown on reported (GAAP / K-IFRS) figures; Western Digital, Seagate, SanDisk and Kioxia headline non-GAAP gross margin, which excludes stock comp and amortisation and typically runs 1–3 points higher. (3) Five different fiscal calendars. Micron's year ends in late August, WDC/Seagate/SanDisk in early July, Kioxia in March, Samsung and SK hynix on the calendar. Every point here is placed on the calendar quarter it actually covers, which is the only way to line them up. (4) Two of these aren't memory makers. Seagate and Western Digital are hard-drive companies with a structurally different cost curve — they are in this fund because HDD is storage, not because they sell DRAM or NAND.
Gross margin %, by calendar quarter — 2023 → 2026 Q2
● marks a cited, reported figure. Dashed segments are interpolated between cited points and are not reported numbers — this is the board's usual convention and it matters more than usual here, because quarterly gross margin is not uniformly disclosed by every issuer in every period. Hover any point to see its basis and whether it is cited or interpolated. Three lines start late and that is the data, not an omission: SanDisk was spun out of Western Digital in Feb 2025 and Kioxia only listed in Dec 2024, so neither has a 2023 quarterly history to plot; Seagate and Kioxia appear as single points where only the latest quarter was cleanly citable. The zero line is drawn heavier because crossing it is the whole story of 2023 — Micron and SK hynix were both selling memory below cost.
What is in the fund but not on the chart
What the dispersion actually tells you about owning this ETF
A 34.6-point margin spread inside a single-theme ETF means the theme is not one trade. At the top, Micron (84.6%), Kioxia (80%) and SanDisk (~80%) are pure-play memory makers capturing the full pricing shock — Kioxia's average selling prices rose roughly 70% quarter-on-quarter, and SanDisk's gross margin went from 51.1% to 78.4% in two quarters. In the middle, SK hynix at 68% — high, but it books HBM through a cost structure that carries far more wafer input per bit (§05·D). At the bottom, Western Digital (54.4%) and Seagate (52.7%) are riding the same AI storage demand through a completely different product, and Samsung (~50%) is diluted by everything else it makes. Two practical readings follow. If the thesis is memory pricing, the fund gives you a lot of things that are not that: over a quarter of the stated weight is a Treasury bill, 3.4% is a company with no published accounts, and roughly 10% is HDD. A note on those percentages, because they do not add to 100: the fund reaches its equity exposure largely through total-return swaps, so the published weights are gross — the charted names come to 90.7%, the unchartable ones to 35.6%, and the sum is 126%. The Treasury bill is not a defensive allocation; it is collateral backing the swaps, which is why the fund can hold 26.6% in bills and still deliver full equity exposure. If the thesis is AI storage broadly, the spread is the point and the diversification is doing work. What the chart cannot tell you is which way it breaks. Margins this far above trend are, definitionally, the part of the cycle that mean-reverts — Micron's own record is the argument: it printed −9.1% three years ago. §05·D has TrendForce expecting the DRAM deficit to widen through 2027 before new capacity lands in 2028; §03·B has the export price still climbing. Those support the level holding a while longer. They do not make it permanent.

Provenance — which points are real, and the four ways this comparison can mislead. Holdings basis [cited]: Roundhill Memory ETF (DRAM) holdings as of Aug 5 2026, 23 lines, $23.92B net assets. Weights quoted here combine each company's direct equity and its total-return-swap exposure, because the fund holds several names both ways. These weights are gross and do not sum to 100% — the fund's own top-10 line is 110.57%, and charted (90.7%) plus unchartable (35.6%) comes to 126%. That is not an error in this table; it is how a swap-based fund reports, with the 26.6% Treasury bill serving as swap collateral rather than as a cash allocation — Micron ~26.8%, Samsung ~24.1%, SK hynix ~21.7%, Seagate 5.0%, Western Digital 5.0%, SanDisk 4.5%, Kioxia 3.6%, CXMT 3.4%. Cited margin prints (●): Micron FQ2'24 19%, FQ3'24 27%, FQ4'24 35.3%, FQ1'25 38%, FQ2'25 37%, FQ3'25 38%, FQ4'25 44.7%, FQ3'26 84.6% GAAP (some outlets report 84.9% non-GAAP for the same quarter — the difference is basis, not a dispute), with FY2023 full-year at −9.1%; Q4 FY26 was guided to ~86% and is not plotted because it has not been reported. SK hynix Q3'23 −9.5%, FY23 −1.6%, FY24 48.1%, FY25 60.4%, Q1'26 68.3%, Q2'26 68% (that quarter's operating margin was 76% — a different line item, not used here). Samsung FY23 30.3%, FY24 38.0%, FY25 39.4%, Q1'26 47.7%. SanDisk FQ2'26 51.1%, FQ3'26 78.4%, FQ4'26 ~80% (guided 79–81%). Kioxia Q1 FY26 80% (82% excluding JV items). Western Digital FQ4'26 54.4%. Seagate FQ4'26 52.7%. Interpolated points, stated plainly: where a company published an annual gross margin but not a quarterly one — most of SK hynix's 2024–25 path and nearly all of Samsung's — the quarterly points were interpolated to hit the cited annual figure and are drawn dashed and tagged in the tooltip. Micron's 2023 quarters and its 25Q4–26Q1 are likewise interpolated between cited anchors. Do not quote a dashed point as a reported number. Roughly a third of the plotted points are cited; the rest exist to give the lines a shape between them. The four distortions, repeated because they are load-bearing: (1) Samsung is whole-company, not memory — the single most misleading line on the chart, and the fund's largest position. (2) GAAP and non-GAAP are mixed — WDC, Seagate, SanDisk and Kioxia are non-GAAP, which runs structurally higher. (3) Five fiscal calendars, mapped onto the calendar quarter each period actually covers. (4) Seagate and Western Digital are HDD makers, not memory manufacturers. And 35.6% of the fund has no line at all — a 26.6% Treasury bill, unlisted CXMT at 3.4% with no published accounts, and 5.6% of small-caps excluded rather than approximated. Cross-refs: §04·E (the ETF's asset gathering and its first drawdown), §03·B (the export price driving these margins), §05·D (the supply-demand balance behind it), §04 (the companies). Sources: Roundhill/Finnhub holdings via stockanalysis, Micron 8-K and 10-Q filings, SK hynix newsroom Q2'26, Samsung Global Newsroom, SanDisk investor relations, Kioxia FY26 Q1 results via TrendForce, Western Digital and Seagate FQ4'26 releases. Not investment advice.

04·JEvolution of the Average Target — Consensus vs the Tape◷ FnGuide · MarketScreener · Aug 6 2026
04·J

Evolution of the Average Target — Consensus vs the Tape

§04·D lists who is calling what. This asks a harder question: has the consensus actually moved? Analyst price targets are supposed to be a forward view, but they are revised by humans with careers, and they tend to move after the market rather than before it. The memory complex just ran that experiment at full scale. Samsung and SK hynix both now trade at almost exactly half their average target — an implied +96% upside each — not because analysts got more bullish, but because the share prices halved while the targets barely moved. The revision data is the tell: Samsung had zero target-cut reports from March through June, SK hynix had none for roughly a year, and then six and eight brokerages respectively cut within days of the Jul 29 plunge. Across the Korean market, July produced 824 downgrade reports against 407 upgrades.

Analyst targets as a percentage of today's share price
Everything is normalised so the current share price = 100, because these four names are quoted in two currencies across three orders of magnitude and are not otherwise comparable. The shaded bar spans the lowest to highest analyst target; the filled dot is the average, labelled with its implied upside; the small dark dot on the 100 line is where the stock actually trades. Where the average has moved since the board's Jul 24 reading, a coloured connector shows the direction — red for a cut, teal for a raise — from a dashed marker at the old level. Read the bar widths first: a wide bar means the analysts covering that stock disagree profoundly about what it is worth, which is itself information.
The numbers behind the bars
The consensus did not lead this cycle — it followed, and it is still catching up
Seoul Economic Daily, citing FnGuide, documents the sequence precisely. Through the first half of 2026 brokerages held Samsung targets in the ₩500,000–600,000 band and kept raising them — and issued not one downgrade report between March and June. SK hynix went roughly a year without a single target cut. Then the stock plunged on Jul 29, and the revisions arrived all at once: six houses cut Samsung, eight cut SK hynix. Mirae Asset took Samsung from ₩500,000 to ₩370,000 on Jul 29. BNK took SK hynix from ₩1.85M to ₩1.48M and moved to Hold, arguing that "demand momentum has peaked out" while "companies are competitively pursuing capacity expansion." The result is a consensus that is internally incoherent. Among the thirteen houses publishing SK hynix targets in July, the highest was ₩4.7M and the lowest ₩1.48M — a ₩3.22M spread, where the low target sits below the current share price and the high is nearly triple it. Micron is wider still in ratio terms: $2,200 against $361, a 6.1× spread across 45 analysts looking at the same company on the same day. When dispersion is this large the average stops being a forecast and becomes an artefact — it is the midpoint of two incompatible theses about whether this is an early-innings supply shock or a late-cycle top, and averaging them describes neither. Cross-read it against §04·D for who holds which view, §04·I for what the margins are actually doing, and §05·D for the supply-demand data both camps are looking at.
Two holdings are missing from this chart, and one of them is a source conflict worth seeing
Western Digital is excluded because the aggregators do not agree on what its consensus is. Four readings surfaced in the same week: this board's own §04·D carried $897 at Jul 24; MarketBeat reports $541.24 across 29 analysts; a second aggregate gives $710.47; a third $497.20. That is a 1.8× spread between data providers describing the same consensus, which is a different and worse problem than analysts disagreeing — it means at least some of these are stale, differently-sampled, or split-adjusted inconsistently after a stock that has run more than 200% this year. Rather than pick one and present it as fact, the line is left off. Seagate is excluded for the simpler reason that no current consensus average could be cited — only individual calls (Wedbush to $1,000 from $825; Wells Fargo upgrading with $1,100). Kioxia, CXMT and the Taiwanese small-caps have no usable English-language consensus coverage at all. This chart therefore covers four of the fund's holdings, not all of them, and §04·I's table of what the ETF contains applies here too.

Provenance — what is cited, what is dated when, and four ways a consensus average misleads. [cited]: Micron — average target $1,507.38 against a $920.95 close, high $2,200, low $361, 45 analysts (MarketScreener consensus page, retrieved Aug 6 2026); prior average $1,492 at Jul 24 from this board's §04·D. SK hynix — brokerages' average target over the past month ₩3,371,538 against a ₩1,718,000 close on Jul 31 (up the daily limit), 13 houses publishing in July, high ₩4.7M (Korea Investment), low ₩1.48M (BNK, cut from ₩1.85M with a Hold); prior board reading ₩3.41M at Jul 24; MarketScreener separately shows ₩3,189,340 across 37 analysts on Aug 6 — a different sample and window, noted rather than blended. Samsung — average ₩514,545 against a ₩262,500 close on Jul 31 (+26.81% that day), 11 houses in July, high ₩650,000 (Korea Investment), low ₩360,000 (DB Securities), Mirae Asset cutting ₩500,000 → ₩370,000 on Jul 29. SanDisk — average $2,116.64 across 23 analysts (MarketBeat), against a Jul 24 price of $1,610 from §04·D — its spot reading is two weeks stale and its implied upside is therefore the least reliable figure on the chart; individual July moves include Goldman to $2,200, Evercore cutting to $2,800 from $3,100, Jefferies to $1,750 from $3,000, Mizuho to $1,900 from $2,200. Revision counts: FnGuide via Seoul Economic Daily, Aug 3 2026 — 407 upgrade reports against 824 downgrades across the Korean market in July. Four ways this metric misleads, stated because the chart is seductive: (1) The average is not a forecast. With a 6.1× spread on Micron and a ₩3.22M spread on SK hynix, the mean sits between two incompatible views and represents nobody's actual thesis. (2) Targets are anchored to price. Analysts revise toward the tape more often than the tape moves toward targets, so a large gap usually closes by the target falling, not the price rising — which is precisely what July showed. (3) The samples are not the same. The Korean figures are FnGuide's count of houses publishing in the past month; MarketScreener and MarketBeat use their own analyst panels over their own windows. Two of the numbers here for SK hynix (₩3.37M and ₩3.19M) differ for that reason alone and both are reported rather than averaged into a false single figure. (4) Spot dates differ by up to two weeks — Micron is Aug 6, the Korean pair Jul 31, SanDisk Jul 24 — and in a market that moved 20% in a session, that matters. Every spot date is on the chart's tooltip. Related: §04·D (individual calls), §04·I (realised gross margins), §04·E (the ETF), §03·C (the crash that triggered the revisions). Sources: MarketScreener consensus; Seoul Economic Daily / FnGuide, Aug 3 2026; MarketBeat; TipRanks/TheFly. Not investment advice.

04·KSOXL — Price & Assets, the 3× Semiconductor Trade◷ Jan 2025 → Aug 6 2026
04·K

SOXL — Price & Assets, the 3× Semiconductor Trade

SOXL is the oldest and largest leveraged bet on this entire thesis — the Direxion Daily Semiconductor Bull 3× ETF, listed since March 2010, seeking three times the daily move of the semiconductor index. Micron is its third-largest equity holding at 5.11%, behind AMD and Nvidia, so this board's memory cycle runs straight through it. From the April-2025 tariff trough of $12.29 to the June-2026 peak of $302.00 is a 24.6× move in fourteen months — and it now sits at $132.33, still up 977% from that low but 56% below the high. But the price is the less interesting axis. Assets are $23.82B, and they did not fall when the price did: July brought $6.9B of net inflows, the largest month in the fund's sixteen-year history, during a month in which it lost 57%. That divergence — the scissors on the chart below — is the whole section.

⚠ This is a daily-reset 3× leveraged product — built for intraday and short-horizon trading, not holding. Over any period longer than a day its return is path-dependent: it tracks 3× the daily move compounded, not 3× the period move. In choppy markets that difference is a permanent, mechanical loss even if the index goes nowhere.
Two crashes and a 24.6× melt-up — Jan 2025 → Aug 2026
Red line, left axis: share price. Teal area, right axis: assets under management. Two separate scales, deliberately, because the point is that they moved in opposite directions. Filled dots mark cited figures; the points between them are interpolated to connect dated anchors and are not published readings — the tooltip says which is which for every point. The dashed line is Jun 22 2026, the day the semiconductor sector peaked. In the six weeks after it, the price fell 63% while assets rose — because investors put money in faster than the fund lost it. That is not a chart of a fund shrinking in a crash. It is a chart of people buying one. Extending back to January 2025 shows this is the second time, not the first. The April-2025 tariff selloff took the fund from ~$28 to $12.29 — the SOX index fell roughly 33% over those months and the 3× wrapper fell more than 70% — and 2025 still finished +54.91% for the year. A positive annual return containing a 70% drawdown is the clearest possible statement of what a daily-reset product is. No 2025 AUM figure was citable, so that half of the teal line is modelled to the ">$12B entering 2026" anchor and every 2025 asset point is tagged as interpolated.
The record since January 2025, in order
The arithmetic of a daily-reset product, which is not intuitive and is not optional
A 3× fund does not deliver 3× the period return. It delivers 3× the daily return, compounded — and compounding is where the money goes. If the index rises 10% then falls 10%, it ends down 1%; the 3× fund rises 30% then falls 30% and ends down 9%. Variance drag scales with the square of leverage, so a 3× fund decays at roughly six times the rate of the unleveraged index against twice for a 1× fund. Nothing about that is a defect or a fee — it is what the product is. 2025 and 2026 each supplied a textbook illustration. In April 2025 the SOX index fell about 33% over several months and SOXL fell more than 70% — and the fund still closed 2025 up 54.91%, because the recovery came through a period of lower volatility than the crash. The same year produced both a 70% drawdown and a 55% gain; neither number describes the experience of holding it. From the Jun 22 peak, SOXL fell 63% while SOXX — the same sector, unleveraged — fell 25%. Three times 25% is 75%, so on that leg the fund actually lost less than naive multiplication implies; but the recovery is where it bites: a −63% drawdown needs a +170% rally simply to return to Jun 22, and if that rally arrives through volatile sessions rather than a straight line, holders need more still. The five-year scoreboard settles the argument about whether the leverage pays: SOXL returned +479% over five years against +404% for unleveraged SMH. Triple daily leverage, 0.75% a year in fees and a beta of 5.76 bought seventy-five percentage points over half a decade. On the way up it looked like genius — +165% in April alone, +433% year-to-date through May. The full cycle is the honest measure, and the full cycle is roughly a draw.
This board has now watched the same behaviour in three markets
Buying leveraged funds harder as they fall is the defining pattern of this cycle, and SOXL is the third and largest instance on this board. In Korea (§03·C, §04·F), retail put ₩14T net into single-stock leveraged ETFs while assets rose and NAV halved — until roughly ₩34T of margin equity was destroyed and the regulator imposed a ₩30M all-cash deposit that cut turnover 94%. In Hong Kong, CSOP's SK hynix 2× went from $16.8B to $3.30B in ten weeks. Now SOXL: −63% from the peak, +$6.9B of inflows in the worst month of its life, assets at an all-time high. The differences are real and worth stating. SOXL is 3× a diversified 49-holding index, not 2× a single stock; it is sixteen years old with a deep options market and a functioning creation/redemption mechanism; and no US regulator has intervened. The similarity is the flow behaviour, and it is the reason §04·E flagged the arrival of RAM and RAMZ — 2× and −2× wrappers on the memory basket — as a signal rather than a product launch. Scale, for calibration: SOXL's $23.82B is almost exactly the DRAM ETF's $23.92B (§04·E) and about half of IBIT's $46.5B. The leveraged semiconductor trade is now as large as the entire pure-play memory ETF it partly contains.

Provenance & the honesty line. [cited — live quote, Aug 6 2026]: close $132.33 (+0.20%; $135.70 after hours), assets $23.82B, 170.45M shares outstanding, volume 57,025,099, expense ratio 0.75%, 49 holdings, beta 5.76, inception Mar 11 2010, 52-week range $22.95–$302.00, one-year total return +439.84%, since-inception average annual return +39.16% (stockanalysis / S&P Global Market Intelligence). Holdings: roughly 52% of the fund sits in Treasury and government cash funds — Dreyfus, Goldman and JPMorgan vehicles — as collateral against an ICE Semiconductor Index swap; the equity sleeve is led by AMD 5.53%, NVIDIA 5.19%, Micron 5.11%, Broadcom 4.59% and Intel 3.35%. As with the DRAM ETF in §04·I, a large Treasury position in a leveraged fund is not a defensive allocation — it is swap collateral, and the fund is fully exposed despite it. [cited — 2025 path]: the year opened near $28 on Jan 2 2025 and was at ~$25 by Jun 30, a move that conceals a round trip of more than 70% in between; the biggest single-session gap-down was −23.0% on Apr 10 2025; the trough close was $12.29 on Apr 28 2025; the 52-week low of $22.95 falls in August 2025; and the full-year 2025 total return was +54.91%. SOXL itself did not reverse split — Direxion split the inverse fund SOXS and several others, not the bull fund — so the price series is continuous across the whole window and no split adjustment has been applied. [cited — 2026 path]: best month April at +165.0%; year-to-date +433% through May against ~81% for the unleveraged index; sector peak Jun 22; −23% in the single session of Jun 23 versus ~8% for unleveraged semi ETFs; ~$5.1B of inflows between Jun 22 and Jul 17; worst month July at −57.0% with $6.9B of net inflows, the largest monthly intake since launch; −63% from the Jun 22 peak by Jul 29 against −25% for SOXX; five-year return +479% versus +404% for SMH. What is interpolated, and it matters here: nine of the sixteen price points and only three of the sixteen AUM points are cited readings. No 2025 assets-under-management figure could be cited at all, so the entire 2025 half of the teal line is a modelled path drawn to meet the one anchor that exists — assets ">$12B" entering 2026 — and should be read as a shape, not a series. Extending the price line back to January 2025 also corrected the 2026 interpolation: anchoring to the cited +54.91% 2025 return puts the end of 2025 at ~$43, which moves the modelled May-2026 point from ~$277 in the first build down to ~$231, consistent with the cited +433% year-to-date. The monthly values between them are interpolated to connect dated anchors and are not published figures — daily AUM history for this fund is not public, and monthly price marks were reconstructed to be consistent with the cited month-return and year-to-date figures rather than taken from a series. Every point's status is in its tooltip. Do not quote an interpolated month. The shape — price down, assets up — rests entirely on the cited endpoints and is not an artefact of the interpolation. Three further caveats: (1) AUM is not flows. Assets move on both creations and the mark-to-market of what is already held; the $6.9B July figure is a separately cited net inflow number, which is why it can be stated alongside a 57% price fall without contradiction. (2) SOXL tracks a semiconductor index, not memory — Micron is 5.11% of it, so this is a broad AI-silicon instrument that happens to contain this board's thesis, not a proxy for it. (3) Past leveraged returns are especially unrepresentative because they are path-dependent; the +439.84% one-year figure describes one particular sequence of days and would differ materially for any other path with the same start and end. Cross-refs: §04·E (the DRAM ETF and its 2×/inverse wrappers), §04·F (the leveraged complex), §03·C (Korea's leveraged unwind), §04·I (holdings-level margins). Sources: stockanalysis SOXL, ETF.com daily flows, Benzinga, TechTimes, 24/7 Wall St, Direxion. Not investment advice — and leveraged daily-reset products are especially unsuitable for non-traders.

04·GMemory vs the Magnificent 7 — Forward PEG◷ Jul 14–22 prints · +AMD
04·G

Memory vs the Magnificent 7 — Forward PEG

Same AI boom, two valuation regimes. The PEG ratio (P/E ÷ expected earnings growth) asks one question: how much are you paying per point of growth? On that screen the seven memory names in §04 and the Magnificent 7 live on different planets — median memory PEG ≈ 0.20 vs Mag-7 ≈ 1.4, a ~7× gap — because the market prices memory growth as cyclical (it won't last) and platform growth as durable (it will). The one crossover is the Mag 7's own memory customer: Nvidia at 0.29 — the only platform name priced like a memory maker. AMD is added as the merchant-accelerator control at ~1.31 — the other big buyer of HBM, and the useful contrast to Nvidia: same AI end-market, same memory dependency, but priced 4–5× richer per point of growth, because its AI earnings are a 2027 promise (MI450/Helios ramp) rather than a booked run-rate. Whether the gap is a gift or a trap is exactly kill-switch #6's question, so the warning box below is part of the chart.

Median fwd PEG — memory 7
≈0.20
SK hynix & Micron at 0.10 are the cheapest screens on the board
Median fwd PEG — Mag 7
≈1.4
META 0.99 → TSLA 5.1 · a ~7× premium per point of growth
The crossover
NVDA 0.29
the only Mag-7 name inside the memory band — it sells the same cycle
The other HBM buyer
AMD ~1.31
⚑ 4–5× Nvidia's PEG — AI earnings still a 2027 promise
The catch
KS#6
memory looked exactly this cheap at the 2018 top — then E collapsed
The PEG ladder — 15 companies, one scale
Forward PEG, log scale. Ember = memory 7 (this board's §04 financials, compiled Jun 24–27) · teal = Magnificent 7 (tracker prints, Jul 14–16) · violet = AMD, the merchant-accelerator comparator (Jul 20–22). Dashed line = PEG 1, the classic "growth fairly priced" rule of thumb. ⚑ = sources disagree on that name (see table). Hover any row for the inputs.
Why cyclicals screen “free” at the top — read before concluding anything
A PEG of 0.10 does not mean the market is asleep — it means the market refuses to annualize this growth. Both of PEG's inputs are at cyclical extremes for the memory names: the E in P/E is a record-margin peak (SK hynix just printed a 72% operating margin), and the G is a once-per-decade repricing (+198%, +346%) that mechanically cannot repeat. At the 2018 peak Micron traded ~5× forward earnings — a memory-maker screen every bit as “cheap” as today's — then fell >50% while getting more expensive on collapsing estimates [approx., §02·D]. That is kill-switch #6's signature: when FY-forward estimates crack (KS#1's price deceleration is the leading edge), PEG re-rates violently upward with no price move at all. The honest reading of this chart: the gap measures the market's disbelief in memory-earnings durability — the thesis bet (§02) is that HBM contracts, SCAs and structural undersupply make this cycle's E stickier than 2018's. That is the bet; this chart just prices it.
The inputs — every number, its source and its date

Provenance & method. Memory side [board compile, Jun 24–27 2026]: forward P/E, latest YoY growth and forward PEG are §04's financials rows (SK hynix 6.8× / +198% / 0.10 · Samsung 6.8× / +69% / 0.20 · Micron 9× / +346% / 0.10 · Kioxia 11× / +58% / 0.30 · SanDisk 12× / +97% / 0.20 · Seagate 17× / +45% / 0.40 · Western Digital 25× / +40% / 0.43). Post-crash caveat: the Jul 8–13 rout cut the Korean names ~25–35% from their peaks with estimates so far intact — their spot P/E and PEG are now mechanically lower than these compiled values; refreshed on the next financials pass rather than guessed here. Mag-7 side [tracker prints, Jul 14–16 2026]: Nvidia fwd P/E ~24.0, PEG 0.29 (financecharts, Jul 14; GuruFocus fwd P/E 22.8 — ⚑ variance); Meta fwd P/E ~20.3, PEG 0.99 (fullratio; GuruFocus 0.82 — ⚑); Microsoft fwd P/E ~19.5, PEG 1.34 (fullratio/GuruFocus; Motley Fool quotes fwd P/E 24.5 — ⚑ basis variance); Amazon fwd P/E ~29, PEG 1.4 and Apple fwd P/E ~33, PEG 2.5 and Tesla fwd P/E ~179, PEG 5.1 (24/7 Wall St, Jul 15); Alphabet fwd P/E ~27.9, PEG 1.77 (stockanalysis, Jul 16; GuruFocus 25.1 fwd P/E — ⚑). The uncomfortable footnote every PEG screen owes you: PEG has no standard basis — publishers mix trailing vs forward P/E and 1-year vs 3–5-year expected growth, and the memory rows' G is YoY revenue-cycle growth while the Mag-7 rows lean on analyst EPS-growth estimates. Treat this as an indicative screen of two regimes, not a precision cross-table; conflicting prints are flagged ⚑, not merged. Cross-refs: §04 (financials), §02·D KS#1/KS#6, §04·D (targets). AMD [added Jul 24, flagged ⚑]: the widest source spread on the board — forward P/E prints of ~57×, ~73× and ~77× coexist across financecharts/fullratio/GuruFocus for the same week, and PEG reads 1.31 (Jul 20) or 0.96 depending on which growth window is used (consensus EPS growth is ~+64% for FY26, ~+58% FY27, ~+38% FY28 — PEG falls as you extend the window). 1.31 is plotted; the 0.96 alternative is disclosed here rather than averaged in. AMD is grouped separately as AI silicon — it is neither a memory maker nor a Mag-7 member, and is included because it is the second-largest merchant buyer of HBM (§06) and the cleanest read on how the market prices accelerator growth versus memory growth. Its premium to Nvidia on this screen is the market paying for booked revenue over promised revenue. Sources: financecharts.com, fullratio.com, GuruFocus, stockanalysis.com, 24/7 Wall St, TIKR, Motley Fool, company disclosures. Not investment advice. Aug 17 2026 — the unadjusted version of this question is now mapped [§04·L]. SK hynix prints a forward P/E of 6.88× against a semiconductor median of 31.58× — a 78.2% discount. §04·L tests the value-trap counter-argument using §04·B's dated estimate cuts and finds the multiple re-rates to 8.1–9.4× on an unchanged price, while still leaving a 70%+ discount. Growth-adjusting it, which is this section's job, is the more appropriate lens for exactly that problem.

04·LGlobal Large-Cap Forward P/E — Heat Map◷ mixed dates · Aug 2026
04·L

Global Large-Cap Forward P/E

Memory trades at a forward multiple no other large cap on this board comes close to: SK hynix at 6.9× and Micron at 10.6×, against a semiconductor median of 31.6×, an S&P 500 at 21× and ASML at 48.8×. The obvious reading is that memory is cheap. The obvious counter-reading is that a single-digit multiple on a commodity cyclical at peak earnings is a value trap, because the E in P/E is about to fall. This board can test that second claim rather than just repeat it — because §04·B recorded the exact estimate cuts, dated, six days before the 6.9× print. Applying them moves the multiple to 8.1–9.4× on an unchanged share price. That is a real re-rating from the denominator alone. It also still leaves a 70%+ discount to the sector — so the honest answer is that both readings are partly right, and the section says so instead of picking one.

Forward P/E across global large caps — sorted cheapest to dearest
Colour encodes the multiple and nothing else. A conventional heat map would shade low multiples green for "cheap"; doing that here would have asserted the very thing the section is questioning, so the ramp is a neutral cool-to-warm scale over 6×–50×. Blank hatched tiles are companies for which no dated, attributable forward multiple could be sourced — they are named and left empty rather than filled from an undated aggregator. Read the as-of stamps on each tile: these figures are not contemporaneous, which is the single biggest weakness of any such grid. Hover any tile.
The value-trap claim, tested against this board's own estimate record
Kiwoom cut these operating-profit estimates on Aug 11; the 6.88× multiple printed on Aug 17. A forward P/E is price over expected earnings, so a cut to the estimate raises the multiple even if the share price never moves. The final column is the discount that survives.
Both readings are partly right, which is the useful conclusion
The trap argument is real and quantifiable here. On the estimates one broker published six days before the multiple was struck, SK hynix's forward P/E is not 6.9× but 8.2× on the 2027 cut and 8.8× on the 2028 cut — a +18% to +29% re-rating with no share-price move at all. On Samsung's steeper 2028 cut the equivalent shift is +37%. Anyone quoting a single-digit multiple is quoting it against an earnings line the sell side was actively marking down that fortnight (§04·B), and §04·B also found the brokers disagreeing by 2.24× on where the stock should trade — so the E is contested as well as falling. But the discount does not disappear. Even at 8.8×, SK hynix sits 72% below the semiconductor median; at 9.4× it is still 70% below. A cyclical discount of that size is not explained away by one round of estimate cuts — it would take an earnings decline of roughly three quarters to bring the multiple to the sector median, which is a far more severe outcome than anything currently forecast. So the defensible statement is narrow: the multiple is genuinely low, it is lower than it looks because the denominator is falling, and it is still low after you correct for that. What it is not is a like-for-like comparison with a compounder at the same number — which is the actual content of the value-trap warning, and is a statement about earnings durability rather than about price.
Four reasons this grid is weaker than it looks
(1) The tiles are not contemporaneous. They carry as-of stamps ranging from Aug 2 to Aug 17 and some only to "2026". Forward multiples move daily; a grid implies a snapshot and this is not one. (2) The providers differ. Figures come from several sources with different consensus panels, different fiscal-year conventions and different treatment of GAAP versus adjusted EPS — an especially large effect for a company like SK hynix whose fiscal year and reporting currency differ from the US names beside it. (3) Coverage is partial. Only 10 of 27 named large caps could be sourced to a dated figure; the rest are drawn blank. A grid of this kind invites the reader to scan for pattern, and 63% of the pattern is missing. (4) One source conflict, unresolved. One report puts Samsung and SK hynix both below 6×; the dated figure used here is 6.88× for SK hynix. Both are described as forward multiples in mid-2026. This board uses the dated one and draws Samsung blank rather than adopt an undated sub-6× figure for either. One check that does pass: the cited "78.2% below the semiconductor median" reproduces exactly from 6.88 ÷ 31.58, so those two figures at least come from a common basis.

Provenance & the honesty line. [cited, with dates where published]: SK hynix forward P/E 6.88× (Aug 17 2026); Micron 10.57×; Meta 21.9×; NVIDIA 22×; Amazon 30.7×; Apple 34.11× (Aug 12 2026); AMD 35×; ASML 48.82× (Aug 2 2026); Alibaba and Tencent described as trading at 10–15× against 25–40× for US technology peers. Benchmarks: global equities 18.89× (Jan 1 2026), S&P 500 ~21×, Magnificent Seven ~28–30×, semiconductor median 31.58×, with SK hynix cited at 78.2% below that median. [cited — §04·B, Aug 11 2026]: Kiwoom cutting SK hynix 2027 operating profit −15.6% and 2028 −22.2%, Samsung 2027 −15.0% and 2028 −26.9%. [derived]: every re-rated multiple in the table (6.88 ÷ (1 − cut)), the +18% to +37% re-rating percentages, and the surviving discounts of 70–72%. These are illustrative arithmetic, not a forecast — they apply one broker's operating-profit cuts to a multiple built on a different (net-income, consensus) earnings base, so they show the direction and rough scale of the denominator effect, not a corrected multiple. A precise version would need the consensus EPS the 6.88× was struck against, which is not published in the sources available here. What this section does not do: (1) no view on whether memory is cheap — it sets out what each reading requires you to believe; (2) no filled cells for unsourced companies, so the grid is explicitly 37% complete; (3) no PEG or growth adjustment, which is §04·G's job and is the more appropriate lens for exactly this problem. On the source that prompted this: it is a short argument that memory's low multiple reflects a commodity cycle rather than cheapness, and contains no valuation table — the grid here was assembled independently. The argument itself is the section's subject and is treated as a hypothesis to test, not a conclusion to relay. Cross-refs: §04·G (forward PEG, the growth-adjusted version of this question), §04·B (the estimate cuts and the 2.24× dispersion in price targets), §04·D (price targets), §04·I (gross margin by holding), §02·B (cycle position), §02·D (kill-switches). Not investment advice.

04·HEarnings Scorecard — the Thesis's Scheduled Tests◷ dated · next: Jul 23
04·H

Earnings Scorecard

The Catalyst Calendar (§02·E) says when; this says what each print has to prove. Every memory earnings date is a scheduled test of a kill-switch (§02·D): consensus/guidance vs actual vs the stock's reaction, in one board. The near-term trio is decisive — SK hynix Q2 (Jul 23) and Samsung full Q2 (Jul 30) land this month, and Micron FQ4 (~Sep 22) is where the guided "moderation" either shows up or doesn't. Reported rows carry the actual; upcoming rows carry the bar to clear.

Provenance. Reported [cited]: Micron FQ3'26 (Jun 24) — revenue $41.5B, EPS $25.11, next-quarter guide ≈$50B (Micron IR; §04·B ratings row). Samsung Q2 preliminary (Jul 7) — record ₩89.4T operating profit ≈19× YoY on ₩171T revenue, stock −6% on profit-taking (Samsung IR; §02·E). Upcoming [scheduled/consensus]: SK hynix Q2 ~Jul 23 (consensus ~₩63–64T operating profit); Samsung full Q2 Jul 30; Micron FQ4 + FY26 ~Sep 22–24. Consensus figures are pre-print estimates and move; each row states the kill-switch it tests rather than predicting the outcome. Cross-refs: §02·D, §02·E, §04·B, §04·D. Sources: company IR, TrendForce, analyst consensus. Not investment advice.

05The HBM Roadmap◷ structural
05

The HBM Roadmap

HBM stacks DRAM dies vertically and links them with through-silicon vias — 5–10× the bandwidth of ordinary GDDR. Each generation is matched to a class of AI accelerator.

HBM3Mature
Bandwidth / stack
~819 GB/s
Status
Workhorse, 2023–24 GPUs
Powers
Nvidia H100 · AMD MI300X
Typical capacity
80–192 GB / GPU
HBM3ECurrent volume
Bandwidth / stack
~1.18 TB/s
Status
Mass production since 2024
Powers
Nvidia H200 · B200 · AMD MI350
Build
12-high stacks
HBM4The new front
Bandwidth / stack
~2 TB/s · 2,048-bit
Status
Volume production Q1 2026
Powers
Nvidia Vera Rubin · AMD MI400
Milestones
All 3 makers qualified Jun 2026 · 12-Hi in volume
16-Hi HBM4 / HBM4ENext front
Bandwidth / stack
~2.5+ TB/s
Status
NVIDIA requested 16-Hi by Q4 2026
Density
16 dies stacked · up to 64GB+
Note
May be branded HBM4E; extreme stacking yield challenge
HBM5On the horizon
Status
Samsung–NVIDIA in talks (Jun 8); mock-up at Computex
Architecture
2nm base die · new thermal design · maxes at 16-Hi
Timing
Late-decade; 20/24-Hi not until ~HBM7 (~2035)
05·BHBM's Bite of DRAM Supply◷ monthly model
05·B

HBM's Bite of DRAM Supply

◷ monthly model

The single chart that explains the shortage: what share of the world's DRAM supply HBM consumes, monthly since 2023, projected to 2035. The denominator matters — HBM is DRAM, and its share looks wildly different depending on how you count. By bits shipped it's still small (~9%). By revenue it's huge (~38%) because of premium pricing. The metric that drives the squeeze is wafer capacity: producing 1 GB of HBM eats roughly 4× the wafer area of standard DRAM (bigger dies, TSV stacking, yield loss), so HBM's ~9% of bits devours ~23% of the world's DRAM wafers — capacity denied to the DDR5 in servers, PCs and phones. That wedge is the shortage.

HBM as a percent of total DRAM supply, by three denominators, monthly Jan 2023 → Dec 2035. Solid = estimate calibrated to reported anchors; right of the divider = projection, with a low/high scenario band on the wafer-capacity series (the band widens because a 9-year memory forecast is guesswork). Hover any month.
Share of DRAM wafer capacity (the supply squeeze) Share of DRAM bits shipped Share of DRAM revenue Wafer-share scenario band (low–high)
Who makes the HBM — vendor share (estimates diverge)

The chart above is HBM versus the rest of DRAM. This is the other split — who supplies the HBM — and it is genuinely contested: firms measure different things (revenue vs bits vs wafer-input) and the #2 slot has flipped quarter to quarter. Shown as a sourced range, flagged, not merged.

EstimateSK hynixSamsungMicronBasis / note
Counterpoint · Q1 202658%21%21%revenue — this board’s cited anchor (Counterpoint / Reuters)
IDC · Q1 202656.4%revenue, rank #1
TrendForce · Q1 2026~70%higher on a different methodology
Counterpoint · Q2 202562%17%21%the “Micron overtakes Samsung” quarter
Counterpoint · Q3 2025~53%~35%~12%Samsung rebounds to #2 (Micron = residual)

The widely-shared “Micron overtakes Samsung in HBM” headline is Q2 2025 Counterpoint data (SK 62 / Micron 21 / Samsung 17), not a 2026 event — Samsung rebounded to ~35% the very next quarter. Q1 2026 reads put SK hynix anywhere from 56% (IDC) to ~70% (TrendForce), with Samsung and Micron roughly tied near 21% on Counterpoint. §04’s per-company “Samsung HBM share ~30–35%” reflects its Q3 2025 peak and 2026 target, above Counterpoint’s Q1 2026 ~21% read — the gap is the point, so both are shown rather than averaged. Sources: Counterpoint Research (Global DRAM & HBM Market Share, Jun 8 2026), IDC, TrendForce, Reuters. Not investment advice. See also: §05·B·V, which holds the contested version of this picture — three research houses giving three different HBM vendor shares for the same quarter. Aug 15 2026 — an absolute anchor for the share this section computes [§05·C·H]. This section works in percentages; §05·C·H supplies the numerator in gigabytes. A bottom-up build of 2027 accelerator demand gives 6.08 bn GB of HBM (48.6 bn Gb). Against this section's ~13% of 2027 DRAM bits, that implies a ~46.7 bn GB total DRAM market — the top-down share path and the bottom-up SKU build are compatible, which was not guaranteed and is worth knowing before either is relied on.

Three honesty notes, in increasing order of importance. (1) Monthly points are interpolated — HBM share is reported quarterly/annually at best, so the monthly curve is a smooth path through the reported anchors, not monthly data. (2) The anchors are solid: TrendForce has HBM at 19% of total DRAM wafer output in 2025 → 23% in 2026; 5% of DRAM bit shipments and 20% of DRAM revenue in 2024; 1 GB of HBM consumes ~ the wafer capacity of standard DRAM (GDDR7 ~1.7×), which is how a single-digit bit share becomes a ~quarter of world wafer supply; per-supplier 2026 HBM wafer shares for the big three run roughly 22–33%; HBM demand is seen growing ~70% in 2026; AI in total (HBM + GDDR7) approaches ~20% of global DRAM capacity-equivalent on ~40 EB of 2026 output. (3) The projection to 2035 is the longest in this dashboard and should be trusted the least — nobody credibly forecasts memory mix nine years out. The base path assumes HBM wafer share keeps rising but decelerates (~30% in 2027 per TrendForce, ~33% in 2028, ~37% by 2030, ~40% by 2035) as new fabs and HBM4/5 base-die efficiency arrive; the band spans a cool-off case (~28% by 2035) to an AI-eats-everything case (~55%). Past the late-2020s, treat every value as scenario, not forecast. Bit- and revenue-share projections are single dashed paths for legibility but carry the same uncertainty. Sources: TrendForce (DRAMeXchange), Commercial Times via TrendForce News, FMS 2025 proceedings, company disclosures. Not investment advice.

05·B·VHBM Vendor Share — Who Actually Makes the HBM (Contested)◷ firms disagree
05·B·V

HBM Vendor Share — the Contested Picture

§05·B is HBM vs the rest of DRAM; this is who supplies the HBM — the single most thesis-relevant, and most disputed, number in the stack. SK hynix leads decisively, but the exact split swings quarter to quarter and disagrees firm to firm. The bars are one internally-consistent source (Counterpoint) so the quarters are comparable; the firm-divergence on SK's Q1'26 share is shown separately. Flagged, not merged.

HBM revenue share by quarter — Counterpoint basis
Stacked to ~100%. The #2 slot flips: Micron over Samsung in Q2'25 (the viral "Micron overtakes Samsung" quarter), Samsung back ahead in Q3'25, roughly tied in Q1'26.
SK hynixSamsungMicron
SK hynix Q1'26 share — the same quarter, three firms

Provenance & the honesty line. Bars [cited · Counterpoint]: Q2 2025 SK 62 / Micron 21 / Samsung 17; Q3 2025 SK ~53 / Samsung ~35 / Micron ~12 (Micron a residual); Q1 2026 SK 58 / Samsung 21 / Micron 21 (Counterpoint “Global DRAM & HBM Market Share,” Jun 8 2026, via Reuters). Firm divergence [flagged]: for Q1 2026 SK hynix, IDC prints 56.4% (revenue rank), Counterpoint 58%, TrendForce ~70% — a ~14-point spread driven by revenue-vs-shipment-vs-wafer methodology and lumpy HBM revenue recognition. The widely-shared “Micron overtakes Samsung” headline is Q2 2025, not a 2026 event — Samsung rebounded the next quarter. §04's per-company “Samsung HBM share ~30–35%” reflects its Q3'25 peak / 2026 target, above Counterpoint's Q1'26 ~21% read; both are shown rather than averaged. Q4'25 is a public-data gap (not charted). Sources: Counterpoint, IDC, TrendForce, Reuters. Not investment advice. Aug 15 2026 — a source that looks like it settles this, and does not. The Morgan Stanley SKU-level exhibit replicated in §05·C·H names HBM suppliers for every 2027 accelerator — but as unranked strings ("Samsung/Hynix/Micron", "Hynix/Micron"), with one entry reading literally "Samsung?". No vendor allocation can be derived from it, so it does not narrow the disagreement this section documents. It is noted here specifically so the exhibit is not mistaken for a resolution.

05·CHBM Cost Per Bit, by Generation◷ estimates · no public market
05·C

HBM Cost Per Bit, by Generation

◷ estimates · no public market

What a gigabyte of each HBM generation actually costs — and the single most important caveat on this entire page: HBM has no transparent market price. Unlike DDR5 (which has published DRAMeXchange spot & contract prices), HBM is sold only through confidential long-term contracts between three suppliers and a handful of buyers, bundled and volume-dependent. Every number below is an analyst triangulation from earnings commentary, teardown cost models, and industry reports — directional, not a quoted price. The pattern that is reliable: each generation costs more per bit (more stacked layers, finer TSV pitch, lower yield), and HBM sits at a large premium to commodity DRAM. The newest hard anchor: on June 5, SK hynix finalized 2026 HBM4 price/volume terms with NVIDIA at >50% above HBM3E — the first concrete HBM4 price signal, replacing earlier guesswork.

Estimated price per GB — bars show low–high range
Approximate contract $/GB by generation, mid-2026 estimates (HBM4 is pre-volume). DDR5 server memory shown as the commodity-DRAM baseline. Bar = reported low-to-high range; number = midpoint. Hover the table below for per-stack prices and sources.

Note the squeeze at the right of the conventional-DRAM bar: DDR5 itself is now ~$10–12/GB in this shortage, up from ~$3–4 a year ago. That's why HBM3E's premium over DDR5 has collapsed from ~4–5× (early 2025) toward ~1–2× by end-2026 (TrendForce) — not because HBM got cheaper, but because commodity DRAM caught up.

Provenance — read this before citing any number. There is no public HBM spot or contract index; the three suppliers (SK hynix, Samsung, Micron) sell HBM under confidential long-term agreements, so all figures are analyst estimates triangulated from earnings calls, teardown cost models, and trade-press reports, and they vary widely by source, buyer, volume and quarter. Per-stack anchors (Silicon Analysts, Apr 2026): HBM3 ≈ $200 / 24 GB stack (~$8–10/GB), HBM3E ≈ $300 / 36 GB stack (~$8–13/GB), HBM4 ≈ $500 / 48 GB stack (est., ~$10–14/GB; pre-volume). HBM3E reportedly peaked at ~$17–20/GB in H1 2025 at the demand peak (Silicon Analysts citing TrendForce / Goldman Sachs / SemiAnalysis), then 2026 contracts were struck lower before a reported ~20% HBM3E price hike for 2026 (TrendForce / Chosun Biz). HBM4 (June 5, 2026): SK hynix finalized 2026 HBM4 price & volume with NVIDIA at >50% above HBM3E per Korean press — implying ~$17–19/GB on the HBM3E midpoint, the first negotiated HBM4 anchor (prior figures were pre-volume estimates). Legacy HBM2/HBM2E ($/GB) are directional single-source estimates — public data is sparse before the 2022 AI inflection. Baseline: server DDR5 ≈ $1.50/Gb (~$12/GB) currently (Counterpoint), elevated by the shortage. Premium dynamic: HBM3E commanded >4× DDR5 in 2Q25, narrowing toward 1–2× by end-2026 as DDR5 spikes (TrendForce). Cost-of-production driver: 1 GB of HBM needs ~3–4× the wafer capacity of DDR5 (Tom's Hardware / TrendForce). System context: a B200 with eight HBM3E stacks carries ~$2,400 of memory, exceeding its logic-die cost (Silicon Analysts). "$/GB" is the practical proxy for cost-per-bit (1 GB = 8 Gb). Sources: Silicon Analysts, TrendForce, Counterpoint Research, Goldman Sachs, SemiAnalysis, Tom's Hardware, KED Global, Chosun Biz. These are estimates of a non-public price. Not investment advice.

05·C·H2027 HBM Demand, Built Bottom-Up — 17 Products, One Total◷ MS supply-chain exhibit · replicated & verified
05·C·H

2027 HBM Demand, Built Bottom-Up

Every other HBM section on this board works top-down — bit growth, wafer share, trade ratios. This one works bottom-up from named silicon: CoWoS wafer allocation → chips per wafer → units shipped → stacks per package → gigabytes. Seventeen products, each with its own HBM generation and supplier list, summing to 6.08 bn GB of HBM in 2027 — the source's "up to 50bn Gb". The board replicated the whole table before using any of it, and it holds: every visible row's arithmetic reproduces exactly, and the two blank cells are recoverable by difference and close both column totals to the unit. But the headline number is not the interesting part. The interesting part is the mix: custom ASICs consume as much HBM as merchant GPUs in 2027 — 49.7% against 49.4% — and Google alone reaches 97% of NVIDIA's HBM consumption. Meanwhile HBM3E is still 48% of demand in a year the roadmap calls the HBM4 era.

2027 HBM consumption by buyer, segmented by generation
Bars are gigabytes of HBM, split by the generation each product carries. The dashed line separates merchant GPU buyers from custom-ASIC buyers, and it lands almost exactly at the midpoint of total demand — which is the structural claim worth extracting from this exhibit. The "Others" bar is drawn hatched because the source prints a wafer allocation for it but no HBM figure; the value shown is recovered as the stated column total minus the visible rows, and the source does not disclose its generation. Hover any bar for the product-level build.
The replicated table — every cell, including the two this board recovered
Starred amber cells are not in the source. They are recovered by difference: the CoWoS allocation for TPU v9 is the only blank in a wafer column whose total is stated, and the HBM demand for "Others" is the only blank in a demand column whose total is stated. Each is forced to a single value, and each closes its column exactly — which is also a check on the transcription.
Two models, one total, three-times-apart components
§05·F builds 2027 accelerator units independently, from a different bank's forecasts. It lands on 21.30M units. This exhibit lands on 21.14M. Those two numbers are 0.7% apart — far closer than either model's own error bars, and a genuine cross-validation of the aggregate. The composition does not survive the same test. On Meta the two differ by 3.1× (1.70M vs 0.55M), on Microsoft 1.8×, on AMD 1.5×, on Google 1.5× — and they disagree in both directions, which is why the totals still meet. The practical rule this implies is worth stating plainly: conclusions on this board that depend on how many accelerators ship in 2027 are well supported by two independent methods. Conclusions that depend on whose accelerators ship are supported by neither. §05·F's per-firm bars, and any HBM-vendor-share inference drawn from them, should be read with that in mind — and this exhibit does not fix the problem, because it lists HBM suppliers as unranked strings like "Samsung/Hynix/Micron" from which no share can be derived (see §05·B·V).

Provenance & validation. [cited — transcribed from a user-supplied exhibit]: Morgan Stanley Research, "AI HBM consumption: Up to 50bn Gb in 2027", compiled from the firm's Asian supply-chain checks, with company data. All 2027 estimates. Every figure in the table above is transcribed, not modeled. Validation performed before use, and it passes: (1) implied shipments = CoWoS allocation × chips per wafer for all 16 rows that carry both — exact, no rounding drift; (2) GB per package = stack density × stack count for all 15 HBM-bearing rows — exact; (3) HBM demand = units × GB per package for all 15 — exact; (4) the unit chain closes: 6,077,256 k GB × 8 ÷ 1,000 = 48,618.048 mn Gb against the highlighted 48,618consistent to the source's own rounding (a 0.0001% difference), not bit-exact. Stated precisely because an earlier draft of this section claimed "exactly" and the validation harness caught it. [derived by this board]: the 148k-wafer CoWoS allocation for TPU v9 and the 53,640 k GB HBM demand for "Others" — each is the unique value that closes a stated column total, and both close to the unit. The implied 2.70 chips per CoWoS wafer for TPU v9 is a consequence of that recovery, not a source figure; it is low but consistent with a very large multi-die package, and it is the least certain number on this page. What this exhibit cannot do, and is not used for: (1) HBM vendor share. The supplier column lists two or three names per product with no allocation between them, and one entry is literally "Samsung?" — no vendor split can be computed from it, and none is attempted here (§05·B·V holds the contested vendor picture). (2) Any year but 2027. There is no time series; this is a single-year snapshot and is not extrapolated. (3) Supply. It is a demand build only — whether 6.08 bn GB can be produced is §05·D and §05·E's question, not this one. Reconciliation to the rest of the board [derived]: §05·D carries HBM at ~13% of 2027 DRAM bits; 6.08 bn GB at 13% implies a total 2027 DRAM market of ~46.7 bn GB, which sits inside the range implied by §06·E's shipment series — so the bottom-up and top-down views are compatible, which is not guaranteed and is worth knowing. One caution on the CPUs: Vera and Venice occupy 520k CoWoS wafers — 20% of the total allocation — and carry no HBM. Any model that infers HBM from CoWoS capacity alone will overstate it by roughly that fraction. Cross-refs: §05·C·G (what these generations actually are), §05·B (HBM's share of DRAM supply), §05·B·V (why vendor share stays contested), §05·D and §05·E (whether the wafers exist), §05·F (the independent unit forecast this is checked against), §06·B (the accelerator memory roadmap), §06·B·T (Google's TPU line, the largest single ASIC buyer here). Not investment advice.

05·C·GHBM Generations — HBM3 → HBM4E: What Changes & What Doesn't◷ JEDEC-anchored · +wafer consumption · HBM4E projected
05·C·G

HBM Generations — HBM3 → HBM4E

Four generations, one family. The similarities are the shared DNA below — every generation is the same fundamental idea: stacked DRAM dies wired vertically with TSVs, sitting on a 2.5D silicon interposer next to the GPU, trading a wide-and-slow bus for GDDR's narrow-and-fast one. The differences are where each generation finds its bandwidth: HBM3→3E pushed pin speed on the same 1024-bit bus; HBM4 is the architectural break — it doubles the bus to 2048-bit (while its per-pin speed actually falls back to HBM3-class), moves the base die from a DRAM process to a logic process, and requires a new PHY and interposer; HBM4E then pushes pin speed again on the doubled bus. Honesty note: HBM4E has no announced product — its figures are projections and are marked as such throughout.

3D TSV-stacked DRAM dies2.5D silicon interposer (CoWoS / EMIB)wide-&-slow bus (vs GDDR narrow-&-fast)channel / pseudo-channel architectureJEDEC-governedknown-good stacks mounted beside the GPU
Stack anatomy, side by side
Each column is one generation's stack, drawn to the same scale: DRAM dies (TSVs dotted through them), the base die, and the bus down to the interposer. The continuous gray strip is the point — every generation sits on the same 2.5D-interposer concept. What changes: stack height (12-Hi → 16-Hi), the base die (gray = DRAM process; green/violet = logic process), and the bus ribbon, which doubles in width at HBM4.
Bandwidth per stack (TB/s) — the headline race
Per-pin speed (Gb/s) — the tell

The tell: HBM4's per-pin speed drops back to HBM3-class (6.4–8 Gb/s spec) — its ~2× bandwidth gain comes from the doubled 2048-bit bus, not the clock. HBM3E and HBM4E are the "speed" generations; HBM4 is the "width" generation.

Spec sheet — generation by generation
Wafer consumption per generation — the number that connects this section to the shortage 1 CITED ANCHOR · REST DERIVED
Bandwidth and capacity are what HBM sells; wafers are what it costs the industry. The single hardest number here is Micron's own disclosure: one gigabyte of HBM3E consumes about three times the wafer capacity of one gigabyte of DDR5. That 3× is why every HBM bit shipped removes roughly three DDR5 bits' worth of fab from the consumer pool (§08·B is what that feels like at the checkout), and it is the mechanism behind the wafer-starts crunch in §05·E. Three effects compound to produce it — and all three get worse with each generation.
① Die-area penalty
TSV tax
Thousands of through-silicon vias plus their keep-out zones consume array area, so an HBM die yields fewer bits per mm² than a DDR5 die on the same node. HBM4's 2048-bit interface widens the periphery again; TSV pitch tightens from 40–45 µm (HBM3) toward sub-30 µm.
② Stack yield (KGSD)
40–60% @ 16-Hi
A stack is only good if every die in it is good. Yields fall 10–20 points going 8-Hi→12-Hi and drop toward a 40–60% band nearing 16-Hi (warpage, cracking, bond integrity). Every failed stack throws away all 16 dies' worth of wafer.
③ Base die leaves the DRAM fab
+ logic wafer
From HBM4 the base die is built on a logic process at TSMC (12nm FFC+ / 4–5nm), not in a DRAM fab. That consumption is real but invisible in DRAM wafer-start statistics — HBM4 draws on leading-edge foundry and CoWoS packaging on top of its DRAM wafers.
Where it lands
18% → 34%
HBM's share of all DRAM wafer starts, 2025→2028 (Deutsche Bank/Gartner; J.P. Morgan has 19%→31%). TrendForce puts AI at ~20% of global DRAM wafer capacity in 2026. Full model in §05·E.
The arithmetic, stated plainly. Wafer per GB ≈ (bits per wafer on a DDR5-equivalent die ÷ HBM die's bit density) × (1 ÷ stack yield). Micron's cited ~3× for HBM3E is the product of those first two terms at 12-Hi maturity. Push to 16-Hi and term ② alone can add ~30–50% more wafer per good stack; widen the die for a 2048-bit bus and term ① grows too — which is why HBM4 is derived here at ~3.5–4× and HBM4E higher. Only the HBM3E ~3× is a disclosed figure. The HBM4/4E multipliers are this board's derivation from the cited yield bands and die-geometry direction — no vendor publishes a per-generation wafer ratio. Treat them as order-of-magnitude with the right sign, not precision: the defensible claim is that wafer intensity rises monotonically across generations, never falls.

Counter-consideration, stated: per-die capacity also rises (24Gb→32Gb dies), which pushes wafer-per-GB down and partly offsets ① and ②. Net direction stays up in every public model, but the generation-over-generation step is smaller than the yield numbers alone imply — which is exactly why a single vendor-published HBM4 ratio would be worth more than this whole panel.

Provenance & the honesty line. Spec anchors [JEDEC + vendor disclosures]: HBM3 standardized Jan 27 2022 (JESD238): 1024-bit bus as 16×64-bit channels, 6.4 Gb/s/pin → ~819 GB/s per stack; SK hynix mass production Jun 2022 for Nvidia's H100. HBM3E is a vendor-enhanced extension of HBM3, not a new JEDEC generation: same 1024-bit bus, 16 channels / 32 pseudo-channels, pin speed pushed to 8–9.6 Gb/s (Micron's 9.6 → ~1.2 TB/s; up to 12.4 in advanced implementations), 36 GB 12-Hi with a 48 GB 16-Hi announced (SK hynix, Nov 2024), ~2.5× perf/W vs HBM2E, backward-compatible with HBM3 controllers (Siemens EDA design guide, Rambus, Wikipedia). HBM4 ratified Apr 15 2025 (JESD270-4): 2048-bit / 32 channels / 64 pseudo-channels — the doubled bus is the generation break; JEDEC per-pin spec 6.4–8 Gb/s (HBM3-class; Samsung has demoed ~13) → ~2.0 TB/s at 8 Gb/s and up to ~3.3 TB/s in advanced configs; up to 16-Hi with 32 Gb dies = 64 GB/stack; the base die moves to a logic process (12nm FFC+ / 5nm at TSMC), opening near-memory compute; first parts shipped late 2025/early 2026, and by mid-2026 all three makers are in HBM4 mass production — SK hynix and Micron first, then Samsung shipping industry-first commercial HBM4 from Feb 12 2026 on its 1c (6th-gen 10nm-class) DRAM with a 4nm logic base die: 11.7 Gb/s (up to 13), 3.3 TB/s per stack, 12-Hi 24–36 GB (16-Hi 48 GB to follow). On Jun 5 2026 Nvidia certified all three (SK hynix, Samsung, Micron) as HBM4 suppliers for Vera Rubin, with supply-chain estimates putting SK hynix ~60–70% / Samsung ~25–30% / Micron the remainder of Rubin volume; Nvidia Rubin R100 (H2 2026, up to 22 TB/s per GPU) and AMD MI400/MI430X are the first major platforms (Siemens, IntuitionLabs, Spheron, HyperPC, Wikipedia). HBM4E [projection — flagged]: in development for ~2027–28; controller IP already supports up to 16 Gb/s (Rambus), implying ~4.1 TB/s per stack on the 2048-bit bus; Micron's roadmap points to customized logic base dies; no GPU product is announced as of mid-2026 and its specs are not manufacturer-confirmed (Spheron, IntuitionLabs) — every HBM4E figure here is hatched/starred as projected. A source conflict, reconciled: Rambus materials describe HBM4/4E controllers as "backward compatible with HBM3" — that refers to multi-mode controller IP; the physical HBM4 interface is a hard break requiring new PHY, a finer-pitch interposer, and CoWoS-L-class packaging ("you cannot drop an HBM4 module into an HBM3E slot" — Siemens/Kynix/AIChipLink). The table reflects the physical reality. Why it's on-thesis: the width-doubling plus logic base die is part of why HBM consumes far more wafer and packaging capacity per GB (see §05·B share and §05·C cost) — and per Micron's FQ3 call (Jun 24 2026), HBM4 12-high is ramping ~2× faster than HBM3E did, with HBM3E + HBM4 sold out through calendar 2027. Wafer consumption [1 cited, rest derived \u2014 added Jul 30 2026]: the anchor is Micron's disclosure that a gigabyte of HBM3E takes roughly 3\u00d7 the wafer capacity of a gigabyte of DDR5 (reported via Tom's Hardware / TechTimes, Jun 2026) \u2014 the cause given is that vertical stacking cuts yield while adding process steps. Stack-yield band [cited]: yields fall 10\u201320 percentage points from 8-Hi to 12-Hi and drop toward 40\u201360% approaching 16-Hi (SemiEngineering; Samsung's Jun 30 2026 \u201cdummy die\u201d patent targets exactly this warpage/cracking failure mode). TSV geometry [cited]: HBM3-era TSV pitch 40\u201345\u202fµm moving toward sub-30\u202fµm for HBM4; the extra area needed for TSVs is why HBM dies are larger than DDR equivalents. Base-die shift [cited]: HBM4's base die is fabricated on a TSMC logic process (12nm FFC+ / 4\u20135nm on Samsung's commercial part), so it consumes foundry wafers that never appear in DRAM wafer-start data. Derived, not disclosed: the ~3.5\u20134\u00d7 (HBM4) and ~4\u00d7+ (HBM4E) multipliers are this board's construction from those cited inputs \u2014 no manufacturer publishes a per-generation wafer ratio, and a rising per-die capacity (24Gb\u219232Gb) partly offsets the penalty, so the multipliers carry real uncertainty even though the direction is not in doubt. Cross-refs: §05·E (wafer starts required, where HBM goes 18%\u219234% of DRAM wafers on the DB/Gartner model), §05·C (cost), §05·B (share), §08·B (the consumer-facing consequence). Sources: JEDEC (JESD238, JESD270-4), Rambus, Siemens EDA, Spheron, IntuitionLabs, Kynix, AIChipLink, Nevsemi, HyperPC, Wikipedia, Micron FQ3'26 earnings call, Tom's Hardware, TechTimes, SemiEngineering, TrendForce, Deutsche Bank/Gartner, J.P. Morgan. Not investment advice. Aug 15 2026 — the 2027 generation mix, quantified [§05·C·H]. This section describes HBM4 ramping roughly twice as fast as HBM3E did, which is accurate and can leave a misleading impression of 2027. A bottom-up build across 17 named accelerator SKUs puts 2027 demand at HBM3E 47.7%, HBM4 40.6%, HBM4E 11.7%HBM3E is still the largest single generation in the year HBM4 is supposed to own, because the highest-volume parts (Google's TPU v8i and v8t, ~7.6M units between them) stay on HBM3E 12-Hi. Fast ramp and majority share are different claims; only the first is supported.

05·DWafer Supply & Demand by Application — 2022→2035◷ 2026 anchored · pre/post modeled
05·D

Wafer Supply & Demand by Application

Two layers, both with surplus/deficit shown in their native units. Memory (DRAM) is in a structural deficit — measured in bits (demand growth outrunning supply growth) — driven by HBM eating a hugely disproportionate share of wafers. All-semiconductor wafers are not in aggregate shortage — measured in wafer-starts/month (kwspm) — but are sharply bifurcated: leading-edge (2/3nm) and advanced packaging (CoWoS) are sold out, while mature nodes sit in glut. The honest caveat up front: wafer capacity cleanly split by end-use is barely published (SEMI segments by device type and node, not application), so segment splits below are triangulated and tagged A anchored / E estimate / T triangulated.

① Memory deficit, in bits — DRAM bit demand vs supply growth
YoY bit-growth %. When demand (red) outruns supply (teal), the market tightens — the shaded gap is the deficit (red) or surplus (teal). 2026 is the one hard-anchored pair (TrendForce: demand +35% / supply +23%); 2022–25 are reconstructed estimates, 2027→2035 are scenario. Revised Aug 3 2026: TrendForce's Jul 30 outlook says the DRAM supply-demand gap widens further in 2027 — new fabs won't ramp meaningfully until 2H27 and substantial output won't land until 2028 — so 2027 was re-cut from a narrowing gap to a wider one and the crossover into surplus pushed to 2028. Those revised bars are this board's translation of a directional statement; TrendForce did not publish 2027 bit-growth percentages, so do not quote them as its numbers. This is a growth-rate proxy for balance, not inventory-adjusted — see the sufficiency-ratio note below, which is the level-based measure and reads very differently.
Bit demand growthBit supply growthDeficit (demand > supply)Surplus (supply > demand)
Two measures of the same shortage that look nothing alike — and why both are right
The chart above shows 2026 DRAM demand growing +35% against supply at +23% — a 12-percentage-point rate gap. TrendForce's Jul 30 note puts the 2026 DRAM sufficiency ratio at −1% to −2%. Those look irreconcilable and are not: they measure different things. A rate gap is the difference in growth over one year; sufficiency is a level ratio of supply to demand after inventory and carry-in. If the market had entered 2026 perfectly balanced, a year at +35%/+23% would end roughly 8.9% short — but it entered with inventory to draw down, so the realised level shortfall is 1–2%. The practical lesson: a 12-point rate gap and a 1–2% level shortfall describe the same market, and quoting either as "the deficit" without saying which one is how this cycle gets over- and under-stated in the same week. Both point the same way for 2027 — TrendForce expects the gap to widen.
New — DRAM and NAND stop moving together in 2027
For three years this board has treated memory as one cycle. TrendForce's July research splits it. DRAM stays tight: sufficiency −1 to −2% in 2026, gap widening in 2027, structural shortage running into 2028, because new capacity ramps only in 2H27 and — the mechanism this whole section exists to show — HBM consumes far more wafer input per bit, so rising wafer starts do not convert proportionally into bits. NAND goes the other way: a 4–5% supply deficit in 2026 turning positive in 2H27 as higher-layer migration and new fabs land into weak consumer demand. Chinese suppliers alone go to ~19% of global NAND bit output. The demand split explains it: servers are now >40% of NAND bit demand but phones and notebooks are still ~40% — and TrendForce has smartphone production −15 to −20% YoY in 2026 with notebooks ~−10%. DRAM's marginal buyer is a hyperscaler; NAND's is still a consumer, and the consumer has stopped buying. §07·C is the hedge on this: if HBF works, NAND acquires a hyperscaler-scale marginal buyer too — which would close exactly the gap that is about to open.
The risk TrendForce itself flagged, quoted rather than buried
The same Jul 30 note that forecasts a widening 2027 DRAM deficit ends on a caveat that cuts against its own conclusion, and it belongs here: CSP capital expenditure "remained at record levels throughout 2026, with some providers even reporting negative free cash flow." If memory prices stay elevated through 2027, memory takes a larger share of CSP infrastructure budgets — and whether that changes CSP investment plans and memory procurement is, in TrendForce's words, "a key variable to watch." That is the demand-side kill-switch (§02·D #3) stated by the same house whose supply-side forecast underwrites the bull case. This board records both halves. Cross-check it against §06·F·D (hyperscaler capex), §06·F·E (the debt funding it, now at rising coupons) and §08·D (whether app-layer revenue is arriving fast enough to justify either).
② All-semiconductor wafer capacity — total vs leading-edge (kwspm)
2026 is the year leading-edge capacity passes one million wafers a month for the first time — SEMI puts it at 1.16M wpm, which is the value plotted here. SEMI also forecasts +69% growth in advanced capacity through 2028, and 2nm-and-below going from under 200K wpm in 2025 to over 500K wpm by 2028. Global installed fab capacity in millions of wafer-starts/month (200mm-equivalent, SEMI basis), left axis. Leading-edge ≤7nm (right axis, note the ~25× smaller scale) is the thin, fully-booked sliver — its scarcity, not total wafers, is the AI constraint. 2022–2026 anchored to SEMI; 2027→2035 extrapolated at ~7%/yr (total) and ~14%/yr (leading-edge).
Total capacity (M wpm, left)Leading-edge ≤7nm (M wpm, right)▼ 2027+ extrapolated
HBM's wafer-vs-bit divergence — why memory breaks differently

1 GB of HBM consumes ~3–4× the wafer area of standard DRAM (TSV stacking, bigger die, 50–60% yield). So HBM is a single-digit share of DRAM bits but devours roughly a quarter of DRAM wafers — starving everything else.

Supply/demand balance by application segment — mid-2026
When relief arrives — capacity coming online (projection)

Aug 3 2026 refresh [cited — TrendForce primary releases]. DRAM, Jul 30 2026: 2026 sufficiency ratio −1% to −2%; supply-demand gap widens further in 2027; new 2027 capacity delayed to 2H27 ramp with substantial output only in 2028; HBM's wafer-input penalty means higher wafer starts do not convert proportionally to bits; 2026 server shipments +17% YoY with 2027 faster; SOCAMM and rising HBM-per-server lift memory content; consumer shipments decline again in 2027; and CSP capex at record levels with some providers reporting negative free cash flow. NAND, Jul 21 2026: 4–5% supply deficit in 2026, balance turning positive in 2H27; 2026 bit growth came from process migration rather than new fabs because makers prioritised DRAM; Chinese suppliers to ~19% of global NAND bit output; servers >40% of NAND bit demand against ~40% for phones+notebooks; smartphone production −15–20% YoY in 2026, notebooks ~−10%. SEMI: advanced-node capacity passes 1M wpm for the first time in 2026 at 1.16M wpm (this is the plotted value, not an estimate); +69% advanced-capacity growth through 2028; 2nm-and-below <200K wpm (2025) → >500K wpm (2028); 200mm capacity +14% 2023→2026 to a record >7.7M wpm. What was changed and what was not: the 2027 and 2028 bars in chart ① were re-cut — 2027 from a narrowing gap (+29/+26, a 3-point gap) to a widening one (+32/+18, a 14-point gap against 2026’s 12), and the surplus crossover pushed into 2028 (+26/+30). TrendForce publishes no 2027 bit-growth percentages; those bars are this board's translation of its directional language and are tagged proj, not anchored. 2029→2035 were not touched — one quarter's guidance is not a reason to redraw a decade. Chart ②'s 2028 leading-edge point moved 1.40 → 1.44 to sit on SEMI's +69%-through-2028 figure from a 2024 base; SEMI does not state its base year explicitly, so that is a reconciliation, not a quote. The metric conflict, resolved in the open: chart ① reads a 12-point rate gap for 2026 while TrendForce reads −1 to −2% sufficiency. Both are correct and they are not the same measure — rate gap vs level ratio, the latter net of inventory. If 2026 had opened balanced, +35%/+23% would end ~8.9% short; it opened with inventory, so the realised level shortfall is 1–2%. Neither figure should be quoted as "the deficit" without naming which one it is. Provenance & method. Anchored [A]: SEMI World Fab Forecast — total capacity ~29M wpm (2022) → 33.6–33.7M (2025), +6.6–7%/yr; leading-edge ≤7nm 850K (2024) → 982K (2025) → 1.16M (2026) → 1.4M (2028); China ~⅓ of capacity. TrendForce — 2026 DRAM bit demand +35% vs supply +23%; HBM wafer-input share 18%/22%/30% for 2025/2026/2027; conventional DRAM contract +90–95% QoQ in 1Q26; DRAM CapEx $53.7B→$61.3B (2025→26). Samsung/SK hynix Q1 2026 earnings — ~70% DRAM order fulfillment (35–40% for smaller buyers), ~72%/~70% operating margins. Estimate [E]: Citi (demand +20.1%/supply +17.5%), IDC (+16% supply), CoWoS 13K→35K→~130K wpm (2023→24→26, TSMC guidance), NVIDIA >50% of 2026 CoWoS. Triangulated/extrapolated [T]: the 2022–2025 bit-growth pairs (primary industry-total pairs weren't published — reconstructed from the 2022–23 downturn + 2024–25 recovery narrative), all 2027→2035 projections, and the application-segment wafer splits (SEMI segments by device/node, not end-market). Reconciled discrepancy: one source listed HBM 2024 wafer share at ~5%, which is internally impossible (below its ~8% bit share, when HBM uses more wafers/bit) — this dashboard uses the internally-consistent ~3× ratio (HBM ~14% wafers / ~5% bits in 2024), matching §05·B. Unit trap: SEMI reports 200mm-equivalent; actual 300mm DRAM starts (~2.25M/mo in 2025) are ~½ the 200mm-equiv number (~4.5M) — always check the basis. The super-cycle end-date is contested (relief "2027" to "past 2028"); pricing/fulfillment move monthly. Sources: SEMI, TrendForce, IDC, Counterpoint, McKinsey, Yole, SemiAnalysis, Silicon Analysts, company earnings. Analysis, not investment advice. See also: §05·G, which asks what shape this capacity gets sold in — MRDIMM, SOCAMM, DDR6 and 3D DRAM all compete for the wafers counted here. Aug 15 2026 — a bottom-up cross-check on the HBM line [§05·C·H]. This section carries HBM at ~13% of 2027 DRAM bits from a top-down share path. §05·C·H reaches 2027 HBM demand independently, from CoWoS wafer allocations and per-package stack counts across 17 named products, and lands on 6.08 bn GB. The two are consistent at a ~46.7 bn GB total DRAM market. Two methods with no shared inputs agreeing is the strongest validation on this board's HBM numbers — and it applies to the total only, not to any per-vendor split. Aug 17 2026 — capacity by producer [§05·D·V]. This section works in whole-industry totals. Vendor-level DRAM capacity is now carried separately: big-3 1,439k → 3,137k wafers/month 2024→2030 (a 13.9% CAGR, SK hynix steepest at 2.50×), plus CXMT at ~350k by end-2026 rising to 600–950k by 2030. NAND vendor capacity appears on the board for the first time there too — and its shape differs sharply from DRAM, with all three majors stalling or contracting through 2025–27 before a late-decade surge.

05·EDRAM Wafer Starts Required — Standard vs HBM, 2023→2030◷ DB/Gartner model · triangulated
05·E

DRAM Wafer Starts Required (WSPM)

The supply question asked in its native unit: how many wafer starts per month does DRAM bit growth actually require, split standard DRAM vs HBM? This replicates a Deutsche Bank Research / Gartner demand-supply model (user-supplied capture, Jul 2026) — the clearest public decomposition of the mechanism this board keeps describing: HBM's trade ratio converts modest bit growth into brutal wafer growth (HBM needs +48–56%/yr more wafers through 2028 to serve its bits), while efficiencies claw back only ~11–13%/yr. The bottom line is the thesis in one row: a shortfall every single year through 2030, peaking at −795K WSPM (129% demand/supply) in 2028, and — notably — never flipping to surplus in the model's window, even after the 2028 capacity wave lands.

Required WSPM vs estimated capacity — the widening, then narrowing, gap
Stacked bars = required wafer starts (standard + HBM), k WSPM. Line = estimated capacity (2025→2030). The red gap labels are the model's annual shortfall. Hatched years are forecast. All values transcribed from the DB/Gartner table below.
Standard DRAM (k WSPM)HBM (k WSPM)— Estimated capacityShortfall
The full model — replicated table (Gartner · Deutsche Bank Research)
Triangulation — three other houses against this model
Where they agree (strongly): J.P. Morgan independently puts total DRAM WSPM at ~2.8M by end-2028, +880K from end-2025 — an end-2025 base of ~1.92M that matches DB's 1,921K exactly, and a 2028 level within ~1% of DB's 2,769K capacity estimate. JPM's HBM share-of-wafers path (19% → 24% → 28% → 31%, 2025→28) and TrendForce's (18–19% → 22–23% → 30%, 2025→27, §05·B/§05·D) bracket DB's (18% → 23% → 27% → 34%) within 1–3 points through 2027. Three houses, one shape.
Where they diverge (flagged, not merged): (1) 2028 HBM share — DB 34% vs JPM 31%: DB assumes a harder HBM mix-shift. (2) Gap magnitude — TrendForce frames the 2026 DRAM supply-demand gap at ~4.9% (bit basis) vs DB's ~10% (wafer basis, 110% D/S): different denominators, and the wafer basis runs hotter precisely because of the HBM trade ratio — the divergence is the mechanism. (3) Duration — UBS says undersupplied "until at least 2Q28," TrendForce sees "meaningful relief unlikely before 2028," while DB's model stays in shortfall through 2030 (easing to 111% D/S). DB is the most extended-bull view of the set; treat the 2029–30 tail as one house's scenario.

Provenance & validation. [cited — transcribed]: every figure in the table is from the Deutsche Bank Research "Summarized DRAM D-S Forecast" (source line: Gartner, Deutsche Bank Research; user-supplied capture, Jul 2026). [validated programmatically]: the transcription passes its own arithmetic — standard + HBM sums to the printed totals for all 8 years; capacity − requirement reproduces all five printed shortfalls (−210/−507/−795/−575/−379K); demand ÷ supply reproduces 110/121/129/118/111%; the "25–30 growth" column recomputes (+887K standard, +967K HBM, +14%/+12% CAGRs). Reading caveat: the per-year efficiency/yield decomposition rows are small type in the capture; they are approximate multiplicative factors in any case (they don't compose exactly to the printed WSPM growth — DB rounds), so treat the level rows as the data and the decomposition as the explanation. Model caveats: capacity basis is the bank's estimate, not SEMI's (this board's §05·D notes ~2.25M actual 300mm DRAM starts/mo in 2025 vs 1,921K here — scope/basis difference, flagged); "requirements" assume Gartner bit-demand paths; forecasts widen to the right. Cross-refs: §05·B (HBM share), §05·D (wafer S&D), §06·E (bits shipped), KS#2 (§02·D — this model says the 2028 supply wave narrows but does not flip the balance, a direct input to that switch). Sources: Deutsche Bank Research / Gartner (capture), J.P. Morgan, TrendForce, UBS, SEMI. Not investment advice. Aug 15 2026 — what the bottom-up build does and does not test here [§05·C·H]. §05·C·H builds 2027 HBM demand in gigabytes from named silicon. It does not validate this section's 774k HBM WSPM figure, because converting gigabytes to wafer starts requires a GB-per-wafer assumption — die size, stack height, known-good-die yield — that neither source publishes. Anyone bridging the two is supplying that assumption themselves, and it dominates the answer. The demand side now has two independent estimates; the wafer conversion still has none. Aug 17 2026 — the strongest evidence against this section, logged here rather than only in §05·D·V. Vendor-level capacity now on the board puts Samsung, SK hynix and Micron at 3,137k wafers/month by end-2030, and CXMT at a further 600–950k. That is 3,737–4,087k from four named producers, against the 3,396k whole-industry capacity this model uses for 2030 — the modelled total is 10–20% below the sum of four of its own components. Re-running this section's own 2030 requirement of 3,775k against those figures gives 92–101% demand/supply instead of the 111% reported above, which narrows the 2030 shortfall to near zero or turns it into a 312k surplus. This section's "shortfall every year through 2030" conclusion does not survive that substitution, and the disagreement is left standing rather than resolved: the capacity definitions may differ, CXMT's wafer starts are not its usable output, and this model's 2029–30 tail was already labelled the most extended-bull view of the four houses triangulated. See §05·D·V for the full working.

05·D·VVendor Wafer Capacity — DRAM & NAND to 2030◷ Nomura + CXMT sources · Aug 2026
05·D·V

Vendor Wafer Capacity to 2030

§05·D and §05·E both work with whole-industry wafer capacity. Neither has ever carried capacity by producer — and adding it turns up a problem with the board's own numbers. Nomura's vendor build has the big three going from 1,439k wafers/month at end-2024 to 3,137k by end-2030: Samsung 655k→1,264k, SK hynix 474k→1,183k, Micron 310k→690k. Add CXMT — which that exhibit does not cover — at the 600–950k other sources put it on by 2030, and four named producers reach 3,737–4,087k. §05·E's whole-industry 2030 capacity line is 3,396k. The modelled total for the entire industry is below the sum of four of its members. That is not a rounding difference, and correcting it does something uncomfortable: it removes most of the 2030 shortfall this board has been describing.

DRAM wafer capacity by producer — and the line it has to fit under
Stacked wafer starts per month. The 2024 column is Nomura's big three only, because that exhibit does not cover CXMT. The 2030 column adds CXMT as a solid block to 600k with a dashed extension to 950k, since the two figures come from different reports and neither is a company disclosure. The dashed red line is §05·E's modelled capacity for the whole industry in 2030. It sits below the top of the stack — which means either the vendor figures are too high, the industry line is too low, or they are measuring different things. This section does not assume which.
The full vendor table, DRAM and NAND
NAND is the half of this the board has never carried at all. The pattern there is different from DRAM and worth reading on its own: all three producers stall or contract through 2025–2027, then move sharply late in the decade.
What this does to the board's own conclusion
§05·E's headline is that DRAM stays in shortfall every year through 2030, easing to 111% demand/supply. That reading depends on a 2030 capacity of 3,396k against 3,775k of required wafer starts — a 379k gap. Substitute the vendor-level capacity and the conclusion changes. At the low end of the CXMT range the 2030 ratio falls to about 101% — a 38k gap, effectively balance. At the high end it goes to 92%, a 312k surplus. Either way the "shortfall through 2030" framing does not survive intact, and this board is stating that plainly rather than burying it, because §05·E is one of the most bullish sections here and this is the strongest evidence against it that has turned up. Three reasons the gap might not be real, none of which this board can settle: (1) definitional — a vendor's "capacity" may be installed tool capacity while an industry model may use effective or qualified output, and the difference at these volumes is easily 10%; (2) CXMT's usable output is not the same as its wafer starts, given node maturity, yield and the tool-access constraints its own coverage flags; (3) the 2030 tail of §05·E's source is one house's scenario, already labelled as the most extended-bull view of the four this board triangulates. What is not in doubt is the direction: the industry line is too low relative to its own named components, and the years where that matters most are 2029–2030 — precisely where §05·E's shortfall was already narrowing.

Provenance & the honesty line. [cited — Nomura, via a user-supplied summary]: Samsung DRAM 655k → 1,264k wafers/month end-2024 to end-2030 on the P4/P5/P6 phases, y/y growth peaking at +22% in 2030; Samsung NAND fluctuating with dips in 2025 and 2027 before reaching 533k by end-2030 at +29% y/y. SK hynix DRAM 474k → 1,183k on the Y1/Y2/Y3 phases with spurts in 2028 (+25%) and 2030 (+26%); SK hynix NAND contracting 2025–2027 then surging from 2028 (+58% y/y) to 580k. Micron DRAM 310k → 690k with Taiwan scaling 195k → 430k and the balance US/Japan; Micron NAND flat at 130k through 2026 then 280k by 2029–2030 via Fab 10B and Fab 11. [cited — separate sources, not Nomura]: CXMT at roughly 300k today across three 12-inch fabs, ~350k by end-2026 (about 25k below Micron's ~375k), buildouts across Hefei/Shanghai/Beijing taking it "over 600k", and around 950k by 2030; CXMT publicly targets 30% DRAM share by 2030; big-3 revenue share 89.7% with CXMT fourth at 7.6% in 1Q26. [derived — this board's arithmetic, and the reason the section exists]: big-3 2030 DRAM sums to 3,137k; adding CXMT gives 3,737–4,087k; §05·E's modelled whole-industry 2030 capacity is 3,396k, i.e. 341–691k (10–20%) below the sum of four named producers. Re-running §05·E's own 2030 requirement of 3,775k against those totals gives 92–101% demand/supply versus the 111% that section reports. Back-solved, not published: Samsung DRAM ~1,036k and SK hynix DRAM ~939k at end-2029, and Samsung NAND ~413k at end-2029, each derived by dividing the 2030 figure by its disclosed 2030 growth rate — these are consequences of two cited numbers, not source figures, and they are not plotted. What is deliberately absent: (1) no year-by-year path — the source gives endpoints and a handful of growth rates, and interpolating six years between them would manufacture precision that does not exist; (2) no CXMT figure for 2024, because none of the sources here gives one, so the 2024 column is big-3 only and is labelled that way; (3) no bit-level conversion — wafers are not bits, and translating these into supply requires per-node density and yield assumptions nobody publishes. One caution about the source chain: the Nomura figures reach this board through a third-party summary of a research note rather than the note itself, so transcription risk sits on top of forecast risk; the arithmetic above has been checked for internal consistency but the underlying numbers have not been verified against the primary document. Cross-refs: §05·D (whole-industry wafer supply), §05·E (the model this contradicts), §05·B (HBM's share of it), §05·C·H (2027 HBM demand bottom-up), §09 (the China front, where CXMT's constraints are tracked), §02·D (kill-switch #2, the supply response). Not investment advice.

05·FAI Accelerator Wafer Consumption — Six Firms vs the World's Wafer Supply◷ modeled · unit anchors to 2026
05·F

AI Accelerator Wafer Consumption

How many of the world's wafers do Nvidia, AMD, Google, Amazon, Meta and Microsoft actually consume for their AI accelerators? The answer depends entirely on which denominator you use, and the gap between the two is the reason this section exists. Against all wafer starts everywhere, in 2026 they take about 0.26% — a rounding error, roughly a quarter of one percent of global output. Against leading-edge capacity — ≤7nm, the only wafers these chips can be built on — they take about 8%. Same silicon, same year, a 30× difference in the answer. That is the whole argument for why "there are plenty of wafers" and "we are out of wafers" are both true statements, and why §05·D insists the constraint is the thin sliver, not the total. Everything below is a model, not a measurement — nobody publishes accelerator wafer consumption, so this is built bottom-up from unit volumes and die sizes, with every input exposed in a table so you can disagree with a specific number rather than the conclusion.

Wafer starts per month consumed by AI accelerator compute die, 2020 → 2035
Stacked areas, left axis: thousand actual 300mm wafer starts a month, by company. Dashed line, right axis: the six firms combined as a share of global leading-edge (≤7nm) capacity. Everything left of the 2026 marker rests on cited unit volumes — JP Morgan on Blackwell and Rubin, Google's 4.3M TPUs, Morgan Stanley's 1.5M Trainium; everything right of it is extrapolated. The shape worth arguing about is the dashed line turning over. It peaks around 2028 and then declines — not because AI demand slows in this model (units keep compounding) but because §05·D's leading-edge capacity is extrapolated at ~14%/yr, which eventually outruns accelerator growth. That crossover is an artefact of two extrapolations meeting, and it is the least trustworthy thing on this chart. If leading-edge capacity disappoints, the line keeps climbing instead.
Every input, 2026 — disagree with a cell, not with the conclusion
The whole model is units × dies-per-package ÷ (gross dies per wafer × yield). Gross dies per 300mm wafer uses the standard Murray–de Vries estimate, π·150²/A − π·300/√(2A), which charges for the wasted ring at the wafer edge. Change any cell and the chart moves in a way you can predict — that is the point of printing it.
The cross-check, and it does not close — which is itself the finding
There is one published figure that should bound this model, and it is much larger. TSMC's N3 capacity is reported at roughly 180K wpm through 2026, with AI-related wafers "just under 60%" of N3 output, rising to 86% in 2027. Sixty percent of 180K is ~108K wpm on N3 alone. This model puts all six firms' accelerators, across every node, at ~42K wpm. The bottom-up number is well under half the top-down one, and the gap is not obviously an error. Two readings, and this board cannot yet distinguish them. (1) The accelerator die is a minority of "AI silicon." An AI rack is not just GPUs: it is Grace and Vera CPUs, NVLink and Ethernet switch ASICs, retimers, NICs, DPUs, voltage regulators and the logic base die under every HBM stack — plus AI customers outside these six (Broadcom's other clients, Tesla, Cerebras, Chinese accelerators). On this reading the model is correct as scoped and simply measures a smaller thing than TSMC's "AI wafers." (2) The unit or die-size assumptions are low. If real accelerators carry more silicon than the table assumes — bigger dies, more chiplets per package, worse yields — the model understates and the true share of leading edge is 12–20% rather than 8%. What the section claims, given that: the numbers here are a lower bound on AI's leading-edge consumption, because they deliberately count only the compute die of six companies. Read the 8% as a floor, not an estimate, and read the cited 60%-of-N3 figure as the ceiling that includes everything else in the rack. The truth sits between them, and the section is drawn at the end it can actually defend.
Why a quarter of one percent of wafers can still break the market
The 0.26% figure is the one people reach for when they want to argue nothing is scarce, and it is the wrong number. Wafers are not fungible. A 28nm line running power-management chips cannot make a Rubin die, and a leading-edge fab costs $20–30B and takes four years. The binding quantity is the ≤7nm sliver — 1.16M wpm in 200mm-equivalents in 2026, about 516K actual 300mm wafers — and that sliver is what these six firms are bidding for, against Apple's phone silicon and everyone else who needs the newest node. This is the same shape as the memory argument this whole board is built on. HBM is a small share of DRAM bits but consumes a disproportionate share of DRAM wafers (§05·E), because it trades roughly 3× the wafer area per bit. AI logic does the same thing one layer up: a small share of units, a large share of the only capacity that can build them. And the two compete for the same capital. §05·D tracks the wafer allocation decision; §06·F·D tracks the capex funding it; §03·F tracks what happens to consumer prices when the allocation moves. This section is the logic-side companion to §05·E's memory-side version of the same arithmetic.

Provenance — this is a model, and here is exactly what it is made of. [cited — unit volumes]: Nvidia — JP Morgan estimates 5.2M Blackwell in 2025, 1.8M Blackwell + 5.7M Rubin in 2026 (7.5M total, used here); TrendForce separately has Blackwell at 70%+ of Nvidia high-end shipments in 2026 and flags Rubin delay risk, which cuts against the JP Morgan mix. 2023 is anchored to Omdia's H100 estimates (~300K units in Q2 2023, ~500K H100+A100 in Q3). Google4.3M TPUs in 2026, with V6 ~1.6M and V5 ~800K. Amazon — Morgan Stanley's >1.5M Trainium in 2026; AWS said >1M Trainium deployed cumulatively at re:Invent 2025. Market context — custom ASIC shipments forecast +44.6% in 2026 against +16.1% for merchant GPUs. [cited — capacity]: TSMC N3 ~120–130K wpm end-2025 → ~180K wpm end-2026 (+40%), AI ~60% of N3 output in 2026 and 86% in 2027, effective N3 utilisation above 100% in H2 2026; HPC is 61% of TSMC revenue. Leading-edge and total capacity come from §05·D's SEMI-anchored series so the two sections cannot disagree: 2026 total 36.0M wpm and leading-edge 1.16M wpm, both 200mm-equivalent, extrapolated past 2026 at ~7%/yr and ~14%/yr respectively. [modeled — everything else, and the four assumptions that matter most]: (1) Die area and dies per package are estimates, not disclosures — Nvidia has never published a Rubin die size, and the values in the table (A100 826, H100 814, Blackwell 2×800, Rubin 2×730, MI300X-class 1017, MI400 1200, TPU 2×600, Trainium 2×600, MTIA 420→700, Maia 820 mm²) are compiled from teardown reporting and reticle-limit reasoning. (2) Yields of 70–82% are assumptions with no public basis; at reticle-limited die sizes a 10-point yield error moves a company's wafer number ~13%. (3) Unit volumes past 2026 are extrapolation, tapering from double-digit to mid-single-digit growth — no source forecasts accelerator units to 2035, and anyone who did would be guessing too. (4) The 2.25× conversion from SEMI's 200mm-equivalents to actual 300mm wafers is a geometric ratio applied uniformly; SEMI's own basis for "leading-edge" may already be 300mm-only, in which case every share figure here is overstated by 2.25× and the 2026 leading-edge share is ~3.6% rather than ~8%. That single ambiguity is the largest uncertainty in the section and it is not resolvable from public SEMI summaries. What this deliberately does not count: CPUs, network and switch silicon, retimers, NICs, HBM base die, packaging interposers and CoWoS substrate — and every AI chip built by anyone other than these six firms. It is therefore a floor. Not a forecast of anything, and not investment advice. Cross-refs: §05·D (the capacity series this is divided by), §05·E (the identical arithmetic on the memory side), §06·B (the accelerator memory roadmap these dies carry), §06·B·T (Google's TPU generations, one of the six), §05·C·G (HBM wafer consumption per generation), §06·E (accelerator GB shipped), §06·F·D (the capex paying for all of it). Sources: TrendForce on TSMC N3 capacity, TrendForce on Blackwell/Rubin mix, JP Morgan via TweakTown, Morgan Stanley, Omdia, SEMI, Tom's Hardware. Aug 15 2026 — an independent test of this section's 2027 numbers, which it half passes [§05·C·H]. A Morgan Stanley SKU-level build for 2027 totals 21.14M accelerators against this section's 21.30M0.7% apart, from unrelated inputs. That is strong validation of the total. The per-company bars do not survive the same test: Meta differs by 3.1× (1.70M here vs 0.55M there), Microsoft 1.8×, AMD 1.5×, Google 1.5×, and the errors point in both directions, which is why the totals still meet. Read this section's aggregate as well-supported and its company split as one plausible allocation among several — the wafer total is far more reliable than any single company's slice of it.

05·GMemory Form-Factor Roadmap — MRDIMM, SOCAMM, DDR6, LPDDR6, 3D DRAM◿ JEDEC + vendor roadmaps · 2026
05·G

Memory Form-Factor Roadmap

Every other memory section on this board asks how much — how many bits, at what price, from whose fab. This one asks what shape they arrive in, because the package decides which demand a bit can actually be sold into. A DDR5 die that can be built into an MRDIMM competes for the same wafers as one destined for a SOCAMM next to an accelerator, and the two serve completely different buyers at completely different margins. Five transitions are in flight at once, and they are not equally certain: MRDIMM Gen2 is dated and conservative, reusing existing dies at 12,800 MT/s; SOCAMM is shipping, 16 dies on copper wire bonding, with Micron first through Nvidia's qualification; DDR6 arrives at 17,600 MT/s into the worst possible pricing environment; LPDDR6 is being deliberately walked out of phones and into the data centre; and 3D DRAM is the one everybody's roadmap contains and nobody has shipped. The chart below shades each bar by how firm the claim is rather than how large the market might be.

Five memory transitions, shaded by how much is actually known
Solid bars are shipping or disclosed; progressively lighter and dashed bars are vendor-dated guidance, standards-committee work, and finally speculation. The visual point is the fade: read left to right and the roadmap dissolves from product into intention somewhere around 2028. 3D DRAM is drawn faintest on purpose — it appears on every major vendor's public roadmap and has no shipping product, no disclosed yield and no announced customer, which is exactly the profile of a technology that can slip five years without anyone formally cancelling it. Hover a row for what each vendor has actually said.
Why form factor is a supply story, not a spec-sheet story
The reason this belongs next to §05·D rather than in a product roundup is that these transitions compete for the same finite wafer output. §05·E's arithmetic holds: an HBM bit consumes several times the wafer area of a commodity DDR5 bit, so every stack diverted to an accelerator removes more than one bit-equivalent from the server and PC pools that §03·F and §08·G are watching get repriced. SOCAMM makes that competition worse in a way HBM alone did not. It is built from LPDDR — the mobile process line — which means AI demand has now reached into a capacity pool that was previously insulated from it by having nowhere else to go. MRDIMM is the counterweight: it raises effective bandwidth without changing the die, so it is the one transition here that adds performance without adding wafer demand. And DDR6's timing is genuinely unfortunate — a platform transition normally needs cheap memory to drive adoption, and 2027 is not currently forecast to supply any. Whether DDR6 slips because of pricing is a testable prediction this board will be able to score.

Provenance & the honesty line. [cited]: MRDIMM Gen2 dated to 2026–2027 at 12,800 MT/s. SOCAMM connects 16 DRAM dies per module using copper wire bonding for thermal reasons; Nvidia commissioned Samsung, SK hynix and Micron to build prototypes and Micron was first approved for mass production; SOCAMM2 moves the module to LPDDR6 and SK hynix places it in the late 2020s. LPDDR6 (JESD209-6) was published by JEDEC in July 2025 at 10,667–14,400 MT/s; in April 2026 JEDEC previewed an extension of LPDDR6 beyond mobile into data-centre and accelerated computing, with 512 GB densities, a processing-in-memory variant and SOCAMM2 all in development — none of which has a published release date, which JEDEC itself stated. DDR6 quoted at 17,600 MT/s, debuting in high-end parts in 2026 with mass adoption in 2027. 3D DRAM: Samsung guiding to an 8–9nm class node in 2027–2028; SK hynix has published a DRAM roadmap running to 2031 covering DDR6, GDDR8, LPDDR6 and 3D DRAM. What is modeled: only the start and end positions of the bars. Vendors give years, sometimes only halves of years; the chart converts those into segments and therefore implies a precision the sources do not have. Read the bar edges as ±2 quarters at best, and worse the further right they sit. What is explicitly not claimed: any market share, unit volume or revenue for any of these form factors. Three caveats: (1) roadmap dates are marketing artefacts as much as engineering ones — every date here comes from a party with an interest in it being believed, and none has been independently confirmed; (2) the 3D DRAM bar should be read as a placeholder, not a forecast — a technology on every roadmap with no shipping part, no disclosed yield and no named customer is a technology whose schedule has not yet met reality; (3) DDR6's 2027 volume date and §03·F's pricing are on a collision course that neither source acknowledges, and this board is flagging the tension rather than resolving it. Cross-refs: §05·D (the capacity all of this competes for), §05·E (the wafer-area arithmetic that makes it a zero-sum fight), §03·F (the contract pricing DDR6 launches into), §06·B (the accelerator memory roadmap), §08·G (where a DDR6 platform transition reaches consumers), §05·F (the accelerators SOCAMM attaches to). Sources: JEDEC, Samsung, SK hynix, Micron, Nvidia, Tom's Hardware, TechPowerUp, VideoCardz. Not investment advice.

06Accelerator Memory◷ Jun 2026
06

Accelerator Memory

The demand engine, by vendor: every major AI accelerator and all the memory on its package or module. Each bar is stacked by memory type — HBM (by generation), on-die SRAM, GDDR, plus the LPDDR5X system memory on Nvidia's Grace/Vera superchips. The merchant GPUs (Nvidia, AMD) and the hyperscalers' own silicon (Google, Amazon, Microsoft, Meta), plus wafer-scale Cerebras and RISC-V Tenstorrent, are all here. Memory per device — not unit volume — is what's exploding.

From 16 GB → 2 TB+ of memory per accelerator

Eight companies now design AI accelerators. HBM is the common thread, but the full memory picture is richer: Nvidia's Vera Rubin superchip pairs 576 GB of HBM4 with up to 1,536 GB of LPDDR5X system memory; Cerebras goes the opposite way with 44 GB of blisteringly fast on-die SRAM and no HBM; Tenstorrent uses cheaper GDDR6. Top-end HBM has grown from the H100's 80 GB (2022) toward 1 TB+ per package by 2027–28. Multiply across every vendor's roadmap and the wafer math becomes overwhelming.

HBM3 HBM3E HBM4 On-die SRAM GDDR6 LPDDR5X (system memory)

Bars are stacked by memory type and scaled to total on-package/module memory — so the LPDDR5X-heavy superchips (GB200/GB300/Vera Rubin) show their full memory pool, and the "GB" figure on each card is the total of all memory types. The chronological view, per-GPU breakdown, and rack-level totals are in the roadmap below. Future/announced parts are marked. Hyperscaler ASIC specs (Google, Amazon, Microsoft, Meta) and the alt-architectures (Cerebras SRAM, Tenstorrent GDDR) are disclosed less fully than merchant GPUs; some capacities are approximate. Sources: Nvidia, AMD, Google, AWS, Microsoft, Meta, Cerebras, Tenstorrent, SemiAnalysis, ServeTheHome. See also: §06·B for the memory roadmap each of these accelerators carries, and §06·B·T for Google's TPU line specifically.

06·BAccelerator Memory Roadmap◷ Jun 2026
06·B

Accelerator Memory Roadmap

Every major AI accelerator in release order — Nvidia, AMD, Google, Amazon, Microsoft, Meta, plus alt-architectures Cerebras and Tenstorrent — with the type and amount of all memory on the package or module: HBM, on-die SRAM, GDDR or LPDDR5X. Each card lists peak compute, and rack-scale systems show their full pooled memory. Use the controls below to switch view (per-module vs per-GPU), filter shipping vs announced parts, or switch to log scale so the older & smaller parts stay legible alongside Rubin Ultra & Feynman.

HBM3 HBM3E HBM4 / HBM4E On-die SRAM GDDR6 LPDDR5X (CPU system memory)
View
Show
Scale Log scale keeps older & smaller parts legible against Rubin Ultra / Feynman.
Total HBM Per Rack / Scale-Up Unit
How much high-bandwidth memory each vendor's rack-scale system pools in one coherent domain. Log scale — Google's 9,216-chip superpod dwarfs everything. (TB; superpod shown at 1,770 TB = 1.77 PB.)

Notes: bandwidth figures are per-GPU package (HBM). "Per superchip" rows include the CPU's LPDDR5X because it ships soldered/socketed on the same module and is coherently shared with the GPU(s) over NVLink-C2C. Standalone GPU rows (B200, MI-series) list HBM only — system DRAM is configured separately by the system builder. Capacities reflect top-bin configurations; salvaged/cut-down SKUs ship with less. Hyperscaler ASICs (Google TPU, AWS Trainium) use HBM but disclose less detail. Sources: Nvidia, AMD, Google, AWS, SemiAnalysis, ServeTheHome.

06·B·TGoogle TPU Generations — Memory & Customers (Broadcom-built)◷ v5–gen8 real · v9+ projected
06·B·T

Google TPU Generations — Memory & Customers

A naming correction first, because it matters: these are Google's TPUs. Broadcom is Google's co-design and ASIC partner — Google owns the architecture and software stack; Broadcom supplies the foundational IP, SerDes/interconnect, and silicon implementation, then coordinates fabrication at TSMC. The two have shipped seven generations together since 2014, under a supply agreement now extending to 2031. Broadcom separately builds custom "XPU" accelerators for other customers (below). So "Broadcom's TPU" is really "the Google TPU that Broadcom helps build." The HBM-per-chip column is the reason TPUs matter to this dashboard: each generation is a growing customer for the same HBM that's in shortage.

HBM per TPU chip, by generation
Real, sourced specs through the 8th generation; everything below the divider is unannounced and modeled. "Memory/chip" is on-package HBM. ↑ multiplier vs the prior comparable generation.

The trajectory that matters: TPU HBM per chip ran 16 GB (v5e) → 95 GB (v5p) → 32 GB (v6e) → 192 GB HBM3E (v7 Ironwood) → 288 GB (gen-8 inference). Ironwood alone is 6× the HBM of Trillium. Multiply by millions of chips deployed (see §06·C) and Google is one of the largest single consumers of HBM on earth — most of it bought from SK hynix and Samsung, the same suppliers serving NVIDIA.

Broadcom's custom-silicon customers

Broadcom disclosed six named XPU customers as of Q1 FY2026, anchoring a >$100B AI-chip revenue target for 2027 against a ~$73B backlog. TPUs (Google) are one program; the others are distinct custom accelerators. Anthropic's compute runs on Google TPUs, so it appears both as a TPU consumer and a Broadcom customer.

Provenance. Real specs (sourced): TPU v5e 16 GB HBM, v5p 95 GB HBM (7nm), v6e/Trillium 32 GB HBM, v7/Ironwood 192 GB HBM3E @ 7.37 TB/s (6× Trillium, GA late 2025), 8th-gen split into 8t (training) and 8i (inference, ~288 GB) for 2026 — per Google Cloud docs/blog, Introl, Spheron, CloudOptimo, Nevsemi. Customers: Broadcom Q1 FY2026 earnings named six XPU customers — Google (TPU, 7 gens since 2014, deal through 2031), Meta (MTIA), Anthropic (~1 GW TPU in 2026 → ~3 GW in 2027, via Google TPUs), OpenAI (first XPU 2027, 10 GW multi-year deal, 3nm+2nm), plus Apple (new 2026 disclosure) and ByteDance/Fujitsu among named/reported — per Tom's Hardware, CNBC, Jon Peddie, hashrateindex. Projected / unannounced: a 9th generation is modeled in §06·C, but "TPU v9, v10, v11" are not public products — no announced specs, memory, or customer assignments exist. The HBM figures for those rows are illustrative extrapolations (continuing the roughly-doubling-every-2-gens trend), not Google roadmap data, and are visibly hatched. Anyone citing a "TPU v11 memory spec" is citing speculation. Not investment advice. Aug 15 2026 — the scale of this line, in HBM [§05·C·H]. A bottom-up 2027 build puts Google's four TPU generations at 2.19 bn GB of HBM — 36.0% of all AI HBM demand, and 96.6% of NVIDIA's. On this estimate a single customer's internal silicon consumes almost as much HBM as the entire merchant GPU leader. Carry one caveat with it: §05·C·H's unit count for Google (8.17M) is 1.46× §05·F's independent estimate (5.60M), so the ranking is robust but the exact figure is not.

06·CAccelerator Shipments◷ annual est.
06·C

Accelerator Shipments

◷ annual est.

How many data-center AI accelerators ship — the volume driving all that memory demand. Important honesty note: vendors don't disclose unit shipments, and the in-house hyperscaler chips (TPU, Trainium, Maia, MTIA) aren't sold commercially. So none of this is reported data — it's the most honest available reconstruction from analyst estimates, rack-shipment reports, and revenue. Below: annual totals (Nvidia vs others), then a more speculative quarterly-by-vendor breakdown and a per-specific-accelerator view (B300, GB300, TPU v7 Ironwood, MI350, Trainium 3…). Every figure is an analyst triangulation with a wide error band — the further into per-quarter/per-chip detail, the more uncertain.

Est. Data-Center AI Accelerator Units Shipped Per Year
Millions of units. Nvidia figures anchored to Omdia data-center GPU estimates (~2.6M in 2022, ~3.8M in 2023); later years and the "others" bucket (AMD + Google TPU + AWS Trainium + Microsoft/Meta ASICs) are analyst consensus. 2026 is a forecast.
Nvidia (data-center GPUs) All others (AMD + hyperscaler ASICs)
Quarterly Units by Vendor — estimated

A quarterly breakdown by vendor, stacked. This is the most speculative chart in the dashboard — quarterly per-vendor unit counts aren't disclosed by anyone, so every bar is a triangulation from rack-shipment estimates, revenue, and supply-chain reports. The shape (Nvidia dominant but custom-ASIC vendors ramping fast) is well-supported; the exact numbers are not. Click a vendor to toggle.

Millions of data-center accelerator units per quarter, by vendor. Q3 2025 onward includes forecast quarters. Counts GPUs/ASICs (not racks): a GB200/GB300 NVL72 rack = 72 GPUs.
By Specific Accelerator — estimated 2026 units

Approximate full-year-2026 unit estimates for the specific accelerators driving demand. Bars scaled within the whole set. Hyperscaler-ASIC volumes (TPU, Trainium, Maia, MTIA) are internal-deployment estimates; merchant-GPU volumes derive from rack and revenue reports. All are analyst triangulations, not disclosed figures.

Per-Model Shipments by Quarter — by Vendor

Every data-center accelerator from each of the six major vendors, stacked by quarter from 2022 to 2030 — read both the running total and how each generation hands off to the next. These are the dashboard's most speculative charts: per-model quarterly units aren't disclosed by anyone, and beyond mid-2026 it's projection. The wave shape is the insight; the numbers are illustrative. Pick a vendor, toggle Units/Share, click models in the legend to hide them.

On the per-model quarterly charts (2022–2030): these are the most heavily modeled visuals in the dashboard — a reconstruction, not reported data. No vendor discloses per-model unit shipments by quarter, and roughly half this timeline (Q3 2026 →) is a forward projection. Each model is shaped as a lifecycle curve (ramp from launch → peak → decline as the next generation takes over), calibrated so annual totals stay near the figures used elsewhere on this page and so generational hand-offs land at the right times. Real timing anchors used: Nvidia — H100 late 2022, Blackwell late 2024, GB300-over-GB200 in 2H 2025, Vera Rubin in full production mid-2026, Rubin Ultra ~2027–28, Feynman ~2028+; AMD — MI300X late 2023, MI350 in 2025, MI400/Helios in 2026, MI500 ~2028; Google — TPU v7 Ironwood announced Apr 2025 / GA late 2025–26 (internal forecast revised ~2M→~4M units; Anthropic contracted up to 1M TPU chips in 2026, 400K Ironwood in phase one), 8th-gen TPU 8t/8i on TSMC 2nm previewing 2H 2026 with GA targeted late 2027; Amazon — Trainium3 (first 3nm, 144 GB HBM3e) shipping Dec 2025, Trainium4 (NVLink Fusion) due late 2026/early 2027, with ~35% of AWS's ~$100B 2026 capex earmarked for Trainium-family silicon; Meta — MTIA v1 announced 2023, v2 "Artemis" in 2024, with a 300-series roadmap extending through ~2027 (inference-first, internal to Meta's fleet); Microsoft — Maia 100 announced Nov 2023 (limited deployment), Maia 200 (3nm, 216 GB HBM3e) announced Jan 2026 (the least-mature hyperscaler program). Note the hyperscaler chips (TPU, Trainium, MTIA, Maia) are never sold as parts — they're deployed internally and rented via cloud — so their "units shipped" are even more inferred than merchant GPUs (anchored via Broadcom/Marvell rack revenue and cloud-capacity disclosures). Peak heights are anchored to those sources, but exact per-quarter splits are interpolated and the projected generations (Rubin Ultra, Feynman, TPU 9th-gen, Trainium5, MTIA next-gen, Maia next-gen) are scenario, not forecast — real volumes could differ by a wide margin. Read the generational shape, not the precise unit counts. Units are individual chips, not racks or revenue. The four hyperscaler programs differ hugely in scale — Google and Amazon ship multiples of Meta and Microsoft — and summing all six charts gives a total in the same ballpark as, but not identical to, the high-level annual figures elsewhere on this page (they're independent estimates built different ways). Sources: Nvidia/AMD filings & keynotes, Google Cloud/Broadcom, AWS, Meta, Microsoft/Azure, SemiAnalysis, Morgan Stanley, Omdia, TrendForce, Tom's Hardware, The Next Web. Not investment advice.

06·DToken Economics — Demand & Price◷ demand modeled · + MW-to-serve · price proprietary
06·D

AI Token Demand

◷ quarterly est.

The root of the whole supercycle: the explosion in AI tokens processed — the raw workload that every accelerator and memory chip ultimately exists to serve. Honesty note: there's no authoritative global token counter. But several providers now disclose real figures — Google especially (9.7T tokens/month in early 2024 → 3.2 quadrillion/month by May 2026), OpenAI's API (6B → 15B tokens/min, Oct 2025 → Mar 2026, ≈650T/mo), plus Microsoft Foundry and others. The bars below estimate total global token run-rate (tokens processed per month, at each quarter-end), built up from those disclosures plus analyst aggregates. The linear scale shows just how brutally recent and vertical the climb has been — the pre-2025 bars almost disappear against the right-hand side, which is the truthful read of an exponential.

Est. Global AI Tokens Processed Per Month — to 2028
Trillions of tokens/month (T), rising into quadrillions (Q = 1,000T). Solid bars are history (orange = a hard provider disclosure); hatched bars are the modeled projection to Q4 2028. On a linear scale the projection still dwarfs the early years (today is ~18% of the 2028 bar) — true to scale, but it crushes the 2022–2024 history; switch to log to read the full trajectory. ChatGPT launched Nov 2022.
Scale
Estimated run-rate (history) Hard disclosure Modeled projection
Megawatts required to serve that demand — H100 · H200 · B200 · B300 DERIVED · ASSUMPTIONS DISCLOSED
The chart above counts tokens. This one converts them into the constraint that actually gates the build-out: electricity. Each line answers a deliberately counterfactual question — if the entire world's token demand that quarter ran on this one accelerator, how much continuous power would it draw? That is why the lines run parallel on the log axis: the ratios between parts are fixed, so the gaps are the generational efficiency gains. The arithmetic is MW = (tokens/sec ÷ tokens/sec per GPU) × all-in kW per GPU, with token demand taken from the series above and every input listed in the provenance. This is a floor, not a fleet plan: it assumes 100% utilisation, no redundancy, no training, no idle capacity — real deployments need several times more.
Why the gaps exist — inference efficiency per kilowatt (all-in)

tokens/second per kilowatt of facility power (GPU + CPU + networking + cooling + PSU losses), indexed to H100

The reason this chart matters to a memory dashboard. Power is the binding constraint on AI build-out (§06·F·C tracks ~31.5 GW of committed compute), and the way you escape it is not more watts — it is more tokens per watt, which comes overwhelmingly from memory bandwidth and capacity per package, not from raw FLOPS. Decode is memory-bound (§06·D·W): a B300 beats an H100 on tokens-per-watt mainly because it streams weights and KV faster and holds more of them on-package. So every efficiency gain on this chart is bought with HBM — which is precisely why serving more tokens on less power still means buying more memory, not less. The power curve bends; the memory bill does not.

Provenance & the honesty line — this is a model, and here is all of it. Demand [from §06·D above]: tokData, trillions of tokens per month, anchored to Google/Azure/OpenAI disclosures and projected past Q3 2026; converted to tokens/second at 30.44 × 86,400 s per average month. Power [cited]: H100 SXM and H200 SXM are 700 W TDP, B200 1,000 W, B300 (Blackwell Ultra) 1,400 W. The chart uses all-in facility power ≈ 1.7 × GPU TDP (1.19 / 1.19 / 1.70 / 2.38 kW respectively) to cover host CPU, NVLink/NIC, cooling and PSU losses — the same NVL72-class basis used in §06·F·C (~120 kW per 72-GPU rack ≈ 1.7 kW per accelerator all-in). Throughput [derived from benchmarks that disagree — the weakest link]: central estimates of 2,200 / 2,860 / 7,000 / 10,500 tokens/sec per GPU. The published spread is wide and is not hidden here: one power-aware benchmark measures H200 at 2,418 tok/s vs H100's 2,191 (a ~10% gain) while long-context tests report H200 at 1.83–2.14× H100; B200 is variously ~2.5× H200 or "up to 47% higher output-token throughput at peak concurrency". This build takes mid-range values (H200 ≈ 1.3× H100, B200 ≈ 2.4× H200, B300 ≈ 1.5× B200 on KV headroom — B200 and B300 share 8 TB/s, so B300's edge shows up at concurrency and long context, not raw decode). Change these assumptions and the levels move materially; the ordering does not. A disagreement worth stating: this bottom-up model puts B300 at ~2.4× H100 tokens-per-watt, while NVIDIA claims GB300 NVL72 delivers 5× throughput per MW vs Hopper (and up to 50× in agentic framings combining latency and power, with 61.4K concurrent agents/MW vs 2,600 on H200). Rack-scale NVL72 with FP4 and NVLink-domain batching plausibly beats a per-GPU model — so the B300 line here is likely conservative (too much power), and the vendor figure is likely optimistic. Both are shown rather than averaged. Four structural caveats: (1) counterfactual, not forecast — no real fleet is single-SKU; the actual installed base is a mix, so no line is "the" answer. (2) 100% utilisation floor — real serving carries peak headroom, redundancy, failed nodes and idle time; multiply by 2–4× for a deployed estimate, and this excludes training entirely. (3) tokens are not fungible — a reasoning token, a cached prefill token and a short chat completion cost wildly different amounts of compute (§06·D·W); a single tok/s number averages across all of it. (4) PUE and cooling vary by site; 1.7× is a mid-range convention, not a measurement. Cross-refs: §06·D·W (why decode is memory-bound), §06·F·C (GW commitments), §05·E (the wafer consequence). Sources: NVIDIA (H100/H200/B200/GB300 specifications and Blackwell Ultra performance claims), SemiAnalysis InferenceX, AIMultiple multi-GPU and concurrency benchmarks, Introl, Spheron, Lyceum Technology, plus this board's own token series. Derived estimate — not a measurement, and not investment advice.

Share of all data-centre power spent processing tokens — by quarter DERIVED · TRIANGULATED AGAINST IEA
The chart above gives token-serving power in absolute megawatts. This one divides it by the whole industry — every data centre on earth, AI and otherwise — to answer the question that actually matters for the build-out: how much of global data-centre power is now going into generating tokens? The denominator is global data-centre electricity converted to an average continuous load (IEA: 415 TWh in 2024 → 945 TWh by 2030, which is ~47 GW → ~108 GW of continuous draw). The shaded band spans the two fleet extremes — everything on B300-class silicon at the bottom, everything on H100-class at the top — because which silicon the work runs on moves the answer by 2.4×. Both lines carry the ×3 deployment factor §06·D·P argues for (utilisation, redundancy, headroom); the dashed grey line is the raw 100%-utilisation floor beneath it.
The number that makes this chart worth trusting. This board built its estimate bottom-up — from token counts, measured GPU throughput, and wattage — with no reference to anyone's top-down figure. The IEA, working entirely top-down from national electricity statistics, puts AI at 5–15% of data-centre power in recent years, rising to 35–50% by 2030. This model lands at ~4–9% today and ~17–40% by end-2028. Two methods that share no inputs land on the same trajectory — which is the strongest form of validation available here, and considerably more reassuring than either number alone. Where they'd diverge is instructive too: the IEA counts all AI compute including training, while this chart counts inference token-serving only, so this model should sit at or below the IEA band — and it does.

Provenance & the honesty line. Numerator [derived — same model as §06·D·P]: tokens/sec from the demand series above ÷ tokens/sec per GPU × all-in kW; the efficient bound uses B300-class (10,500 tok/s at 2.38 kW all-in), the legacy bound H100-class (2,200 tok/s at 1.19 kW). Every assumption and its benchmark spread is documented in §06·D·P and is not re-litigated here — but it carries forward, and it is the largest uncertainty in this chart. Denominator [cited]: IEA — global data-centre electricity consumption of 415 TWh in 2024 (~1.5% of world electricity), growing ~12%/yr since 2017 and ~15%/yr from 2024, reaching 945 TWh by 2030 (base case) and 1,200 TWh by 2035; converted here to average continuous gigawatts (TWh × 10¹² ÷ 8,760 h), giving 37.8 GW (2022) → 82.9 GW (2028). Total installed global capacity is separately reported near 100 GW, which is higher than average load because facilities do not run flat out — this chart uses the load basis so numerator and denominator match. Deployment factor [modeled]: ×3, the midpoint of the 2–4× range §06·D·P argues for; at ×2 the current share falls to ~2.6–6.2% and at ×4 it rises to ~5.2–12.5%, so this single assumption moves the answer by about 2× and is stated rather than buried. Five caveats: (1) Inference only — training, fine-tuning and non-token AI workloads are excluded, so this is a floor on "AI share", not a measure of it. (2) No real fleet is single-SKU; the true answer sits inside the band, not on either edge. (3) The denominator is a projection past 2024 — the IEA's own base case, and other houses differ. (4) Quarterly resolution is synthetic: the IEA publishes annual figures, so each quarter within a year shares that year's denominator, which is why the lines step rather than curve. (5) Tokens are not fungible (§06·D·W) — a reasoning token and a cached prefill token cost wildly different power, and a single tok/s figure averages across all of them. Cross-refs: §06·D·P (the absolute megawatts), §06·D·W (why decode is memory-bound), §06·F·C (committed gigawatts), §06·F·L (who physically wires it). Sources: IEA Energy and AI (2025) and its data-centre demand analysis, S&P Global, Carbon Brief, Scientific American, plus this board's token series and the accelerator assumptions in §06·D·P. Derived estimate — not a measurement, and not investment advice.

By Provider — modeled share of global tokens
Only Google is anchored to a disclosed token figure (3.2 quadrillion/mo, May 2026); OpenAI's API throughput (~15B tokens/min ≈ 650T/mo, Mar 2026, up from 6B in Oct 2025) is partially disclosed. The other five providers are modeled from proxies — monthly active users, web-traffic share, revenue and usage intensity — not reported token counts. Each provider's confidence is flagged in the legend. Default view is share % because the share is the modeled quantity; the absolute view inherits the same linear-scale flattening of the early years.
View
Per-provider basis — Google: disclosed token volume, but it counts AI Overviews across ~2B Search users, which inflates its share versus chat-only peers. OpenAI / ChatGPT: ~15B tok/min API (≈650T/mo, Mar 2026 — up from 6B in Oct 2025) + ~900M weekly-active consumer estimate; still leads consumer token consumption. Anthropic / Claude: ~7–9% and rising — a $30B revenue run-rate (Apr 2026, up from ~$9B end-2025), weighted up for token-heavy coding (Claude Code). Meta AI: ~1B MAU but light, free, casual usage. Microsoft Copilot / Foundry: ~420M MAU; Foundry token throughput +30% QoQ (300+ customers past 1T tokens/yr) — but its tokens often run OpenAI/Anthropic models, so there's definitional overlap. xAI / Grok: bundled with X, fast-growing from a small base. Open-source / local: the roughest figure — distributed across clouds and private devices, fundamentally unmeasurable. Shares are normalized to the anchored global total; the split is a model, not disclosed data.
Memory Required to Serve That Demand — by quarter
Approximate HBM (accelerator memory) needed to serve the token run-rate, in exabytes (EB = Bn-GB). Serving tokens needs GPUs, and each GPU's value for inference is mostly its HBM — holding model weights plus the KV-cache for every concurrent request. This converts tokens/mo → sustained tokens/sec → the fleet of accelerators (and thus HBM) needed to serve it. This is a derived model resting on one big efficiency assumption, shown below and adjustable in your head: if tokens/GPU/sec doubles (better models/hardware), the memory need halves.
Efficiency assumption
HBM needed (actual era) HBM needed (modeled / projection) exabytes
The conversion chain, made explicit — tokens/mo ÷ (30×24×3600) = tokens/sec; ÷ tokens/GPU/sec = GPUs needed; × HBM/GPU = total HBM. The base case assumes a sustained ~2,500 output-tokens/sec per high-end accelerator (blended across batching, prefill/decode, model sizes) and ~250 GB HBM per accelerator (a Blackwell-class average rising toward Rubin's 288–576 GB). Real serving is far messier — KV-cache grows with context length, MoE models activate only some weights, prefill and decode have very different memory profiles, and a large share of "tokens" are cheap small-model calls. Treat this as an order-of-magnitude sanity check on why token growth forces HBM growth — not a procurement model. It deliberately estimates active serving HBM, which is only part of total memory shipped (see §06·E).
Scenario builder — stress-test the 2028 memory need [interactive]
Drag the three assumptions; the active-serving HBM recomputes live. Same chain as the chart above (tokens/mo → tokens/sec → accelerators → HBM), exposed so you can test it yourself. If efficiency doubles, the memory need halves.
Accelerators needed
Active-serving HBM
Illustrative order-of-magnitude only — active-serving HBM, not total memory shipped (§06·E). 1 Q = 1,000 T tokens; 1 EB = 1,000 PB = 10⁹ GB. Real serving is messier (see the chain note above). Not investment advice.

Real disclosed anchor points behind the token estimate: Google — 9.7T/mo (Apr 2024), 480T/mo (May 2025), 980T/mo (Jul 2025), ~1.3 quadrillion/mo (Oct 2025), 3.2 quadrillion/mo (May 2026, I/O). Microsoft Azure — 100T+/quarter through 2025, a record 50T in March 2025. OpenAI — API throughput 6B tok/min (Oct '25) → 15B tok/min (Mar '26) ≈ 650T/mo, plus ~900M weekly active users. Microsoft Foundry — 300+ customers past 1T tokens/yr, accelerating +30% QoQ (FY26 Q3). Anthropic — $30B revenue run-rate (Apr '26, from ~$9B end-'25), coding-heavy token mix. A mid-2025 cross-provider aggregate (Tunguz) put the world at ~88 trillion tokens/day ≈ 2.6 quadrillion/month. Projection to 2028: the run-rate is extrapolated from 7 quadrillion/mo (mid-2026) on a decelerating growth curve (~30%/qtr easing toward ~15%/qtr) to ~54 quadrillion/mo by Q4 2028 — roughly 8× over two-plus years. That is an aggressive model, not a forecast; demand could stall on a capex pause or compound faster — the error band widens every quarter to the right. The memory-required chart is a second-order model built on top of the token model (token assumptions × serving-efficiency assumptions), so its uncertainty is compounded — use it for the shape of the argument, not the level. Tokens/month run-rate, not cumulative. Sources: Google I/O & earnings, Microsoft earnings, OpenAI, SemiAnalysis, Tomasz Tunguz, First Page Sage, Similarweb, company disclosures. Not investment advice.

06·D·PToken Price Index — Silicon Data SDLLMTK◷ Jun 2026
06·D·P

↓ Token Price Index — Silicon Data SDLLMTK

Silicon Data publishes the LLM Token Expenditure Index (ticker SDLLMTK) daily on the Bloomberg Terminal — an expenditure-weighted average of what the whole market pays per million LLM tokens, across frontier APIs, open-weight platforms, and brokered/self-hosted inference (>90% of global inference spend). It's a proxy for the market's marginal willingness to pay for AI, and one macro strategist (Andreas Steno Larsen, Jun 9 2026) called it "the most important chart for the entire market" — because if token spend rolls over, the memory/GPU/datacenter trade this whole dashboard tracks loses its demand engine. Critical honesty note: the actual daily index values are a paid Bloomberg/Silicon Data product — I can't reproduce the proprietary series. The curve below is a schematic reconstruction of the index's publicly-described shape, pinned to a handful of cited public reference points. Treat it as the story of the index, not its values.

Two curves, one paradox — per-token price vs total expenditure index
The whole point is the divergence: a single token's price has fallen >90% since 2023 (Apollo's Torsten Slok), yet the expenditure index has roughly doubled since late 2025 as usage exploded — Jevons' paradox: cheaper tokens → far more agents/workflows → higher aggregate spend. Schematic shapes pinned to cited public points (●); both axes are indexed/normalized, not exact. Recent reading ~2.1 (early-2026 spike, then a downtick).
Expenditure index (SDLLMTK, left)Per-token price, log scale (right)● cited public reference point

Provenance & the honesty line. SDLLMTK is a real, proprietary index — Silicon Data, published daily on Bloomberg Terminal; full history is a paid product (sales@silicondata.com). I have not reproduced the actual series (I don't have the values), so the chart is a schematic of the publicly-described shape, anchored to these cited public statements: per-token price down >90% since 2023 and expenditure ~2× since late 2025 (Torsten Slok / Apollo, via X, Jun 2026); index "more than doubled since December," rose sharply through May 2026, then a downtick/stagnation (Seeking Alpha citing Andreas Steno Larsen, Jun 9 2026; corroborated by Citadel Securities "Tokenomics," Jun 2026, and 36Kr); a public reading ~2.12 with May acceleration (third-party chart commentary, mid-2026). The "stagnation" is Silicon Data's own framing — possibly slowing migration to premium closed models, not a confirmed reversal. The index is expenditure/usage-weighted price (marginal willingness to pay), not token volume and not simple list price; Silicon Data notes it would more precisely be called a "Token Expenditure Price Index." Why it's here: token billing ties AI usage directly to GPU-hours, DRAM bandwidth, and datacenter demand, so a durable roll-over would be an early warning for the memory thesis — exactly the leading indicator this dashboard exists to watch. For the real-time series, see Silicon Data / Bloomberg SDLLMTK. Schematic of a proprietary index. Not investment advice.

06·D·EMemory Demand by End-Use — Who Consumes Each Type (2010→2030)◷ modeled share · anchored reference points
06·D·E

Memory Demand by End-Use

One story runs through all four memory types: consumers built the market; data centers took it over. Each chart below splits 100% of that memory type's annual supply by end-use application, from 2010 to today with a projection to 2030. In 2010 every one of these was dominated by PCs, phones, or consumer gadgets. By the mid-2020s — and overwhelmingly by 2030 — the data center is the center of gravity for all of them. Honesty note: like the token-share chart above, these are modeled shares — smooth reconstructions anchored to cited industry reference points (marked in the captions), not a continuously-reported annual series. The dashed line marks where history ends and projection begins.

DRAM

By bit consumption. Anchors: server overtook mobile in 2023 (~37.6% vs 36.8%, TrendForce); HBM ~5% of bits 2024; AI/HBM ~20% of DRAM wafers by 2026.

NAND Flash

By bit consumption. Anchors: smartphones ~40% / SSD ~25% in 2023; enterprise SSD becomes the #1 segment in 2025–26 (TrendForce).

HBM

By bit consumption. HBM didn't exist before 2015 (AMD Fiji). Anchors: AI/ML ~55%+, HPC ~25%, graphics ~12%, emerging ~8% in 2026 (PatSnap/JEDEC).

HDD (Storage)

By capacity (exabytes) shipped. Anchors: nearline ~54% of capacity today → >90% by 2029 (Coughlin); client/consumer collapsing as PCs move to SSD.

Provenance & the honesty line. These four charts are modeled shares — the same epistemic class as the "By Provider" token chart above. No vendor publishes a continuous 2010–2030 annual end-use breakdown by bit/exabyte, so each series is a smooth reconstruction fitted to cited anchor points and the well-documented qualitative arc. Anchors used — DRAM: server DRAM overtook mobile in 2023 at ~37.6% of bits (mobile 36.8%), server the largest segment since (TrendForce); HBM ~5% of DRAM bits / ~20% of revenue in 2024 (TrendForce); AI projected to consume ~20% of DRAM wafer capacity in 2026 (Commercial Times/TrendForce); the 2010–2015 mobile surge and PC decline are well-documented. NAND: smartphones ~40% and SSD ~25% of demand in 2023 (industry reports); enterprise SSD becomes the single largest NAND application segment in 2025–2026 on AI-server demand, with suppliers shifting capacity from client/mobile to data-center SSD (TrendForce, Jan 2026); 2010-era dominance of cards/USB + mobile is well-documented. HBM: first commercial use 2015 (AMD Fiji, graphics), data-center accelerators from Nvidia P100 (2016); by 2026 AI/ML training+inference ~55%+, HPC ~25%, graphics ~12%, emerging (autonomous/edge/networking) ~8% (PatSnap citing JEDEC); early years were graphics/HPC-weighted before AI dominance. HDD: nearline ≈54% of capacity shipped today, projected >90% by 2029 (Tom Coughlin); consumer/client HDD displaced by SSD (flash is "the default choice outside the data center"); 2010 was PC/consumer-dominated before the nearline/cloud shift. The caveats: (1) shares are by the noted unit (DRAM/NAND/HBM by bits, HDD by exabytes) — not revenue, and not unit count; (2) category boundaries blur (e.g. "server" vs "AI server," "graphics" GDDR vs HBM) and different sources slice differently; (3) everything past ~2025 is projection and widens with time; (4) HBM is a subset of DRAM shown separately because its end-use profile is distinct. Sources: TrendForce, PatSnap/JEDEC, Tom Coughlin/Forbes, industry market reports, company disclosures. Modeled shares — analysis, not investment advice.

06·D·WThe Memory Wall — Why Idle Accelerators Buy More Memory◷ measured anchors · roofline math shown
06·D·W

The Memory Wall — Idle Accelerators & Memory

FOR A CHILD

An AI chip is like a super-fast cook. It can chop a million carrots a second — but the carrots live in a pantry down the hallway, and it can only carry a few armfuls per trip. So the cook mostly just… stands there, waiting for carrots. "Memory" is the pantry and the hallway. Give the cook a bigger pantry right next to the stove and a wider hallway, and it almost never waits. That's the memory wall — and it's why everyone is racing to buy more memory.

FOR A TEENAGER

Here's the weird part: to write each single word, an AI has to re-read its entire brain — all of its billions of stored numbers — from memory chips sitting next to the processor. The math units are so fast they finish almost instantly, then wait for the next full re-read. So chatbot speed isn't set by how fast the chip does math; it's set by how fast memory can feed it (bandwidth) and how much fits next to the chip (capacity). The proof: Nvidia sold a chip with the exact same math speed as its old one but bigger, faster memory — and AI output nearly doubled. Rule of thumb from this section: generating one word costs as much chip-time as reading ~300.

FOR AN ENGINEER

Precisely: autoregressive decode at batch B has arithmetic intensity ≈ B FLOPs per weight-byte (weight reads amortize across the batch; per-request KV reads never do), against ridge points of ~281 (B200) to ~568 (Rubin) FLOPs/byte — so the tensor-core utilization ceiling ≈ min(1, B/ridge), B is capped by HBM capacity (weights + KV must be resident), and at long context the KV term dominates and pins utilization below 1% at any batch. Prefill is the compute-bound phase; decode is the memory-bound one; measured MFU on H100/Llama-3-70B stays under 20% below B≈100 for exactly the capacity reason. Everything below quantifies this — the roofline chart, the frontier and open-model scenarios, and the 16-prompt verdict table. (The analogies above compress; the numbers below don't.)

The dirty secret of AI inference: the world's most expensive chips spend most of their time waiting. During the decode phase — generating each new token — the GPU must re-read the entire model's weights from HBM for every single token. A 70B FP16 model means ~140 GB read per token. The H100 pairs ~1,979 FP8 TFLOPS of compute with 3.35 TB/s of bandwidth — a 591:1 compute-to-memory ratio — so at low batch sizes the tensor cores are mathematically capped at single-digit (even sub-1%) utilization while they wait on memory. More memory is the fix, through two distinct levers: more bandwidth steepens how fast utilization rises with batch size, and more capacity lifts the ceiling on batch size itself (every concurrent request needs its KV cache resident in HBM). That is the causal engine of this whole dashboard: idle compute is the most expensive thing in the data center, and memory is what un-idles it.

The causal chain, step by step.
1. At batch 1, each token's math touches every weight byte once — arithmetic intensity ≈ 1 FLOP/byte, vs the ~300 FLOPs/byte an H100 needs to keep its FP16 tensor cores busy → ~0.3% utilization ceiling. 2. Batching amortizes each weight-read across B concurrent requests — intensity scales ≈ linearly with B. 3. But every request must keep its KV cache in HBM (≈ 2 × layers × KV-heads × head-dim × context × bytes), so HBM capacity caps B. 4. Therefore: more HBM GB → bigger B → higher intensity → compute actually utilized; more HBM TB/s → more tokens/sec at any B. The proof: the H200 has identical TFLOPS to the H100 — only more, faster memory (141 GB @ 4.8 TB/s vs 80 GB @ 3.35) — and delivers up to 1.9× the inference throughput on Llama-70B (NVIDIA's own benchmark). Nvidia sold a memory-only upgrade, and the market paid a premium for it.
Tensor-core utilization ceiling vs batch size — and where memory capacity cuts it off
Roofline-derived utilization ceiling for 70B-class FP16 decode (util ≈ min(100%, bandwidth × B ÷ peak), log-log). Slope = bandwidth (faster HBM lifts the whole line). Vertical cut-off = capacity — the batch where weights + KV cache exhaust HBM (illustrative: FP8 weights ≈70 GB, KV ≈1.3 GB/request at 8K context). Roof = compute, rarely reached in decode. The ✕ marks the measured cross-check: a published H100 / Llama-3-70B profile found MFU stays below 20% for batches under ~100 — right where this model puts it.
Same 70B model, same-or-similar compute — memory did this (tokens/sec, single GPU)
Which prompts idle what — sixteen queries across the workload map, scored DERIVED · BATCH-1 · 70B-CLASS REF
Every request has two phases: prefill (ingesting the prompt — all input tokens in parallel, compute-bound) and decode (generating output — one token at a time, memory-bound). Which resource a prompt idles comes down to its input:output ratio — and the crossover is the ridge from the chart above: on this reference (70B FP16, H100, batch 1) one generated token costs as much accelerator-time as ~295 ingested tokens (281 on B200, 469 B300, 568 Rubin). So a request only tips compute-bound when it reads ~300× more than it writes. The table runs sixteen realistic prompts through that math — a span chosen to cover the big categories in public usage analyses: coding and agentic work (the largest API category), chat, writing, reasoning, RAG/support, summarization, translation, extraction, research, and vision. The honest takeaway: at interactive batch sizes almost everything is memory-bound — even a 60-token extraction from 3,000 tokens of contract — and the reasoning/agentic era (huge thinking outputs, long cached contexts) pushes the mix further toward memory. Only massive-ingest/tiny-output jobs tip the other way, which is exactly why serving stacks disaggregate prefill and decode onto separate hardware pools. Verdict tally: 13 memory-bound · 3 compute-bound.
Frontier reality check — Blackwell & Rubin running a Fable-5-class model MODELED SCENARIO
"Fable 5 Max" read here as the top-end deployment of Anthropic's Claude Fable 5 ("Max" is Anthropic's subscription tier, not a separate model). Anthropic discloses nothing about Fable 5's architecture — parameter count, MoE structure, KV design and serving precision are all non-public — so this is a schematic scenario: a frontier-class MoE (~2T total / ~120B active parameters, NVFP4 serving, KV ≈250 KB/token, NVL72-class pod), evaluated at two operating points — everyday chat (8K context) and the "Max" workload at Claude's public 200K-token context. Bars show decode-phase utilization ceilings vs each GPU's dense-FP4 peak; the idle share is hatched red. Two tells: the ceilings fall each generation (FP4 compute grows 5.6× from B200 to Rubin while bandwidth grows 2.75× — the wall gets higher even as tokens/sec improves), and at 200K context no batch size can fix it — batching amortizes weight reads, but every request's KV cache is unique, so long-context decode is irreducibly bandwidth-pinned below 1%. What extra HBM capacity buys there is concurrency (how many 200K requests fit at all), which is throughput and cost-per-token — not utilization.
Open-weights reality check — DeepSeek V4 Pro on the same silicon PUBLISHED ARCH · DERIVED MATH
The largest and highest-performing open-weight model as of mid-2026: DeepSeek V4 Pro — 1.6T total / 49B active MoE (61 layers, hidden 7168, 384+1 experts, 6 active/token), 1M-token context, MIT license, released Apr 24 2026. It is the largest open model by a wide margin and ranks #1 among open weights on agentic Elo (GDPval-AA 1554) with the highest open SWE-bench Verified (80.6%); on the raw Intelligence Index it scores 52, two points behind GLM-5.2 / Kimi K2.6 / MiMo-V2.5-Pro — the disclosure matters, and the honest read is "largest + top-tier," the closest single fit to the brief. Unlike the closed-model scenario above, this math is real: DeepSeek's own paper states V4 Pro needs just 27% of the FLOPs and 10% of the KV cache of V3.2 at 1M context → KV ≈ 7 KB/token (≈2% of GQA-class caches), and the FP4/FP8 checkpoint fits on an 8×B200 node. Two tells, opposite directions: at chat scale, engineering around the wall works — V4's KV is so small that on the B200 the ceiling actually reaches the compute roof (a first in this section, though it takes ~281 concurrent requests, which now genuinely fit); at its headline 1M context, even 98% KV compression leaves 98.6–99.3% idle — capacity stops binding (thousands of requests fit) and pure bandwidth-per-request becomes the whole wall.
Which prompts idle what — 75 queries on DeepSeek V4 Pro OPEN MoE · DERIVED
The same lens as the sixteen-prompt table earlier, now on the open model characterized just above — DeepSeek V4 Pro (1.6T total, 49B active per token, 7 KB/token KV), served batch-1 on an 8×B200 node in FP8. Two things change versus the dense-70B reference. (1) MoE decode is lighter: each token reads only the ~49 GB of active experts, not a full dense model — so absolute latency drops, but the compute:bandwidth balance is a hardware property, so the crossover is still ~281 input tokens per generated token (vs ~295 on the H100 table). (2) Compressed KV barely bites: at 7 KB/token, even a 250K-token context adds only ~3–4% to decode — so unlike a GQA model, long context does not push the verdict harder; the active-weight read dominates. The result mirrors the dense case: 56 of 75 are memory-bound. Only massive-ingest / tiny-output jobs — classification, moderation, single-field extraction, reranking, needle-in-haystack — tip to compute(prefill)-bound, and the crossover shows up as borderline cases near the ~281:1 line (invoice extraction 321:1, ticket-routing 300:1).

DeepSeek-V4 prompt table — derived math. Reference: DeepSeek V4 Pro (49B active of 1.6T MoE, 7 KB/token KV — see the open-weights block above) served batch-1 on an 8×B200 node, FP8 (aggregate ~36 PF dense FP8, 64 TB/s HBM — the node that holds the ~1 TB checkpoint). Per-token: prefill ≈ 2×49e9 ÷ 36 PF ≈ 0.0027 ms/input-token (compute-limited); decode ≈ (49 GB active-expert read + 7 KB×context KV) ÷ 64 TB/s ≈ 0.77 ms/output-token, rising only to ~0.79 ms even at 250K context — the compressed-KV signature. Crossover ≈ 281 input:output, a hardware compute:bandwidth property (hence close to the H100 table’s 295), not a model property. MoE caveat, sharper than dense: batching amortizes the active-weight read, but different tokens route to different experts, so a large batch touches most of the 384 experts — amortization is weaker than for a dense model and real decode stays memory-bound to higher batch sizes. Verdicts are batch-1 per-request; prefill idealized at 100% MFU; token counts illustrative. Not investment advice.

So what share of all token usage is memory- vs compute-limited? ESTIMATE · TWO LENSES · REFRESHED JUL 2026
The tables above score individual prompts; this is the aggregate across all real inference. The honest answer depends entirely on how you weight it — a prefill (compute) token and a decode (memory) token differ ~300× in cost, so counting tokens and counting GPU-time give near-opposite answers. Both are shown, with the governing rule underneath. Jul 2026 refresh — the question got genuinely harder: agentic serving has pushed raw input:output ratios through the crossover (agents now routinely send 50K–500K input tokens against a few hundred out — 100:1 to >1,000:1, i.e. nominally compute-bound), while prefix caching deletes most of that prefill work and reasoning tokens pile on decode. The two forces are drawn out below, because they very nearly cancel.
By GPU-time / cost / energy — the metric that buys memory
~70–90% memory-bound
Decode dominates wall-clock time unless a request reads >~280–300× more than it writes. Almost none do. This is the economically meaningful split — it's what the HBM bill actually pays for.
By raw token count — the less-meaningful metric
compute-leaning
Production is input-heavy (RAG, agents, long context) and 2026 agentic traffic made it more so — but each prefill token is worth ~1/300th of a decode token, and a cached one costs no prefill at all. The majority by count remains a sliver of the cost.
Force 1 — pushes toward COMPUTE (prefill)
Agentic context explosion. A ReAct agent making 10 tool calls can consume ~800K input tokens to emit ~500 output — a ~1,600:1 ratio, far above the ~281–295:1 crossover, and re-billed every round. Gartner (Mar 2026) puts agentic tasks at 5–30× the tokens of a chatbot turn. On raw ratios alone, much of 2026's fastest-growing workload screens prefill-bound.
Force 2 — pushes back toward MEMORY (decode)
Caching deletes the prefill; reasoning adds decode. A cached prefix token is never prefilled again — hit rates run 84–92% in tuned agent stacks, cutting effective prefill ~5–10× and dragging that 1,600:1 back under the line. Meanwhile reasoning models emit 20–30× more output tokens (a 200-token answer can hide 10–30K thinking tokens) — all of it decode. Net: the memory-bound share holds.
Share of GPU-time (illustrative central estimate)memory-bound decode vs compute-bound prefill
~80% MEMORY (decode)
~20% compute
Share of tokens processed (illustrative · nudged prefill-ward for 2026 agentic mixes)counted before cache hits remove prefill work
~75% prefill (compute)
~25% decode (memory)
Bars are illustrative central points within wide ranges, not precise measurements — the aggregate input:output mix is not publicly reported, and 2026's cache-hit rates vary hugely by stack (1.7%→92.2% on the same workload after tuning). What's anchored is the mechanism, the crossover, and the cited workload/cache figures; the weighting is estimated. The GPU-time split is the one that has held up — it is the raw token mix that moved this year.

How the estimate is built, and its limits. The mechanism [well-established]: prefill (ingesting the prompt) is compute-bound; decode (generating tokens) is memory-bandwidth-bound — now "the default playbook across nearly every major LLM serving" stack, with disaggregation shipping at Perplexity, Meta, Mistral, and via NVIDIA Dynamo/vLLM/SGLang/TensorRT-LLM. Measured, prefill runs GPUs at 90-95% utilization while decode craters to 20-40%, and prefill saturates tensor cores at 30-45% MFU while decode sits mostly idle waiting on memory bandwidth (Towards Data Science / InfoQ Sept 2025; buildmvpfast; HuggingFace). The governing rule [derived, shown above]: a request's decode-time exceeds its prefill-time whenever its input:output ratio is below the crossover — ~295:1 on H100, ~281:1 on B200 — so unless average prompts read ~300× more than they write, GPU-time is decode-dominated, i.e. memory-bound. Why time ≠ tokens: output tokens cost far more to produce — providers price them at 3:1 to 5:1 over input tokens (e.g. Claude Sonnet $3 input / $15 output per million, a 5:1 ratio), reflecting the sequential per-token generation cost. So even where input tokens dominate by count, output tokens dominate by time and money. Workload weighting [cited anchors, still an estimate]: real mixes span a wide range — a typical support-ticket workload is ~3,150 input / 400 output tokens (~8:1, input-heavy), while high-throughput models like DeepSeek V3.2 show an output-to-input ratio near 1.6× and Llama 4 Maverick around 3× (output-heavy); programming rose from 11% to over 50% of all LLM token usage on OpenRouter by late 2025, and agentic coding workflows average 1–3.5M tokens per task — long, cached, input-heavy contexts by count but with substantial generation. Crucially, reasoning/thinking tokens are billed and processed as output (decode) — reasoning models use 10–20× more tokens — pushing the time-share further toward memory. The batch caveat [bounds the estimate]: at high concurrency decode GEMMs can turn compute-bound — past roughly batch 32 the decode GEMMs become compute-bound while attention stays bandwidth-bound — but the attention/KV reads remain memory-bound, and at long context KV dominates and keeps decode memory-bound at any batch (per the frontier and DeepSeek blocks above). This is the main reason the time-share isn't ~100% memory: heavy-batch, short-context decode and bulk-ingest prefill claim the compute-bound remainder. Jul 2026 refresh [cited]: hardware-level measurements have sharpened — prefill drives GPUs to 70–100% of peak power while decode draws only 20–40%, and tensor-core utilization runs ~92% in prefill versus ~28% in decode milliseconds later (Spheron 2026 guide; Towards Data Science). Disaggregation is now a costed engineering default: an H100-prefill + H200-decode split costs ~45% more per hour but delivers ~75% more throughput. The agentic counter-pressure [cited, and the main revision]: 2026 agents routinely send 50,000–500,000 input tokens per request against only a few hundred output, and a ReAct agent running 10 tool calls can consume ~800K input tokens while emitting ~500 output — ratios of 100:1 to ~1,600:1, above the ~281–295:1 crossover, which on raw ratios alone would flip those requests to prefill-bound. Gartner (Mar 2026) puts agentic tasks at 5–30× the tokens of a standard chatbot turn. Why it doesn't flip the answer [cited]: (1) prefix caching removes the prefill entirely for repeated context — tuned stacks reach 84–92% cache-hit rates (one workload went 1.7%→92.2%; ProjectDiscovery 7%→84%), cutting effective uncached prefill roughly 5–10× and pulling those ratios back under the line; the market prices this directly, with cached input billed at ~10% of normal (Anthropic), ~25% (Google), ~50% off (OpenAI). (2) reasoning tokens are decodereasoning models emit 20–30× more tokens per task, and a 200-token visible answer can hide 10,000–30,000 thinking tokens, every one of them memory-bound generation. (3) KV pressure is itself now a reported constraint on agentic economics (WEKA, 2026). The two forces very nearly cancel; the token-count bar above is nudged prefill-ward to reflect the agentic shift, while the GPU-time split is left unchanged. Bottom line: by the metric that matters for this dashboard — GPU-time, and therefore memory purchased — inference is majority memory-bound, on the order of ~70–90%; by raw token count it leans compute (prefill) because production is input-heavy, but those tokens are individually ~300× cheaper. The mechanism and crossover are anchored; the exact percentages are an estimate — the aggregate input:output distribution across all providers is not publicly disclosed. Sources: Towards Data Science, buildmvpfast, Medium (Chen; Patel), HuggingFace, arXiv 2512.22066, Silicon Data, iternal.ai, onprem.ai, Grizzly Peak, Mobisoft; 2026 refresh: Spheron (prefill/decode disaggregation & context-engineering guides), yage.ai (KV cache-hit economics), WEKA (KV eviction & agentic economics), vLLM/Mooncake, Gartner (Mar 2026), Digital Applied, newline; plus the roofline anchors cited above. Estimate, not a measurement — not investment advice.

How much memory bandwidth per megawatt maximizes tokens/sec? ESTIMATE · MODELED
The capacity question (GB/MW) turned out to saturate — a rack already holds enough KV to reach the roofline, so past the ridge more GB buys context, not throughput. The question that does not saturate is bandwidth. Decode is memory-bandwidth-bound: every generated token requires streaming the model’s weights (and its KV) out of HBM, and the arithmetic to process them finishes faster than the read. So in the regime where inference actually lives, tokens/sec per megawatt tracks bandwidth per megawatt — nearly one-for-one. Here is what a MW streams today, and what the HBM roadmap adds.
HBM bandwidth per MW — today (GB200 / HBM3e)
~4.4 PB/s
≈ 546 B200s × 8 TB/s per facility-MW. This — not FLOPS, not GB — is the decode speed limit.
Decode tokens/sec per MW — at the roofline
~25M tok/s
Compressed-KV MoE, weight-streaming bound. Scales ~linearly with bandwidth, so every figure here moves with the column below.
The scaling law [measured]
TPS ∝ BW
H100→H200 raised bandwidth 1.43× (3.35→4.8 TB/s) and delivered ~1.4× tokens/sec on the same model. Bandwidth converts almost directly to throughput.
Additional bandwidth — HBM4 / Rubin
~7.5 PB/s
~1.7× GB200 per MW → ~42M tok/s/MW. Per chip HBM4 is 2.75× (8→22 TB/s), but ~2× the watts mutes it to ~1.7× per MW.
Why per-MW lags per-chip: each generation raises bandwidth and power. Rubin’s 22 TB/s is 2.75× a B200’s 8 — but at ~1.8–2.3 kW/GPU vs 1 kW, a MW holds fewer of them, so the per-MW bandwidth (and TPS) gain lands at ~1.7×. GB300 makes the point in reverse: same HBM3e at 8 TB/s, so despite more capacity and FP4 compute, its bandwidth-per-MW barely moves. The tokens/sec jump comes from the HBM generation, not the GPU badge.
Bandwidth per megawatt across HBM generations
Decode throughput per MW rises with aggregate HBM bandwidth per MW. The relative column indexes decode TPS/MW to GB200 = 1.0 — valid because, in the memory-bound regime, tokens/sec tracks bandwidth.
AcceleratorHBMBW / GPU~BW / MWrel. decode TPS/MW
H100HBM33.35 TB/s~2.4 PB/s0.55×
H200HBM3e4.8 TB/s~3.4 PB/s0.8×
B200 · GB200HBM3e8 TB/s~4.4 PB/s1.0× (ref)
B300 · GB300HBM3e8 TB/s~3.9 PB/s~0.9×
Rubin · VR200HBM422 TB/s~7.5 PB/s~1.7×
Rubin UltraHBM4e(2027)higher>2×
Absolute anchor: GB200’s ~4.4 PB/s → ~25M tok/s/MW for a compressed-KV MoE at the roofline; the relative column scales that. Real systems reach a fraction, capped by achievable batch and latency — but the ceiling moves only with bandwidth.

Why bandwidth is the tokens/sec lever, and how these are built. The mechanism [established]: most production inference is memory-bandwidth-bound — generating each token requires reading the model’s weight parameters out of HBM, and the compute to process them takes less time than the read, so adding FLOPS without proportional bandwidth yields diminishing returns. In that regime decode tokens/sec is set by how fast weights (and KV) stream from memory, so it scales ~linearly with bandwidth: the H100→H200 step lifted bandwidth 1.43× (3.35→4.8 TB/s) and produced almost exactly 1.4× tokens/sec on the same model. Per-GPU bandwidth by generation [anchored]: H100 is 3.35 TB/s on HBM3, H200 4.8 TB/s and B200 8 TB/s on HBM3e, and Rubin (R100) reaches up to 22 TB/s on HBM4; B300/GB300 holds at 8 TB/s HBM3e (288 GB, 1,400 W) — capacity and FP4 compute rise, per-chip bandwidth does not. A Rubin NVL72 rack aggregates 1.6 PB/s across 20.7 TB of HBM4, and NVIDIA states the Rubin GPU nearly triples memory bandwidth compared to Blackwell. The HBM generation is the engine [anchored]: HBM4 delivers ~2 TB/s per stack versus HBM3e’s ~1.2, via a doubled 2,048-bit interface, yet bandwidth still grows only ~1.6× every two years while FLOPS scale ~3× — widening the memory wall; HBM4’s per-stack uplift is roughly +75–120%. Rubin Ultra (H2 2027) moves to HBM4e with 1 TB per GPU in 600 kW NVL576 racks. Per-MW conversion [modeled]: bandwidth/MW = (GPUs/MW) × (BW/GPU), with GPUs/MW from all-in power at a liquid-cooled PUE ~1.15 (~714 H100/H200, ~546 B200, ~490 GB300, ~340 Rubin per MW). The per-MW gain trails the per-chip gain because power climbs each generation — Rubin draws ~1,800–2,300 W per GPU versus Blackwell’s 1,000 — so HBM4’s 2.75× per-chip bandwidth lands at ~1.7× per MW. Why it’s the thesis: a megawatt’s token output is gated by HBM bandwidth, and the only way to raise it is the next HBM generation — supplied by exactly three firms, SK Hynix (~62%), Micron (~21%) and Samsung (~17%), with capacity sold out. That is the memory supercycle in one line: tokens/sec per watt is an HBM-bandwidth story (see §05·C·G). All MODELED: per-GPU bandwidths and the H200 scaling point are anchored to published specs; per-MW figures fold in power/PUE assumptions and a representative weight-streaming model, none of which is a single number. Estimate, not a measurement — not investment advice.

Provenance & the honesty line. Measured / cited anchors: decode is memory-bandwidth-bound, prefill is compute-bound (arXiv 2601.11822 and standard serving literature); a 70B FP16 model reads ~140 GB from memory per generated token, and on an H100's 3.35 TB/s that sets a ~42 ms/token floor ≈ ~24 tok/s regardless of FLOPS; H200 ≈ 34, B200 ≈ 57 on the same model (GMI Cloud, Spheron — B200 figures trace to GTC-2024 disclosures/estimates); the H100 pairs 1,979 FP8 TFLOPS with 3.35 TB/s — a 591:1 compute-to-memory ratio (Introl); measured MFU stays below ~20% at batch sizes under ~100 on H100 / Llama-3-70B, explicitly because VRAM cannot hold the KV cache needed for larger batches (arXiv 2405.01814 — the single best empirical statement of this section's thesis); one profiled inference kernel ran at 23% compute and 47% memory-bandwidth utilization (arXiv 2504.06319); H100s have been measured using only ~32% of peak memory bandwidth in distributed decode, with HBM accesses alone 30–50% of energy (arXiv 2602.18568); even large-batch serving stays DRAM-bound with significant compute idle (arXiv 2503.08311). The H200 proof point [NVIDIA product brief via GMI]: H100 and H200 share identical FP16 (989) and FP8 (1,979) TFLOPS; the H200's up-to-1.9× Llama-2-70B inference speedup comes entirely from memory (141 GB HBM3E @ 4.8 TB/s vs 80 GB HBM3 @ 3.35). KV-cache formula per request: 2 × layers × KV-heads × head-dim × seq-len × bytes (GMI). Derived [formula shown, labeled]: the chart's utilization curves are roofline ceilings — util ≈ min(100%, BW×B/peak) with FP16-dense peaks (H100/H200 989 TFLOPS, B200 ~2,250 est.) — real kernels achieve less (see the measured 23%/47% figures), so the idle problem is worse than the curves, not better. Illustrative [assumptions stated]: the capacity cut-offs assume a 70B-class model, FP8 weights ≈70 GB, FP8 KV ≈160 KB/token, 8K average context → ≈1.3 GB per request; real deployments shard across GPUs, page KV (vLLM-style), quantize, and disaggregate prefill/decode, so exact caps vary widely — the mechanism (capacity caps concurrency, concurrency drives utilization) is the anchored claim, not the specific integers. Prompt-verdict table [derived — reference math shown]: per-token costs on the section's reference config (70B FP16, H100, batch 1): prefill ≈ 2×70B FLOPs ÷ 989 TFLOPS ≈ 0.14 ms per input token (compute-limited); decode ≈ 140 GB ÷ 3.35 TB/s ≈ 41.8 ms per output token (bandwidth-limited) — the agent row adds the 200K-context KV read (~64 GB at GQA-class 320 KB/token) → ~61 ms/token. The crossover ratio (input tokens per output token at which prefill time equals decode time) is decode÷prefill ≈ 295 — algebraically identical to the ridge batch, since both reduce to peak ÷ (bandwidth × FLOPs-per-byte). Caveats: verdicts are per-request at batch 1 — production batching divides the weight-read share of decode by ~B until context/KV dominates, pulling crossovers down; prefill is idealized at 100% MFU (real prefill runs ~40–70%, which lengthens prefill's time share but doesn't change which resource limits each phase); token counts are illustrative (~4 chars/token; thinking budgets vary widely); prompt caching converts recomputed prefill into stored KV — trading compute time for HBM capacity occupancy, the section's thesis in miniature. Coverage note for the sixteen-row table: the shapes span the categories that dominate public usage analyses — coding/agentic work (the largest API category), chat, writing/rewriting, reasoning, RAG & support, summarization, translation, extraction/classification, research, and vision — with illustrative token counts, not measured shares (platform mixes vary; one image counts ≈ 1.5K prefill tokens). Methodology consistency: rows whose decode runs at long context (the 10-K, agent turn, agentic-coding loop, deep research, needle-in-haystack) include the per-token KV read in the decode cost; the offline-classification row is the one deliberate exception to batch-1 — scored at batch 256 to demonstrate the flip that production batching produces. Frontier scenario [modeled — every workload number is an assumption]: Anthropic publishes no architecture details for Claude Fable 5 (the model that drafted this dashboard does not know its own parameter count — it isn't disclosed even to it), so the scenario uses round frontier-class figures in line with widely-reported analyst estimates of GPT-4-class systems (~1.8T-parameter MoE): ~2T total / ~120B active, NVFP4 weights (~1 TB, sharded ~14 GB/GPU across an NVL72-class pod), KV ≈250 KB/token with GQA/MLA-style compression; Claude's 200K context window is the one public spec used. Utilization here = decode-phase ceiling vs dense-FP4 peak: at short context, batching amortizes weight reads and the ceiling ≈ min(compute ridge, KV asymptote); at 200K the KV read per token (~50 GB/request) dominates and the ceiling collapses to ~0.3–0.6% regardless of batch — added capacity then buys concurrency (how many 200K requests fit), not utilization. MoE routing makes real ceilings lower still (large batches touch most experts, weakening amortization), prefill mixing raises blended MFU, and production kernels run below all of these ceilings (see the measured anchors above). Frontier GPU specs [published / reported]: B200 192 GB HBM3E · 8 TB/s · 9 PF dense FP4 (shipping); B300 288 GB (12-Hi HBM3E) · same 8 TB/s · 15 PF dense FP4 — its advantage is "purely about fitting larger models or longer contexts… rather than serving each token faster" (Spheron; SemiAnalysis; Tom's Hardware; IntuitionLabs; shipping since H2'25); Vera Rubin 288 GB HBM4 · 50 PF dense FP4 · bandwidth reported at 22 TB/s in the latest GTC-2026 coverage after evolving 13 → 20.5 → 22 TB/s across reporting as NVIDIA pushed HBM4 pin speeds (tech-insider, SemiAnalysis, IntuitionLabs) — sampling Q4 2026, volume Q1 2027, so all Rubin figures are preliminary. Corroborating context from the same reporting: memory is set to consume ~30% of hyperscaler data-center spend this year (~4× 2023) and NVIDIA system memory costs are up ~485% (Tom's Hardware). Open-weights block [published architecture + derived math]: DeepSeek V4 Pro specs are from its Hugging Face model card / paper abstract (Apr 24 2026, MIT): 1.6T total / 49B active, 61 layers, hidden 7168, 384 routed + 1 shared experts (6 active/token), native 1M context; the abstract states the hybrid CSA+HCA attention needs 27% of single-token inference FLOPs and 10% of the KV cache vs V3.2 at 1M context. KV/token is derived from that ratio: V3.2's MLA cache ≈ 576 dims × 61 layers × 2 B ≈ 70 KB/token → V4 Pro ≈ 7 KB/token; cross-check: reporting puts it at ~2% of an 8-head GQA BF16 cache, and that GQA reference works out to ~250 KB/token — the same figure assumed in the closed-model scenario above, an independent validation of that assumption. Weights ≈1 TB mixed FP4/FP8, anchored to Lambda's note that the checkpoint fits on an 8×B200 node (≈14 GB/GPU sharded across an NVL72 pod). Directional caveat: the 1M-context bars model the per-token KV read as the full compressed cache — CSA/HCA's sparse reads touch less, so V4's true long-context ceilings sit above the bars shown (the conservative direction); capacity math uses full storage, which is correct regardless. "Largest and highest-performing" disclosure: GLM-5.2 (744B) and Kimi K2.6 / MiMo-V2.5-Pro (1T) edge V4 Pro on the Intelligence Index (≈54 vs 52), but V4 Pro is the largest open model, #1 on open agentic Elo, and highest open SWE-bench — the closest single fit; swapping the block to any of those is a config change. The meta-point the two blocks make together: the best open architecture is explicitly engineered around the memory wall (MoE sparsity + compressed/sparse attention) — demand-side adaptation to memory scarcity — and cheap 1M-token context is exactly what unlocks more long-context usage, so per-request memory efficiency feeds total memory demand rather than reducing it (sources: DeepSeek V4 model card, Latent Space/Lambda, Artificial Analysis, morphllm, aimadetools, codersera). On-thesis: this is the machine that converts token demand (§06·D) into memory demand (§05·B, §06·D·E) — and why TrendForce frames the supercycle as a "memory wall" problem: as AI shifts to inference, compute demand per token falls while memory demand rises, with inference set to be the primary AI-server driver by 2029. Sources: arXiv 2405.01814, 2504.06319, 2602.18568, 2503.08311, 2601.11822; GMI Cloud; Spheron; Introl; NVIDIA H100/H200 datasheets; TrendForce "Memory Wall" insight (Jan 2026). Not investment advice.

06·D·DMemory Demand Drivers — Every Trend Pushing Capacity & Bandwidth◷ synthesis
06·D·D

Memory Demand Drivers

One map of every trend compounding memory demand — C = capacity (GB), B = bandwidth (TB/s). The point of seeing them together: these multiply, not add. Longer contexts × more reasoning tokens × more agents × more users × bigger models — each is a separate factor on the same HBM bill, which is why token volume can grow 24× while memory demand grows faster.

TrendC / BScale of effectTracked in
Context windows — KV cache grows linearly with tokens in contextC + B128K ≈ 19 GB KV per sequence (GQA) vs ~0.9 GB (MLA); decode re-reads it every token§06·D·W
Reasoning / thinking tokens — answers now include long hidden tracesB10–20× more output tokens per query; all decode = all bandwidth§06·D·W agg.
Agentic loops — full context re-sent every tool step; agent teams run parallel instancesC + B0.4–2M tokens per task; per-dev use +18.6× in 9 months§08·E
Model scale / MoE totals — all experts must be resident even if few activateC1.6T-param checkpoint ≈ 1 TB in FP8 — a full NVL72 rack to hold one model§06·D·W open block
Batch / concurrency — saturating compute needs KV for ~281 sequences at onceCroofline batch × context × KV/token; 5.4 TB at 128K standard-KV§06·D·W per-MW
Multimodality — images ~1.5K tokens each; video/audio are token streamsC + Bvision rows land memory-bound in both prompt tables§06·D·W tables
Prompt caching — the fix that eats capacity: cached KV kept resident/tiered to DRAMC90% input discount only exists because the KV is stored, not recomputed§06·D·W · §01
Token volume growth itself — more users, more queries, Jevons dynamicsC + BGoldman: 24× enterprise tokens by 2030 (120 quadrillion/mo)§08·E · §08·D
FLOPS outrunning bandwidth — compute grows ~3×/2yr, HBM BW ~1.6×/2yrBthe wall widens each generation → bandwidth carries a structural premium§06·D·W BW/MW · §05·C·G
On-device AI floors — edge inference raises minimum DRAM per phone/PC/carCiPhone base 8→12 GB; Tesla 300 GB+/vehicle trajectory§08·C · §07·B·T
Multi-tenancy extras — LoRA adapters, draft models for speculative decoding, embeddings/RAG storesCeach resident alongside weights; RAG shifts GB into server DRAM too§01 · §06·D·E

Label: SYNTHESIS. This section introduces no new figures — every number is sourced and cited in the section its row links to; this is the one-glance union. The compounding claim is the analytical content: demand drivers are multiplicative (context × reasoning × agents × users × model size), while supply-side relief is per-factor (MLA compresses KV ~20× but touches only one factor; HBM4 lifts bandwidth ~1.7× per MW but not capacity per user). The bear case is the same table read backwards: efficiency gains (compression, distillation, caching, smaller routed experts) attack individual rows — kill-switch #5 (§02·D) watches whether efficiency ever outruns compounding. It has not yet. Not investment advice. See also: §06·D·D is the qualitative synthesis; §05·F puts numbers on the silicon those drivers consume.

06·D·RAI Model Usage — OpenRouter Rankings◉ live extract · Jul 17 2026
06·D·R

AI Model Rankings — OpenRouter

Where the tokens in §06·D actually get spent. OpenRouter routes a large slice of the independent LLM-inference market, so its weekly usage is one of the cleanest public reads on which models are consuming compute — and therefore memory. The numbers are now fully cited — extracted live from openrouter.ai (Jul 17 2026): weekly throughput ran 3.41T (Jul 28 2025) → 62.8T (Jul 13 2026), an 18.4× ramp in a year, with the steep leg after January. Every week carries the site's own per-model split (top-9 + Others). Every token here is a KV-cache read against HBM (§06·D·W), so this curve is the demand side of the supercycle, measured in the wild.

Top Models — weekly usage across OpenRouter
tokens / week · T = trillion
Now fully cited, every week. All 52 weeks were extracted from openrouter.ai itself (Jul 17 2026) by programmatically hovering the site's own chart and transcribing its tooltips — so every bar's colour split is the site's real per-model data (each week's top-9 + Others, exactly as OpenRouter reports it), and hovering any bar reproduces the site's own tooltip verbatim, including the Total. Colours are stable per model across weeks. The final bar (Jul 20) is the in-progress week — the site's pace estimate is ≈58T. Segment sums match the printed totals within the site's own display rounding (<2%).
LLM Leaderboard — most popular models

Live extract · Jul 17 2026 · sort: Trending · All models — the Trending sort surfaces fast-growers (note Hy3, also #2 by weekly volume), not the volume leaders in the chart above

Provenance & method. Source: OpenRouter LLM Rankings, openrouter.ai/rankings?view=trending, extracted live in-browser on Jul 17 2026. The page is client-rendered with no public per-model-per-week endpoint, so the series was captured at the render layer: the site's own chart tooltips were triggered programmatically for each of the 52 weekly bars and transcribed verbatim — date, top-9 models with values, Others, and Total. [cited]: all 52 weeks (Jul 28 2025 → Jul 20 2026), 520 model-week values plus totals; segment sums reconcile to printed totals within display rounding (<2%); the Jul 13 2026 total (62.8T) matches the site's rendered chart peak. Window notes: Jul 20 2026 is the in-progress week (27.6T so far; the site's own pace estimate ≈58T). Model sets change per week (each week shows its own top-9); a handful of names repeat within a week where the site lists paid and free variants under one display name — kept as displayed. Superseded: this section's earlier illustrative bands and image-read totals (2.5T→62T) are replaced by this extract (3.41T→62.8T actual); a stale 11-column hydration dataset found in the page (year-ago leaders + Others, totals ~19T) did not match the rendered chart and was discarded in favour of the tooltip transcription, which does. The leaderboard is the live Jul 17 Trending list. Not investment advice.

06·D·SPaid AI Subscriptions — The Monetization Curve◷ 2 firms cited · 3 modeled · proj to company targets
06·D·S

Paid AI Subscriptions

How many people and companies actually pay for AI — the monetization mirror of the token-usage and backlog sections. This stacks the paid base (consumer chatbot subscriptions and enterprise seats) across the five biggest platforms, quarter by quarter for the last two years, then projects to 2030. Two segments are anchored to hard disclosures: ChatGPT (OpenAI reported ~15.5M paid subs end-2024, ~35M mid-2025, ~47M end-2025, ~50M by Feb 2026) and Microsoft 365 Copilot (15M paid seats reported Jan 2026, 20M reported Apr 2026). The other three — Gemini, Claude, and the long tail (Perplexity, Grok, …) — do not break out paid-subscriber counts, so those bands are modeled from revenue and bundling disclosures and labeled as such. The projection leans on the one company that published a long-range target: OpenAI guides to ~122M consumer subscribers by end-2026 and ~220M by 2030; the rest grow at roughly the ~37%/yr rate the market-research houses use for generative AI. By mid-2026 the cited-plus-modeled paid base is ~125M; the illustrative 2030 total is ~600M.

Paid AI subscriptions & seats by quarter, then projected to 2030
Stacked bars = total paid AI subscriptions + enterprise seats, millions. Solid bars are the last eight quarters (the actual window); hatched bars are year-end projections. White markers sit inside the ChatGPT and Copilot bands on quarters backed by a hard company disclosure (or, for 2026 & 2030, OpenAI's own published target). The Gemini / Claude / Other bands carry no ● — they are modeled, since those firms don't report paid-sub counts. Every value is transcribed in the table below.
ChatGPT (OpenAI)Microsoft CopilotGoogle GeminiClaude (Anthropic)Other (Perplexity, Grok, …)● disclosed / company targethatched = projection
The cited anchors — the hard numbers under the curve
ChatGPT paid subs · end-2024
15.5M
nearly tripled in 2024 (The Information)
ChatGPT paid subs · mid-2025
~35M
Plus + Pro (The Information)
ChatGPT paid subs · Feb 2026
~50M
all tiers, +9M business users flagged
Copilot paid seats · Jan 2026
15M
Q2 FY26 call, +160% YoY (Microsoft)
Copilot paid seats · Apr 2026
20M
Q3 FY26 call, +5M in one quarter
OpenAI's own target
122M → 220M
consumer subs, end-2026 → 2030 (The Information)
The full series — transcribed table (millions)
How to read the honesty here
Cited (● anchors): the ChatGPT and Copilot bands are pinned to hard company disclosures on the quarters marked ●; the values between those quarters are straight interpolation. Modeled: Gemini, Claude and the long tail are not disclosed as subscriber counts, so they are derived — Gemini from Google One / Gemini subscription revenue (~$1.2B in 2025) and Workspace bundling; Claude from Anthropic's revenue ramp ($3B mid-2025 → ~$14B annualized early 2026, of which consumer + Claude Code subs are a slice); Other from Perplexity's ~$100M+ ARR plus Grok/other tails. Treat those three bands as order-of-magnitude, not precise. Projected (hatched): OpenAI's bars sit on the company's own published targets (122M end-2026, 220M by 2030); every other segment is grown at ~35–40%/yr, the generative-AI market CAGR the research houses cluster around — an illustrative scenario, not a forecast. Two conflicts, flagged not merged: (1) OpenAI's "50M" is variously reported as all-tiers total or as consumer-only-plus-9M-business — this board uses ~50M as the consumer-tier anchor and notes the business seats separately rather than silently double-counting. (2) "Seats" (Copilot, enterprise) and "subscriptions" (consumer) are different units bundled into one paid base here — deliberately, to show total monetized demand, but the table keeps them separable.

Provenance & the honesty line. ChatGPT / OpenAI [cited — reported to shareholders / announcements, via The Information & Reuters]: 5.8M paid subs (end-2023) → 15.5M (end-2024, "nearly tripled") → 20M (early Apr 2025) → ~35M (Jul 2025, Plus+Pro) → ~47M (end-2025) → ~50M across all tiers + ~9M paying business users (Feb 2026); company target ~122M consumer subs by end-2026 (driven by a new ~$8/mo tier) and ~220M by 2030 (≈8.5% of a projected 2.6B weekly users). Microsoft 365 Copilot [cited — FY26 earnings calls]: 15M paid seats (reported Jan 28 2026, Q2 FY26, +160% YoY) → 20M paid seats (reported Apr 29 2026, Q3 FY26, +5M QoQ); ~5.8M implied a year earlier from the 160% YoY; still only ~3.3% of Microsoft's 450M+ commercial seats. Google Gemini [modeled/derived]: Google doesn't break out Gemini subscribers; derived from Google One (150M+ paid subs) with AI Premium/Gemini bundled, ~$1.2B of 2025 Gemini subscription revenue, Workspace Gemini add-ons, and Alphabet's 350M total paid subscriptions (Q1 2026). Claude [modeled/derived]: Anthropic discloses revenue, not sub counts; derived from the $3B→~$14B annualized revenue ramp (mid-2025→early-2026) and Claude Code (~$2.5B annualized), with consumer chat well below ChatGPT. Other [modeled]: Perplexity (~$100M+ subscription ARR, 2025), xAI Grok / SuperGrok (largely bundled with X), and smaller assistants. Projections [illustrative]: OpenAI bars = company targets; all other segments compounded at ~35–40%/yr, consistent with generative-AI market forecasts of a ~37% CAGR to 2030 (Grand View, MarketsandMarkets) and Counterpoint's ~$700B consumer-GenAI-spend-by-2030 view. Honesty caveats: (1) only two of five segments are cited; three are modeled and drawn without ● markers. (2) consumer subscriptions and enterprise seats are summed into one "paid base" — different units, disclosed. (3) the ~$225B/$700B consumer-spend figures are spending, not subscriber counts, and are used only to set the projection growth rate, not the levels. (4) interpolation between disclosed quarters is linear. Cross-refs: §06·D (token demand), §06·D·R (OpenRouter usage), §08·D (app-layer revenue), §06·F·B (RPO backlogs). Sources: The Information, Reuters, Seeking Alpha, Microsoft FY26 IR, Alphabet IR, Counterpoint, Grand View Research, MarketsandMarkets, Business of Apps. Not investment advice.

06·D·MAI Model Size — Parameters Over Time, and the Disclosure Blackout◷ open = disclosed · closed = estimated
06·D·M

AI Model Size Over Time

First — why "how big is the model?" is really a memory question

Every parameter is a number the accelerator has to store and read back for each token it generates. So the first-order memory a model needs is simply how many parameters it has × how many bytes each one takes — and that memory has to sit in the fast memory bolted to the chip, HBM, the exact supply this whole dashboard tracks. Bigger model → more bytes → more HBM. That is the reason model size belongs here at all.

weight memory  =  parameters  ×  bytes per parameter

bytes / parameter by precision:  FP16 / BF16 = 2  ·  FP8 = 1  ·  INT4 = 0.5  (lower precision = smaller, up to a quality limit)

70B dense · FP8
~70 GB
≈ one H100 (80 GB HBM)
405B · Llama 3.1 · FP8
~405 GB
≈ 6× H100 · fits one 8-GPU node
671B · DeepSeek-V3 · FP8
~671 GB
≈ 9× H100 · 4× B200 (192 GB)
2.8T · Kimi K3 · FP8
~2.8 TB
≈ 35× H100 · ~20% of an NVL72 rack

And weights are only the floor. Serving a model also holds a KV-cache — the running memory of each active conversation — that grows with context length × concurrent users. At long context and high concurrency it can rival or even exceed the weights, so real serving memory ≈ weights + KV-cache + activation overhead, all of it living in HBM next to the chip. (Training needs several times more again, for gradients and optimizer state.)

Why Mixture-of-Experts is the twist. An MoE model computes with only a few billion active parameters per token — but every expert's weights must still sit resident in HBM in case the router selects it. So DeepSeek-V3 needs ~671 GB of HBM capacity even though it only computes with 37B. That is the crucial link the charts below make visible: total parameters (the top of each stem) set the HBM capacity bill even while active compute stays flat — so total-parameter growth drives HBM demand directly. Ties to §05·E (wafer starts) and §06·D·W (the memory wall).

"How big is the model?" got harder to answer over time, not easier. This plots the parameter count of major model releases since 2022 on a log scale — the only honest way to show a range that runs from Mistral's 7B to Kimi K3's 2.8T. Two things make a naïve "models keep getting bigger" chart misleading, and both are drawn explicitly. (1) The disclosure blackout: OpenAI last published a parameter count for GPT-3 (175B, 2020); Google last for PaLM (540B, 2022); Anthropic has never disclosed one. Every closed frontier model since — GPT-4, 4o, 4.5, the GPT-5 line, Claude, Gemini — is an estimate (hollow markers); the open-weight labs (Meta, DeepSeek, Alibaba, Mistral, Moonshot) are now the only ones reporting real numbers (solid markers). (2) Total ≠ active. Modern frontier models are Mixture-of-Experts: Kimi K2 holds 1.04T parameters but fires only 32B per token; DeepSeek-V3 holds 671B but activates 37B. The stem under each MoE dot drops to that active count. The surprises the chart makes visible: the largest dense model ever shipped is still PaLM (540B) from 2022; GPT-4's leaked ~1.8T (2023) was followed by a smaller GPT-4o (~200B est., 2024); and the parameter frontier only kept climbing because sparsity, not dense scale, pushed totals into the trillions. Added Aug 10 2026 — and read them differently from everything else on the chart. GPT-5 (Aug 7 2025) is plotted at ~1.2T, the geometric midpoint of a published range that runs 300B to 5T — a 17× spread between named sources, drawn as the dashed bar rather than hidden inside a single dot. The high end is the interesting one for this board: it comes from Samsung's own head of memory, Dr Jung-Bae Lee, who put GPT-5 at 3–5T parameters in a SemiCon Taiwan slide — a memory executive sizing the customer's model in public. Claude Mythos (Apr 7 2026) is the weakest point on this page and is marked accordingly with a dashed ring: the only figure in circulation is ~10T total / ~0.8–1.2T active, and it traces back not to Anthropic — which has never published a parameter count for any model — but to third-party cost-and-throughput reverse-deduction, republished by outlets that contradict themselves about whether it was ever disclosed at all. It is on the chart because it was asked for, and it is dashed because it should not be quoted. Two more followed: Claude Fable 5 (Jun 9 2026) — Anthropic's own description, "the first publicly available Mythos-class model" — inherits Mythos's ~10T because that phrase is the only thing connecting it to any number; and GPT-5.6 Sol (Jul 9 2026), for which no estimate has been published by anyone, derived here from GPT-5 by API-price ratio and drawn with a 900B–20T band — a 22× span that is the honest width of what price can tell you. All four new points are guesses with different amounts of scaffolding, and the chart now says which is which.

Parameter count of notable model releases, 2022 → 2026 (log scale)
Each mark = one model release. Y = total parameters (log, billions → trillions); X = release date. Solid mark = a disclosed/official count; hollow mark = an estimate or leak (closed labs don't publish counts). The stem below a mark drops to that model's active parameters per token (Mixture-of-Experts); dense models have no stem. Color = developer. Four closed models were added on Aug 10 2026 by request — GPT-5, Claude Mythos, Claude Fable 5 and GPT-5.6 Sol — reversing this chart's earlier policy of plotting only models with a solid public number. They are drawn with the machinery that reversal requires: a dashed vertical bar spanning the range of published estimates where one exists, and a dashed ring (◌) for a figure that no independent source corroborates. The rest of the closed 2026 line-up (the GPT-5.x point releases, Claude Opus 4.8, Gemini 3.x) still has no usable estimate and remains unplotted; the release timeline below tracks every OpenAI, Anthropic and Google model since 2022, sized or not.
● disclosed count○ estimated / leaked◌ uncorroborated figurestem = MoE active paramsdashed bar = span of published estimates
The price lens — the only quantitative method available, and a dated proof that it does not work
Every estimate on this chart from a closed lab ultimately comes from one idea: serving cost scales with active parameters, so API price is a proxy for size. It is how Alan Thompson derives his GPT-5 range, and with no disclosure to work from it is the only lever anyone has. Here is the whole ladder this board could assemble, in US$ per million tokens: GPT-5 $1.25 / $10  ·  Claude Opus 4.8 $5 / $25  ·  Claude Fable 5 $10 / $50  ·  GPT-5.6 Sol $5 / $30  ·  Claude Mythos ~$30 / $150 (unofficial) From which: Fable 5 is exactly 2× Opus 4.8 on both input and output, and Sol is 3–4× GPT-5. Those ratios are what put Sol at ~4.2T and set Fable 5's active count. Now the fact that ruins it. On Jul 30 2026 OpenAI cut GPT-5.6 prices: Luna −80% ($1.00 → $0.20 input), Terra −20% ($2.50 → $2.00), and Sol unchanged. No model's parameter count changed that day. A method that would have "measured" Luna as five times smaller overnight is measuring commercial strategy, not silicon. TechCrunch made the same point about the base model at launch — "OpenAI priced GPT-5 so low, it may spark a price war" — which means the price lens does not merely add noise, it is biased downward exactly when a lab is buying share. Serving efficiency compounds the problem: better kernels, quantization and batching all cut cost per token while the weights sit unchanged. One independent cross-check that does hold up. §06·D·U measures what it actually costs to run these models through a fixed benchmark: Claude Fable 5 $2.75 per task against GPT-5.6 Sol's $1.04 — Fable 5 is 2.6× more expensive to actually use, for one index point more. That is a different and better measure than list price because it includes reasoning-token behaviour, and it points the same way as the sizes plotted here. It is corroboration of the ordering, not of the numbers.
Provenance for the four added points — four evidence tiers, and these sit in the bottom two
GPT-5 — tier 3, named third-party analysis. Released Aug 7 2025. OpenAI disclosed no parameter count and has not published one since GPT-3 (175B, 2020). Two independent named estimates exist and they do not agree: Alan D. Thompson (LifeArchitect.ai) gives ~300B–3T, derived by inferring size from API price; Samsung's President and Head of Memory Business, Dr Jung-Bae Lee, showed a slide at SemiCon Taiwan (Sep 2024) putting GPT-5 at 3–5T parameters, trained on 7,000× NVIDIA B100. Looser web figures cluster at 1.7–1.8T (mostly recycled from the GPT-4 leak) and one outlet claims ~1T. This board plots 1,200B — the geometric mean of the 300B floor and the 5T ceiling — and draws the whole span, because the honest content of the estimate is the range, not the midpoint. Architecture is not disclosed either. GPT-5 ships as a router across GPT-5 Instant / Thinking / Pro, so "the parameter count of GPT-5" may not name a single network at all — the stem is omitted rather than guessed. Claude Mythos — tier 4, uncorroborated. Anthropic disclosed Mythos on Apr 7 2026 after its existence leaked on Mar 26, described it to Fortune as a "step change in capabilities," and never released it publicly — access went to ~40 firms under Project Glasswing. Anthropic published nothing about its size, and has never published a parameter count for any Claude. Because the model was never publicly served, there is no pricing or throughput signal to reason from either. The ~10T total / ~0.8–1.2T active figure plotted here comes from secondary coverage attributing it to "reverse-deduction analysis" on unnamed technical forums; the most prominent article carrying it labels the same number "disclosed via third-party" in one table and "also-unverified" three paragraphs earlier, and contains unedited drafting notes to itself. That is not a source, it is a rumour with a citation format. It is plotted at the request that prompted this build, at the lowest tier the chart can express. Claude Fable 5 — tier 4, inherited. Released Jun 9 2026, $10/$50, 1M-token context, 128K max output, knowledge cutoff Jan 2026; roughly 80% on SWE-Bench Pro; falls back to Claude Opus 4.8 in cybersecurity, biology and chemistry. Anthropic describes it as "the first publicly available Mythos-class model" — and that sentence is the entire basis for the 10T plotted here. It is an estimate of an estimate: Mythos's own figure is uncorroborated, and "Mythos-class" is a marketing tier, not an architecture statement. The active count is scaled from Mythos's reported pricing and is weaker still. GPT-5.6 Sol — tier 4, derived by this board. Released Jul 9 2026 after a Jun 26 preview, the top of a three-model family (Luna / Terra / Sol), $5/$30, 1.05M-token context, updated again in ChatGPT on Aug 6. No one has published a parameter estimate for it. The 4.2T plotted is this board's own arithmetic — GPT-5's midpoint × the 3–4× price ratio — and the band carries GPT-5's uncertainty through, giving 900B to 20T. That bar is deliberately enormous. It is drawn at full width because the width is the finding: on the only method available, the answer for OpenAI's current flagship is "somewhere between smaller than Kimi K3 and twice the largest number anyone has ever claimed." Two consequences worth naming. First, Anthropic now has three points on this chart — Claude 3.5 Sonnet at ~400B and Mythos and Fable 5 both at ~10T — implying a 25× jump that no number involved is strong enough to support, and a flat Mythos→Fable 5 segment that is an artefact of one inheriting the other rather than a measurement. Do not read a trajectory from any of it. Second, three of the four points added today sit at the bottom evidence tier, which is a fair summary of what is publicly knowable about frontier model size in 2026. The section's original point survives all of this intact: the reason these two dots need this much apparatus is that no frontier lab has published a parameter count in four years, and the chart's job is to show that blackout rather than paper over it.
Every OpenAI, Anthropic & Google release since 2022 — and how few have a published size
The comprehensive view the scatter above can't show: all 48 notable releases from the three biggest closed labs, on one timeline (one lane each). = size disclosed · = size estimated/leaked · = size never disclosed. The punchline is the ratio the lane badges count for you: across all three labs, only four parameter counts were ever published — every one of them a Google open PaLM/Gemma model. Not a single Gemini, GPT-4/5, or Claude has an official size — including the two now carrying estimates on the scatter above, which is exactly why they are marked ◐ here and not ●. Hover any mark for its date and status.
● size disclosed◐ estimated / leaked○ never disclosedbold label = a number exists · faint = size unknown
Who still tells you the size — the disclosure blackout, by lab
OpenAI · last disclosed
GPT-3 · 175B
2020 — GPT-4 onward estimated only
Google · last disclosed
PaLM · 540B
2022 — Gemini line undisclosed
Anthropic
never disclosed
Claude 1 → Fable 5, no count ever
Meta
still discloses
Llama & Llama 4 herd, full specs
DeepSeek · Qwen · Kimi
disclose
open weights — total & active given
xAI
partial
Grok-1 314B open; Grok 3–5 estimated
The full set — transcribed table

Provenance & the honesty line. [disclosed — official papers / model cards / open weights]: Chinchilla 70B, PaLM 540B (dense, Google/DeepMind); OPT-175B, BLOOM 176B, LLaMA 65B, Llama 2 70B, Llama 3.1 405B (dense, Meta/BigScience); Falcon 180B (TII); Mistral 7B, Mixtral 8×7B (46.7B total / 12.9B active); Grok-1 314B MoE (~79B active, weights opened Mar 2024); Nemotron-4 340B (Nvidia); Qwen2.5 72B, Qwen3-235B (235B / 22B active, Alibaba); DeepSeek-V2 (236B / 21B), DeepSeek-V3 (671B / 37B), DeepSeek-V4-Pro (1.6T / 49B); Kimi K2 (1.04T / 32B), Kimi K3 (2.8T total, active not broken out at capture, Moonshot). [estimated / leaked — closed labs, no official count]: GPT-3.5 ~175B (inherited from GPT-3); GPT-4 ~1.8T total / ~280B active (MoE, per The Information leak); PaLM 2 ~340B (reported); GPT-4o ~200B and Claude 3.5 Sonnet ~400B (Epoch AI estimates); Grok 5 ~6T (unverified vendor claim — treat with caution). Not plotted (no credible public number): GPT-4.5, the GPT-5 line, Claude 3.x/4.x/Opus 4.8/Fable 5, Gemini 1–3.x — parameter counts withheld. Honesty caveats: (1) "parameters" is one axis of size — training compute (FLOP) and data are others and have kept rising even where parameters fell; this chart is deliberately the parameter view. (2) MoE totals and dense totals are not like-for-like capability — a 671B sparse model ≠ a 671B dense model; the total/active stem is drawn precisely so the two aren't conflated. (3) estimates are point-in-time and contested; they're marked hollow and excluded from any "record" claim. (4) release dates are first public availability, approximated to the month. Cross-refs: §05·E (wafer starts), §06·D·W (memory wall / KV-cache), §06·D (token demand). Sources: Epoch AI (notable-models data & size estimates), original model papers/cards, The Information (GPT-4 leak), Meta AI, DeepSeek, Moonshot AI, Alibaba, Mistral, xAI. Not investment advice.

06·D·QModel IQ Over Time — and the Contamination Gap◉ live extract · trackingai.org · Jul 30 2026
06·D·Q

Model IQ Over Time

§06·D·M measures models by size. This measures them by score — using Tracking AI (Maxim Lott), which has administered IQ tests to every frontier model on a continuous schedule and publishes the raw daily log. The numbers below are extracted from that log directly: 5,315 dated score rows across 85 models. What makes this source unusually useful is that it runs two tests — and the difference between them is the finding. The "Offline Test" is private: written by a Mensa member, never published, absent from any training corpus. "Mensa Norway" is a public online test that has been on the internet for years and is therefore almost certainly in the training data. Tracked side by side, the public test now reads 151 while the private one reads 136 — and, more importantly, the private score has been flat in a 123–136 band since April 2025 while the public score kept climbing.

Best score achieved by any tracked model, by month — private vs public test
Each point is the highest IQ any model scored that month on that test — a frontier, not an average, so it cannot fall just because a weak model was added. Amber = Offline Test (private, uncontaminated) · violet = Mensa Norway (public, in training data). The shaded band between them is the contamination premium. The dashed line marks April 2025, after which the private frontier stops rising. Hover any month for both scores and which model set them.
Why a memory dashboard cares what an AI scores on an IQ test
Not for the headline. For the shape. This board's entire demand case (§06·D) rests on tokens consumed, and the two ways that grows are more users and more thinking per user. If raw model capability were still compounding, you would expect the private frontier to keep climbing; instead it has been flat for fifteen months in a 123–136 band while the labs shipped GPT-5, Grok-4, Claude Opus 4.x, Gemini 3.x and now the GPT-5.6 line. That is consistent with the thesis this board already runs on: the gains of the last year came from reasoning at inference time — more tokens spent per question (§06·D·W) — rather than from a smarter base model. Memory demand follows the tokens, and tokens have kept compounding (§06·D·R) even as the IQ line went sideways. The bear reading is equally available: a flat capability frontier is exactly what you would expect before a spending re-appraisal, which is kill-switch #3 territory (§06·F·D). This section deliberately does not adjudicate between those two readings — it just makes the flatness visible.
Before the daily log — the earliest models the tracker carries

These are scores for models by their stated release date. The continuous daily log begins May 2024 — Tracking AI was not running in 2022, so no 2022 data exists on this source and none is invented here. The earliest model it carries at all is Bing (Feb 2023).

Provenance & method. Source: Tracking AI (trackingai.org/iq), a project by Maxim Lott. [cited — extracted live]: the page is client-rendered, so the series here was taken from the site's own underlying data file, app/database/proj_IQ/score_logs/iq/daily_logs_iq.csv, read in-browser on Jul 30 20265,315 score rows, 85 models, 2024-05-09 → 2026-07-30, joined to the site's model metadata for release dates. Each plotted point is the maximum iq_score recorded in that calendar month for that test_source; the tooltip names the model that set it. The two tests, in the site's own words: the Offline quiz "is a test made by a Mensa member that has never been on the public internet, and is in no AI training data"; Mensa Norway "is a public online IQ test." Tests are verbalised for text models; vision models are shown the image, and vision variants are scored separately (both appear in the frontier). Honesty caveats — several, and they matter: (1) An IQ score is not a capability measure for a machine. These tests were normed on humans; a model that scores 136 is not "smarter than 99% of people" in any transferable sense, and the number should be read as performance on this specific instrument, nothing more. (2) The public/private gap is evidence of contamination, but not proof — it could also reflect the two tests having different difficulty or item mix. The site's own framing supports the contamination reading; this board reports the gap and the alternative explanation together. (3) Month-to-month noise is real — models are re-tested continuously and refuse questions occasionally (the log carries a num_refused_questions field); single-month moves of ±6 points are common and should not be over-read. (4) Frontier ≠ typical — this is the best score any model achieved, so it says nothing about the median deployed model. (5) No 2022 coverage: the daily log starts May 2024 and the earliest tracked model dates to Feb 2023; the request for a 2022 start cannot be met from this source and has not been faked. (6) Long-lived endpoint names (e.g. "Gemini Pro") are re-tested as the vendor updates them, so a name's score reflects today's endpoint, not its original release — which is precisely why this chart is built on measurement dates rather than release dates. Cross-refs: §06·D·M (size), §06·D·W (inference-time reasoning), §06·D·R (token usage), §06·F·D (the spend this capability is meant to justify). Not investment advice.

06·D·TAI Capability Milestones — A Dated Timeline◷ curated · dated · Aug 6 2026
06·D·T

AI Capability Milestones — A Dated Timeline

Thirty years of AI firsts, each with a date and a specific verifiable claim. This is deliberately not a list of model releases — §06·D·M already charts those by parameter count, and §06·D·Q tracks scores over time. This tracks things that had never been done before: a machine beating a world champion, a fifty-year biology problem falling, a benchmark built to resist AI being solved anyway. The shape is the point. Eight of these eighteen milestones landed in the last two years, against six in the twenty-three years from Deep Blue to GPT-3 — which is why the time axis below is deliberately stretched rather than linear. And the most recent entry cuts the other way: when a new reasoning benchmark launched in March 2026, every frontier model scored under 1% while humans solved all of it.

Capability firsts by date and domain — 1997 → 2026
The time axis is piecewise, and that is a deliberate distortion you should know about: 1995–2015 gets 22% of the width, 2015–2022 another 22%, and 2022–2026 the remaining 56%. On a true linear axis the last four years would occupy an eighth of the chart and nothing in them would be legible. The two amber dashed lines mark where the axis stretches. Larger rings mark the milestones this board considers genuinely discontinuous — where something moved from impossible to done, rather than from good to better. Hover any point for the specific claim and date.
The pattern this timeline actually shows — benchmarks fall, then get rebuilt
Read the red lane on its own and a cycle appears. ARC-AGI was designed specifically to resist memorisation — problems a child can do that models could not. In Dec 2024 o3 scored 87.5% on it, using enormous test-time compute, and the benchmark was declared broken. ARC-AGI-2 was built harder; by Mar 2026 GPT-5.4 was at 73.3% base and 83.3% Pro — largely solved in about four months. ARC-AGI-3 launched Mar 25 2026 and every frontier model scored below 1% — GPT-5.4 0.26%, Claude Opus 4.6 0.25%, Grok-4.20 0.00% — while humans solved every environment in it. Both readings of that are legitimate and this board holds them together. The optimistic one: each benchmark falls faster than the last, and the gap between "impossible" and "saturated" is now measured in months. The sceptical one: a system that scores 83% on one reasoning test and 0.26% on another, while humans do both, is not reasoning in the way the first score implies — it is very good at the shape of problems it has seen. §06·D·Q is the corroborating evidence for the sceptical read: the private, uncontaminated IQ test has been flat in a 123–136 band since April 2025 while the public one kept climbing. Neither observation settles it. What both rule out is the simplest story, in which capability rises smoothly and benchmarks just measure it.
Why a memory dashboard tracks this at all
Every dollar of capex in §06·F·D, every gigawatt in §06·F·C and every bit of the DRAM deficit in §05·D is ultimately a bet that capability keeps arriving. This timeline is the scoreboard for that bet, and it is deliberately kept separate from the spending so the two can be compared rather than assumed to move together. The bull evidence on this page is real: the IMO went from formal-methods silver in Jul 2024 to natural-language gold twelve months later — 5 of 6 problems, 35 points, under the same conditions as human contestants, when only 67 of 630 humans took gold. A diagnostic system solved 85.5% of hard NEJM cases against 20% for experienced physicians. Two Nobel Prizes went to AI work in a single week. The bear evidence is on the same chart: ARC-AGI-3, and the flat private IQ frontier behind it. The honest position is that the capability curve and the spending curve are both real and only one of them is contractually committed. §02·D kill-switch #3 is exactly this question — if capability visibly stalls while §06·F·E's debt coupons keep rising, the financing case gets re-examined long before the compute does.
Every milestone, newest first

Provenance, and the two biases built into any list like this. [cited — dated events]: Deep Blue defeats Kasparov May 11 1997; IBM Watson wins Jeopardy! Feb 16 2011; AlexNet wins ILSVRC Sep 30 2012, cutting top-5 error from ~26% to ~15%; AlphaGo beats Lee Sedol 4–1, Mar 15 2016; “Attention Is All You Need” Jun 12 2017; AlphaGo Zero Oct 18 2017; GPT-3 May 28 2020; AlphaFold 2 at CASP14 Nov 30 2020; ChatGPT Nov 30 2022; GPT-4 Mar 14 2023; Jul 25 2024 AlphaProof + AlphaGeometry 2 score 28/42 (4 of 6 problems) for IMO silver; Oct 8–9 2024 Nobel Prizes in Physics (Hopfield, Hinton) and Chemistry (Hassabis, Jumper, with Baker); Dec 20 2024 o3 scores 87.5% on ARC-AGI-1; Jan 27 2025 the DeepSeek R1 cost shock, with NVDA down ~17% in a session; Jul 2025 Microsoft's MAI-DxO solves 85.5% of 304 NEJM cases versus a 20% average for 21 experienced physicians; Jul 19–21 2025 Gemini Deep Think and an experimental OpenAI model each score 35 points (5 of 6) for IMO gold in natural language under human exam conditions, against 67 golds among 630 human contestants; Mar 2026 GPT-5.4 at 73.3%/83.3% on ARC-AGI-2; Mar 25 2026 ARC-AGI-3 launches with every frontier model under 1%. Bias one — curation. There is no objective list of “breakthroughs.” This one favours events that are dated, singular and checkable, which systematically under-weights slow diffuse progress that mattered more: the steady march of translation quality, speech recognition crossing usability, the compounding of open-source tooling, and the entire history of hardware that made any of it possible. An honest reader should assume the list is incomplete in ways that flatter the dramatic. Bias two — survivorship. A timeline of successes cannot show the failures, and AI has a long record of confident predictions that did not land: two prior AI winters, self-driving timelines missed by roughly a decade, and IBM Watson's own pivot from Jeopardy champion to a healthcare business that was eventually sold off. The 1997 and 2011 entries on this chart are both systems whose commercial successors disappointed — which is worth holding in mind when reading the 2026 entries. Progress on a benchmark is not the same as a durable business, and this board's §08·D exists precisely to measure the gap between the two. Not tracked here, deliberately: model releases and parameter counts (§06·D·M), measured IQ over time (§06·D·Q), token consumption (§06·D), and revenue (§08·D). Cross-refs: §02·D (kill-switches), §06·F·D (the capex this underwrites), §06·D·W (the memory wall these capabilities run into). Sources: DeepMind blog (AlphaGo, AlphaFold, AlphaProof, IMO gold), NobelPrize.org, OpenAI, ARC Prize Foundation, Microsoft Research (Sequential Diagnosis with Language Models), Scientific American, VentureBeat, TechCrunch, and the primary papers. Not investment advice.

06·D·UThe Pareto Frontier — Intelligence vs Cost◷ Artificial Analysis · Jul 17 2026
06·D·U

The Pareto Frontier — Intelligence vs Cost

Every other model comparison on this board ranks by capability alone. This one asks the question a buyer actually faces: for a given budget, what is the smartest model you can run — and which models are simply never the right answer? A model sits on the Pareto frontier if nothing else is both at least as capable and cheaper. Everything off the frontier is dominated: there exists another model that beats it on both axes at once, so no rational buyer picks it on price-performance grounds. Of the eight frontier-class models with both figures published, five are on the frontier and three are dominated — and the frontier's shape carries the finding: the cheap end is astonishingly cheap, and the last point of intelligence costs more than the first nine combined.

Intelligence Index vs cost per Index run — the frontier and what it dominates
Up is smarter, left is cheaper, so the best place to be is the top-left. The teal staircase is the Pareto frontier: at every price level it marks the most capable model money can buy. Filled circles with a ring are on the frontier; hollow circles are dominated, each joined by a dashed line to the model that beats it on both axes. The shaded region is everything that is simultaneously worse and dearer than some available alternative. The x-axis is log scale — necessary because the frontier spans 13× in price, and on a linear axis the four cheapest models would sit on top of each other. Hover any point for its exact position and, if dominated, by what.
The shape of the frontier is the whole argument — intelligence is cheap until it is the best
Walk the staircase from the cheap end and the marginal price of capability does something violent. Points 51 → 54 cost about $0.033 each — Grok 4.5 buys three index points over GPT-5.6 Luna for ten cents. Points 57 → 59 cost $0.05 each. Then the final point, from GPT-5.6 Sol at 59 to Claude Fable 5 at 60, costs $1.71 — a 164% price increase for 1.7% more index, and roughly 51× the marginal cost of the points at the bottom. That convexity is the single most important fact for anyone modelling AI spend. It means the token-demand curves in §06·D are not one market but at least two: a volume market where near-frontier intelligence is nearly free and usage scales with almost anything, and a frontier market where the last increment is priced like a luxury good and is bought only where it decides the outcome. Both are real and they behave completely differently under a price shock — which is exactly the distinction §06·D·W draws between tokens that must run on HBM and tokens that can run anywhere.
What being "dominated" does and does not mean
Three of the eight are dominated, and one of them is awkward: Claude Opus 4.8 is beaten by GPT-5.6 Sol, which is three index points better and 42% cheaper — so Anthropic holds both the frontier leader (Fable 5) and a strictly dominated model at the same moment. Muse Spark 1.1 and GLM-5.2 both sit at index 51, where GPT-5.6 Luna does the same job for less. But "dominated on this chart" is a narrow claim and should not be over-read. The Index is one composite score; a model can lose on it and still be the right choice for a specific task — Artificial Analysis publishes separate coding-agent and long-horizon knowledge-work indices where the ordering differs. Price per Index run is also not price per your workload: it embeds a particular input/output token mix, and a very output-heavy or cache-friendly application will reorder these. Nothing here accounts for rate limits, context length, latency, data residency, deployability on your own hardware, or whether the weights are open — GLM-5.2 and Kimi K3 being open-weight is a category of value this axis cannot see at all. The honest claim is narrow and still useful: on published composite intelligence per published dollar, these three are beaten outright, and that is a fact a buyer should have to argue against rather than ignore.
Every model, ranked by index

Provenance, and four reasons this frontier moves under you. [cited — single source, both axes]: all sixteen numbers here (eight models × index and cost) come from Artificial Analysis, “Four frontier launches in eight days,” published Jul 17 2026. Using one source for both axes is deliberate — mixing a benchmark from one house with pricing from another is the most common way this chart gets built wrong, because the cost figure has to embed the same task set the score was measured on. Cited values: Claude Fable 5 (max) 60 / $2.75; GPT-5.6 Sol (max) 59 / $1.04; Kimi K3 57 / $0.94; Claude Opus 4.8 (max) 56 / $1.80; Grok 4.5 (high) 54 / $0.31; GPT-5.6 Luna (max) 51 / $0.21; Muse Spark 1.1 (xhigh) 51 / $0.26; GLM-5.2 (max) 51 / $0.32. Frontier membership is computed here, not taken from the source — a model is marked dominated only if another in this set is at least as high on the index and no more expensive. Four reasons to treat this as a snapshot with a short half-life: (1) The index itself was revised after these numbers were published. Artificial Analysis shipped Intelligence Index v4.1.1 on Aug 6 2026, upgrading grader models and one benchmark component — so these Jul 17 scores are pre-patch and the ordering could shift on re-scoring. (2) The frontier moved four times in eight days. The source article exists because Grok 4.5, GPT-5.6, Muse Spark 1.1 and Kimi K3 all launched inside a week, taking the count of labs above index 50 from two to six. Anything built on this chart should assume it is stale within weeks. (3) One model is missing for a good reason. GPT-5.6 Terra (max) scores 55 but no cost-per-task was published, and a Pareto analysis with a guessed coordinate is worse than one with an acknowledged hole. (4) Reasoning-effort settings are part of the identity. “max”, “high” and “xhigh” are different operating points of the same model with different costs and scores; comparing a max-effort configuration against another model's default is not a like-for-like comparison, which is why each label carries its setting. Related but distinct sections: §06·D·M ranks by model size, §06·D·Q by measured IQ over time, §06·D·R by actual usage share on OpenRouter, and §06·D·T by dated capability firsts. This is the only one that prices capability. Source: Artificial Analysis, Jul 17 2026; index revision notice, Aug 6 2026. Not investment advice.

06·EMemory Shipped in Gigabytes◷ quarterly model
06·E

Memory Shipped in Gigabytes

◷ quarterly model

Quarterly shipped capacity by memory/storage type, 2022 → 2030E, in billion gigabytes (Bn-GB) — numerically equal to exabytes (EB). The tiers differ by orders of magnitude (HDD ships ~10× the bytes of all DRAM), so they're split into two charts: compute / active memory below, then storage media beneath it — otherwise HBM would be an invisible sliver next to nearline HDD. This is an industry model, not audited company-reported shipment data.

Compute / Active Memory — quarterly shipped capacity
HBM · DDR (server + desktop) · LPDDR (mobile + AI superchips) · GDDR (graphics) · CXL (memory-expansion modules). Bn-GB per quarter. Click a legend item to toggle it.
Storage Media — quarterly shipped capacity
NAND / SSD (enterprise + client) · HDD / nearline · Tape / archive. Note the y-axis scale is ~10× larger than the active-memory chart above — storage media ships far more raw capacity. Bn-GB per quarter.

Scale anchors (real): HDD shipped ~900 EB in 2023 and ~250–300 EB/quarter (Coughlin/Nidec); SSD ~90 EB in Q1 2024, roughly 12× DRAM capacity and on par with total HDD bytes (Yole/Forward Insights); DRAM ~30 EB/quarter with HBM ~6% of DRAM bits (industry estimates); tape a small but growing archive tier. Forward quarters (2026Q3→2030Q4) extrapolate each tier's bit-growth CAGR — DRAM/HBM fastest, HDD steady, tape slow. These are modeled proportions calibrated to published scale relationships, not audited shipment reports. Quarterly splits use observed seasonality. Values are shipped capacity (supply), not installed base. 1 Bn-GB = 1 EB = 1 billion gigabytes. Sources: Yole Intelligence, Forward Insights, IDC, Coughlin Associates, TrendForce, company filings. Not investment advice.

06·FCompany CapEx → New Supply◷ model
06·F

Company CapEx → New Supply

◷ model

A section per maker: past and projected capital expenditure, and a model translating that capex into estimated new memory supply by type. The logic is the industry's own: capex buys wafer-fab equipment and cleanroom; at a known capital intensity (dollars per EB of annual bit capacity) that yields new bits — but with a 6–8 quarter lag, which is exactly why 2025–26's record spend doesn't relieve the shortage until 2027–28.

The model, made explicit — applied per company, per memory type:
new annual supply (EB) = capex allocated to type ÷ capital intensity ($/EB) × efficiency arrives ~6–8 quarters later
Capital intensity anchors (calibrated to industry bit-growth): DRAM ≈ $0.7B per new EB/yr · HBM ≈ $2.1B per equivalent EB (≈3× DRAM wafer intensity per usable bit) · NAND ≈ $0.32B per new EB/yr (cheaper per bit, denser 3D stacking). These are blended/approximate — real intensity varies by node, yield, and how much capex is maintenance vs. greenfield. They're tuned so the aggregate reproduces the industry's real ~20–30 EB/yr of net-new DRAM.
Aggregate Modeled New Supply — all makers combined

Summing the per-company models: estimated net new annual bit supply coming online each year (lagged from capex), by type. This is the supply side of the bit-gap — the wave that closes the shortage in 2027–28.

EB of net-new annual supply arriving each year, by type, from modeled capex. Forward years reflect committed + projected capex.

CapEx anchors (real, from filings & TrendForce): Micron FY22 $12.0B → FY23 ~$7.0B (cut ~40% in the downturn) → FY24 ~$8.1B → FY25 ~$13.8B → FY26 ~$27B (revised sharply up; $7.1B in FQ3 alone per the Jun 24 FQ3 report), with further increases planned for FY27; ~$200B US long-term plan (Idaho greenfield output now expected calendar 2028). SK Hynix ~$12.5B (2024) → ~$21B (2025, +30% on HBM) → ~$20.5B (2026E); M15X output begins 2H 2026, Yongin cluster from 2027. Samsung ~$18B → $20B (2026E); P4L expansion. Industry DRAM capex $53.7B (2025) → $61.3B (2026, +14%); NAND $21.1B → $22.2B. The capex figures are well-sourced; the capex→supply translation is a model. Capital intensity is blended and approximate, lag is a simplification (real fab ramps are gradual, not a step), and "efficiency" bundles yield, node mix, and maintenance-vs-growth capex into one factor. Directionally sound — the supply wave is real and lands 2027–28 — but the EB figures are estimates, not forecasts. Sources: company 10-Ks/8-Ks, TrendForce, The Elec, Coughlin Associates. Not investment advice.

06·F·DHyperscaler CapEx — New Demand◷ Q1–Q2 2026 disclosed · guidance-projected
06·F·D

Hyperscaler CapEx — New Demand

The mirror of the section above. §06·F tracks what the memory makers spend to create new supply; this tracks what the buyers spend to create new demand — quarterly data-centre capital expenditure at Microsoft, Alphabet, Amazon, Meta and Oracle. The scale has stopped being comparable to anything in prior computing history: these five spent ~$233B in 2024, ~$406B in 2025, are guiding to ~$808B in 2026, and on current trajectory cross $1 trillion a year during 2027. The single most important datapoint for this dashboard arrived on Jul 30 2026: Amazon raised its 2026 capex from $200B to $220B and told investors the increase was driven by higher memory prices — the supercycle this board tracks, showing up as a line item in a hyperscaler's budget.

Jul 22–30 2026 — the earnings fortnight that broke the consensus
For two years every hyperscaler raised capex together. This reporting cycle, they diverged for the first time — and the market noticed, selling the sector after Alphabet's print on capex-scrutiny grounds. Alphabet raised FY26 guidance to $195–205B (from $180–190B), its CFO citing "an acceleration in the delivery of capacity to meet growing demand." Meta raised to $130–145B. Amazon raised to $220B — explicitly on memory costs. But Microsoft cut its calendar-2026 plan from ~$190B to ~$175B, even while reporting a record $41B quarter and an $678B backlog (§06·F·B). Read together: demand is not slowing, but at least one buyer is now optimising cost rather than racing on volume — and Amazon's raise says part of what's left of the race is being paid to the memory industry rather than converted into capacity. That is kill-switch #3's exact tension: capex that keeps rising in dollars can still stop rising in bits if input prices are what's climbing.
Quarterly data-centre capex, five hyperscalers — 2024 → 2027
Stacked quarterly capital expenditure, $B (including finance leases where companies report them that way). The teal band (Q1–Q2 2026) is disclosed quarterly prints; bars to the left are modeled from audited annual totals with a within-year ramp (companies did not all break out quarterly capex consistently in 2024–25); bars in the amber band are projected by allocating each company's own full-year guidance across the remaining quarters. Hover any quarter for the company split and its tier.
The guidance moves that made this cycle's first divergence
From dollars to memory — the bridge, and why the ratio is the whole argument. Industry teardowns put memory at roughly 8–15% of an AI server's bill of materials, rising with HBM content per accelerator (§05·C). Applied to the ~$808B of 2026 hyperscaler capex — of which the large majority is data-centre and IT equipment rather than land and shell — that implies something on the order of $45–90B of memory purchased by these five firms in 2026 alone, against a total DRAM industry that was ~$100B in 2024. [Derived, wide, and deliberately shown as a range] — the BoM share is the contested input, the split between shell/power and silicon is not disclosed per company, and Oracle's gross-vs-net figures differ by $20–25B of customer prepayments. The reason to compute it anyway: it is the only line that connects this section's dollars to §05's wafers. And Amazon's Jul 30 raise is the first time a hyperscaler has publicly attributed a capex increase to that ratio moving against it.

Provenance & the honesty line — three data tiers, labelled on the chart. [● cited — disclosed quarterly prints]: Q2 2026 — Amazon $54.2B (vs $32.1B a year earlier), Alphabet $44.9B, Microsoft $41B including finance leases (FQ4, +69% YoY), Meta $31.1B including finance leases; Q1 2026 — Amazon $44.2B, Alphabet $35.7B, Meta $19.84B. Those five firms' Q2 spend alone totals roughly $190B in a single quarter, an annualised run-rate above $760B. [◷ FY guidance — company-stated]: Amazon $220B for 2026, raised from $200B on Jul 30 with CEO Andy Jassy attributing the increase to rising memory prices; Alphabet $195–205B, raised Jul 22 from $180–190B (CFO Anat Ashkenazi: acceleration of capacity delivery); Meta $130–145B, narrowed upward from $125–145B; Microsoft ~$175B for calendar 2026, reduced from ~$190B guided in April; Oracle FY2027 gross $90–95B (net ~$70B after $20–25B of customer prepayments), after $55.7B in FY2026 — itself 2.6× the prior year's $21.2B — funding ~3 GW of new GPU capacity. [modeled]: 2024 and 2025 quarterly bars are apportioned from audited annual totals (~$233B and ~$406B for the five) using a rising within-year ramp; Microsoft's and Oracle's Q1/Q2 2026 bars are likewise derived because their fiscal calendars do not align to calendar quarters (Microsoft's FY ends June, Oracle's May). Four caveats: (1) capex ≠ AI capex — these totals include non-AI data centre, network, office and fulfilment spend (materially so for Amazon), so treat them as an upper bound on AI infrastructure. (2) Fiscal-calendar mismatch — Microsoft and Oracle are re-cut to calendar quarters, an approximation. (3) Finance leases are included by some firms and not others; where a company reports "capex including finance leases" that is the figure used, which inflates comparability slightly. (4) 2027 is extrapolation, not guidance — only Oracle has published an FY27 number; the other four are grown at ~25%, and the "$1 trillion" crossing is this board's arithmetic, not a company statement. Estimates for the combined 2026 figure range $700B–$900B across houses depending on whether Oracle is included and how finance leases are treated; ~$808B here sits mid-range. Cross-refs: §06·F (the supply mirror — memory-maker capex), §06·F·B (backlogs — the contracted revenue this capex serves), §06·F·C (the same spend measured in gigawatts), §02·D KS#3 (a capex cut is the thesis's clearest early warning). Sources: company Q2 2026 earnings releases and calls (Amazon, Alphabet, Microsoft, Meta, Oracle), CNBC, Reuters, Sherwood, Fortune, Statista, Epoch AI, CreditSights, Futurum, FactSet. Not investment advice. See also: §06·F·G, the revenue line underneath all of this spending — whether the cloud businesses these budgets serve are still growing. See also: §06·F·N, the depreciation assumption that decides what this capex actually costs in reported earnings, and §06·F·P, whether the power to run it can physically arrive on schedule. See also: §06·F·O, the optical layer inside these budgets, where two published forecasts for the same component in the same year differ by six times.

06·F·EHyperscaler Debt — On and Off Balance Sheet◷ recent quarters cited · off-BS on mixed bases
06·F·E

Hyperscaler Debt — On & Off Balance Sheet

§06·F·D shows what the hyperscalers are spending. This shows how they are paying for it — and the answer has changed. For a decade these were famously net-cash businesses that funded everything from operating cash flow; that "unspoken contract" with investors broke in 2025–26. On-balance-sheet financial debt across Microsoft, Alphabet, Amazon, Meta and Oracle has gone from ~$204B to ~$499B in five years, the five sold $159B of bonds in the first five months of 2026 alone (+47% YoY), and Alphabet — long effectively debt-free — now carries $98.2B. But the balance sheet is the smaller half of the story: Moody's counts ~$662B of signed data-centre leases sitting off balance sheet, and Meta has pushed its largest project into a $27.3B special-purpose vehicle whose debt will never appear in its accounts.

Why this section exists — the memory cycle now has a credit channel
Every other section of this board asks whether AI demand is real. This one asks a different question: what happens to memory orders if the financing stops? The DRAM cycle has historically been broken by demand collapse or supply flood — but this build-out is increasingly funded by credit, which introduces a third failure mode that has nothing to do with tokens or wafers. Moody's warned on Jul 24 2026 that "unprecedented" AI spending threatens the credit quality of Amazon, Meta and Alphabet; S&P has already cut Oracle to BBB−, one notch above junk, with 2026 capex running at roughly twice its operating cash flow. Alphabet raised $84.75B of equity in June 2026 — the largest equity transaction ever by a listed company — rather than add more debt. If spreads widen or the SPV market closes, capex plans get cut before demand does, and §06·F·D's $808B becomes the swing factor for §05's wafers. That is kill-switch #3 arriving through the bond market rather than the order book.
On-balance-sheet financial debt, quarterly — five hyperscalers, 2021 → 2026
Stacked quarterly financial debt — bonds, loans and finance-lease liabilities — in $B. The teal band marks quarters anchored to reported figures; earlier bars are modeled from annual filings, since not every issuer breaks total debt out quarterly on a consistent basis. This series deliberately excludes operating leases, purchase commitments and SPV debt — those are counted separately below, because mixing them is exactly how the headline numbers in this topic get to disagree by 3×.
Five years, five balance sheets
Off balance sheet — what the accounts don't show EACH ON ITS OWN BASIS — DO NOT SUM
These figures come from different studies using different definitions, and are listed side by side without being added together or divided into the chart above. Read each on its own terms.
What each bond was issued at — coupon by pricing date CITED TRANCHES ONLY
The charts above measure how much these companies borrowed. This one measures what it cost them — the coupon each bond was actually issued at, plotted against the day it priced. Each deal prices on a single date, so its tranches stack into a vertical ladder, and the height of that ladder is the deal's term structure: short paper at the bottom, 30- and 40-year paper at the top. Dot area is proportional to tranche size. The dashed amber line joins the longest tranche of each deal — the cleanest like-for-like read, because it holds tenor roughly constant and lets the date do the work. This is the coupon at issuance, fixed for the life of the bond — not the current yield and not the spread. That matters: spreads over Treasuries stayed relatively contained through this window, so a rising coupon is mostly the risk-free curve repricing rather than the market deciding these are worse credits. Both effects are real; only the coupon is what the company actually pays.
The trend the ladders show — borrowing got more expensive while the borrowing got bigger
Hold tenor constant and the direction is unambiguous. Forty-year paper went from Meta at 5.75% (Oct 2025) to Amazon at 6.25% (Jul 2026) — +50bp for the same maturity, eight months apart. Amazon can also be compared to itself: its Nov-2025 three-year priced at 3.90%; the equivalent tranche in July came at 4.60%. Alphabet's century bond is the emblem of the window — £1B at 6.125% to 2126, tech's first 100-year bond since Motorola in 1997, drawn nearly ten times oversubscribed at 120bp over gilts. It has since traded below 90 pence, which is a lesson in duration risk rather than credit: nothing about Alphabet changed, the discount rate did. And the July Amazon deal is where the demand side finally showed strain. Amazon had to offer 18–21bp of extra yield on its longest bonds, and orders came in at just 2.5× the size on offer, down from 3.2× in March — what BofA called the weakest hyperscaler new-issue performance since Meta's October 2025 deal, with hyperscaler spreads widening 6–15bp that day and the 10-year Treasury yield lifting 8bp on the back of it. "Investors are pushing back," BofA wrote. Scale for context: AI-related issuance ran $136B for all of 2025, reached ~$270B by mid-July 2026 (~$194B of it hyperscalers), and is forecast near $570B for the full year. Amazon alone had issued $92B in 2026 by July — more than Alphabet, Meta or Oracle individually. This is the financing channel behind the capex in §06·F·D and the backlogs in §06·F·B, and its price is now visibly rising.
The definitional trap, stated plainly. The widely-quoted headline is that these five carry $1.65 trillion of "hidden" debt, equal to 122% of what's on their balance sheets. That ratio implies an on-balance-sheet base of about $1.35T — roughly 2.7× the ~$499B of financial debt plotted above. The gap is not an error in either number: the study's base includes lease liabilities, purchase obligations and other items that this chart deliberately excludes. So the one thing you must not do is divide $1.65T by this chart's $499B and report a 3.3× ratio. Compare like with like, or state the basis. This board plots the narrow, unambiguous measure and lists the broader ones separately — the disagreement between them is the finding, and it is why "how levered is Big Tech?" currently has answers ranging from "barely" to "dangerously" depending entirely on who is counting.

Coupon chart — provenance & what is deliberately absent. [cited — each tranche individually reported]: Oracle Sep 24 2025, $18B/7 tranches, coupons reported as spanning 4.45%–6.1% with the 2032 at 4.8%, the 10-year at 5.2% and the 20-year at 5.875%. Meta Oct 30 2025, $30B/6 tranches — 4.20% (5yr), 4.60% (7yr), 4.875% (10yr), 5.50% (20yr), 5.625% (30yr), 5.75% (40yr); $125B of peak orders, the biggest corporate deal of 2025. Alphabet Nov 6 2025, $17.5B USD — 3.875% (2028), 4.10% (2030), 4.375% (2032), 4.70% (2035), 5.35% (2045), 5.45% (2055), 5.70% (2075), plus a floating tranche at SOFR+0.52%. Amazon Nov 17 2025, $15B — 3.90%, 4.10%, 4.35%, 4.65%, 5.45%, 5.55%. Alphabet Feb 9 2026, ~$20B multi-currency — 40-year at 5.75% and the £1B 100-year at 6.125% due Feb 13 2126, priced +120bp over gilts. Amazon Jul 7 2026, $24.923B across 8 tranches, coupons reported as running 4.600% (2029) to 6.250% (2066). What is NOT plotted, and why: (1) Un-reported middle tranches. Oracle's 30-year and the six intermediate tranches of Amazon's July deal had no individually-published coupon in the sources used, so only the reported endpoints appear — the gaps in those ladders are missing data, not missing bonds, and nothing has been interpolated onto the chart to fill them. (2) Floating-rate tranches (Alphabet's 2028 FRN, Amazon's $750M SOFR+58bp) have no fixed coupon and cannot be placed on this axis. (3) Alphabet's €6.5B euro tranches priced the same week as its dollar deal at 2.375%–4.375% — roughly 150bp below the dollar ladder. That gap is currency, not credit, so folding them into the same ladder would have manufactured a fake improvement; they are held out and noted here instead. Alphabet also issued in yen (1.965%–4.599%) and Swiss francs, excluded for the same reason. (4) Microsoft appears in the debt chart above but has no individually-cited AI-era bond tranche in these sources, so it is absent from this chart entirely — an absence of data, not an absence of borrowing. Basis note: the y-axis is the fixed coupon at issuance, which is what the issuer pays for the life of the bond; it is not yield-to-maturity, not current market yield, and not spread over Treasuries. A bond issued at 6.125% and now trading at 90 pence still pays 6.125% on par. Sources: Bloomberg, Fortune, PitchBook, Cleary Gottlieb, Simpson Thacher, cbonds, Law360, CNBC, Forbes, US News, cryptobriefing, IFR, TwentyFour AM, and BofA research as reported. Not investment advice. See also: §06·F·N — a firm that shortens its useful-life assumption reports lower earnings against this same interest bill, which is the configuration in which a capex plan gets revisited. See also: §06·F·R — the same physical assets financed by borrowers who cannot issue investment-grade paper against a balance sheet, only against the hardware and the contracts on it.

Provenance & the honesty line. [cited — reported]: Alphabet total debt $98.165B at Jun 30 2026; Meta $84B at Q1 2026, from zero before its first bond in Aug 2022; Microsoft $26.0B of finance-lease right-of-use assets added over the trailing four quarters; Oracle raised $43B of debt and $5B of equity in FY2026 with ~$40B combined planned for FY2027; the five sold $159B of bonds in the first five months of 2026, +47% YoY, with UBS projecting full-year issuance of $230–240B and the group potentially exceeding 5% of the entire IG index by year-end; 2024–26 issuance totals of roughly Meta $30B, Alphabet $31B, Amazon $24.9B; Alphabet's June 2026 equity raise priced at $84.75B. [cited — off balance sheet, mixed bases]: Moody's counts ~$662B of signed data-centre leases off balance sheet and ~$1.2T of total lease commitments, of which >$820B has not yet commenced; Meta's own 10-Q discloses $182.88B of operating and finance leases not yet commenced at Mar 31 2026; Meta's Hyperion / "Project Beignet" JV with Blue Owl is $27.3B (Blue Owl 80% / Meta 20%; $27B of A+ rated debt plus $2.5B equity, anchored by PIMCO $18B and BlackRock $3B), with Meta retaining operational control and leasing the campus back — converting what would have been capex into operating expense; a further ~$13B SPV is being raised for a Texas site; Oracle's total disclosed obligations reach $273.3B (May 2026). [modeled]: quarterly bars before Q1 2026 are apportioned from annual filings — issuers do not all disclose total debt quarterly on a consistent basis, so read the shape before any single early bar. Four caveats: (1) definition drives the answer — see the box above; the $1.65T figure and this chart are not on the same basis and must not be combined. (2) Finance vs operating leases: finance leases sit inside this chart, operating leases do not, and the boundary between them is an accounting judgement that materially changes reported leverage. (3) SPV debt is genuinely non-recourse in form but economically supported by the tenant's lease — whether that is "hidden" leverage or ordinary project finance is a live argument, not a settled fact, and this board does not adjudicate it. (4) Fiscal calendars differ (Microsoft June, Oracle May), so quarters are calendar-approximated. Cross-refs: §06·F·D (what the debt is funding), §06·F·B (the contracted revenue meant to service it), §02·D KS#3 (financing stress as the leading indicator of a capex cut). Sources: company 10-Q/10-K filings, Moody's, S&P Global, Nikkei, CNBC, FactSet, UBS, CreditSights, Bisnow, Global Data Center Hub, Business Standard. Not investment advice.

06·F·GCloud Revenue Growth — AWS · Azure · Google Cloud, 5 Years◷ company results · to 26Q2
06·F·G

Cloud Revenue Growth — AWS · Azure · Google Cloud, 5 Years

This is the revenue line underneath everything else on this board. §06·F·D tracks what the hyperscalers spend, §06·F·B what they have booked, §06·F·E the debt funding it — this tracks whether the business those bets are placed on is actually growing. Twenty quarters, three clouds, year-on-year growth. The shape is a V. All three peaked near or above 40% in late 2021, collapsed through the 2023 "optimisation" trough — AWS bottomed at +12% — and have since re-accelerated hard on AI. The latest quarter reads Google Cloud +82%, Azure +43%, AWS +37%, with all three faster than the quarter before. AWS’s +37% is its best since 2021 by four-tenths of a point — 22Q1 was +36.6% — so the honest statement is that AWS has clawed back to where it was before the digestion trough, not that it has broken new ground. But the most useful fact here is one the percentages hide: at +37% and +82% respectively, AWS and Google Cloud added almost exactly the same number of dollars in the quarter.

Year-on-year revenue growth by calendar quarter — 21Q3 → 26Q2
Each line is year-on-year percentage growth, so a falling line still means revenue rising — just more slowly. The shaded band is 2023, when enterprises paused and "optimised" cloud commitments and every provider's growth roughly halved; it is the most useful reference point on the chart because it is the only recent period where cloud demand genuinely broke. Filled dots mark figures cited individually in this build; the earlier points are compiled from reported quarterly results and carry the lighter provenance noted below. Read the right-hand end carefully: the three lines are diverging, and the divergence is mostly about denominator size rather than momentum — see the block below the chart.
82% versus 37% — and the same dollars. The most misread comparison in cloud reporting
In 26Q2 Google Cloud grew 82% to $24.80B and AWS grew 37% to $42.23B. Those headlines invite the conclusion that Google is running away with the market. Do the subtraction instead. Google Cloud's prior-year quarter was $13.63B, so it added $11.17B. AWS's was $30.82B, so it added $11.41B. A 45-point gap in growth rate, and a 2% gap in dollars. This is not a trick, it is arithmetic that percentage reporting systematically obscures: the same absolute increment is a far larger percentage of a smaller base. Both readings are true and they answer different questions. If the question is who is winning share, the growth rate is right and Google Cloud is winning it — its share of the $142B quarterly cloud-infrastructure market rose to 15% while AWS held 28%. If the question is whose incremental demand is driving the semiconductor order book — which is this board's question — the dollars are what matter, and on that measure AWS and Google Cloud are contributing equally. Azure cannot be put through this test at all, because Microsoft does not publish Azure revenue quarterly.
The five-year arc, at the moments that mattered
Why this is the section to watch if you want to know when the capex stops
Cloud revenue growth is the closest thing this board has to a demand signal that cannot be pre-booked. Backlogs (§06·F·B) are contracted future revenue and can grow while current demand softens. Capex (§06·F·D) is a decision about the future made today. Memory prices (§03·B) are a supply story as much as a demand one. Recognised cloud revenue is money that customers have already spent, and it is the one series that would break first if AI demand were more announced than real. Right now it is not breaking — it is accelerating on all three clouds simultaneously, which is the strongest single piece of evidence for the bull case on this board. The 2023 band is why it is drawn. That episode is the proof that this line can halve in four quarters without any AI narrative changing, and it is the specific shape kill-switch #3 (§02·D) is watching for. TrendForce's own caveat in §05·D points at the mechanism: CSP capex is at record levels with some providers reporting negative free cash flow, and if memory prices stay elevated, memory takes a larger share of a budget that has to be funded from somewhere. If cloud revenue growth rolls over while §06·F·E's debt coupons keep rising, the financing case gets re-examined long before the compute does. That is the sequence to watch, and this chart is the first place it would show.

Provenance — two tiers, and the difference matters. [cited — verified in this build]: 26Q2: AWS +37% at $42.23B, described as its strongest expansion since 2021, accelerating from +28% in 26Q1; Azure +43%, from +40% in the March quarter; Google Cloud $24.80B, +82%, from +63%. Total cloud-infrastructure revenue $142B in the quarter, +43% YoY, with shares of AWS 28%, Azure 22%, GCP 15%. Google Cloud back-history: 25Q1 +28% ($12.3B), 25Q2 +32%, 25Q3 above 30%, 25Q4 +47.8% ($17.66B), with backlog more than doubling to $240B entering 2026. AWS: FY2024 +19% ($107.6B, first year above $100B), FY2025 +20% ($128.7B), 25Q4 +24% — its fastest in 13 quarters. Azure: FY2025 (to Jun 30 2025) +34%, passing $75B; FY2026 +41%, passing $100B. One independent check on the compiled tier: Amazon described 25Q4’s +24% as its fastest in 13 quarters. That claim was not used to build the series — but reading it back off the finished chart, the last quarter faster than +24% is 22Q3 at +27%, exactly 13 quarters earlier. A cited qualitative claim reproducing itself from independently-assembled numbers is weak evidence the back-history is coherent, and it is offered as exactly that: weak, but real. The one place rounding mattered: AWS 22Q1 is carried at +36.6% (reported $18.44B against $13.50B), not the rounded 37% — because at 37% it would tie 26Q2 exactly and make the cited "strongest since 2021" claim untestable on this chart. It is the closest comparison on the page and the margin is four-tenths of a point. [compiled — lighter provenance]: the quarters before 25Q1 are reported figures assembled from company results rather than individually re-verified in this build, and are drawn without dots. They are widely published and the shape they describe — a ~40% peak in late 2021, a 2023 trough, a 2024 plateau — is not in dispute, but an individual pre-2025 quarter here should be treated as approximate to within a point or two and checked against the source filing before being quoted. Every point's tier is in the tooltip. The Azure series was additionally fitted to reconcile with the two cited fiscal-year figures (+34% FY25, +41% FY26), since Microsoft's fiscal year runs July–June and its quarterly rates must average to them. Three structural caveats: (1) Azure has no level. Microsoft reports an Azure growth rate and, since 2025, occasional annual milestones — but no quarterly revenue. Its line therefore cannot be converted to dollars, cannot be checked by subtraction, and is the least verifiable of the three. (2) The three are not the same product. AWS and Google Cloud report a cloud segment including software and services; "Azure and other cloud services" is a Microsoft-defined grouping inside Intelligent Cloud. Segment definitions differ and none of the three is a pure infrastructure number. (3) Growth rates are not comparable across different bases — the point of the block above — and a decelerating line on this chart can still mean accelerating dollars. Cross-refs: §06·F·D (capex), §06·F·B (backlogs/RPO), §08·G (where the cost lands on consumers), §06·F·E (the debt), §05·D (the supply side), §02·D (kill-switch #3), §06·G (what that capacity rents for), §05·F (the wafers underneath it). Sources: CNBC on AWS 26Q2, Alphabet and Microsoft quarterly releases, Computer Weekly, Data Center Dynamics, Statista. Not investment advice. See also: §10·C, the provenance index, which lists every section's sourcing tier in one table — including this one's.

06·F·NGPU Depreciation & Useful Life — The Assumption Under the Whole Trade◷ filings + Q1 2026 · Aug 14 2026
06·F·N

GPU Depreciation & Useful Life

This board has tracked what the hyperscalers spend (§06·F·D), what they have booked (§06·F·B), what they borrowed (§06·F·E) and how much power it needs (§06·F·C). It has never tracked the single accounting choice that decides whether any of it was profitable. A GPU bought for $30,000 is not an expense — it is an asset, expensed a slice at a time over an assumed useful life. Choose six years and you book $5,000 of cost a year. Choose three and you book $10,000. Same chip, same cash out the door, double the reported cost. In the four quarters to March 2026 these four firms bought $433.9B of property and equipment and recognised about $149B of depreciation — a 2.9× gap that is pure timing, and whose eventual size depends entirely on a number each company picks for itself. They used to agree. As of 2025 they no longer do: Amazon shortened its assumption and took a charge, Meta extended hers, Microsoft and Alphabet held at six. When the people who own the assets stop agreeing on how long they last, that disagreement is the story.

One year of buying, five possible depreciation bills
Each bar takes the same $433.9B of purchases and divides it by a different assumed life — straight-line, which is what these companies use. Amber bars are lives somebody actually books today; the red bars are Michael Burry's claim that the real economic life of AI silicon is two to three years; the grey bar is a life nobody uses. The teal line is what was actually recognised — and it sits below even the six-year bar because most of that $433.9B has not been placed in service yet. That is the honest reading of the gap: it is a lag, not a lie. But the lag ends, and when it does the annual bill is whichever bar the company chose.
Who assumes what — and who changed their mind
Between 2020 and 2024 every one of these firms moved the same direction, extending server life from three or four years toward six and cutting reported depreciation by roughly $18B a year collectively. 2025 broke the consensus, and the two firms that moved went opposite ways.
The argument, stated by both sides, because this board does not get to settle it
The bear case (Burry and others): Nvidia ships a materially better part roughly every year — §06·D·M's own chart shows the cadence — so a GPU's economic life is two to three years, not the four to six being booked. On that view, depreciation is understated, earnings are overstated, and the cumulative overstatement across the majors runs to more than $176B over 2026–2028. The bull case (Nvidia, and the hyperscalers' own disclosures): customers set lives from observed utilisation, and old accelerators do not become worthless — they cascade down to inference, batch and internal workloads. An A100 from 2020 is still earning. Six years reflects what the fleet actually does, not what the marketing cycle says. What would actually settle it, and is worth watching: not opinion but more accelerated-depreciation charges. Amazon's $920M write-down in 2025 was a company telling the market its own prior estimate had been wrong. One firm shortening is a judgement call; three firms shortening inside a year is the bear case being conceded in the filings, and it would arrive as a sudden step-up in reported cost with no change in cash. That is the specific, falsifiable thing to watch for — and it is why this section exists.
Why an accounting footnote belongs on a memory dashboard
Because it is the mechanism by which the memory cycle could end without demand falling at all. The chain this board tracks runs: AI revenue (§08·D) → hyperscaler capex (§06·F·D) → accelerator orders (§06·C) → HBM and DRAM (§05·C·G, §05·D) → prices (§03). Depreciation sits between the first and second link. If useful lives shorten, reported profit falls without any change in cash flow or demand — and the pressure lands squarely on the capex budget that funds everything downstream. That is kill-switch #3 (§02·D) arriving through the income statement rather than through the market. And it interacts with the debt. §06·F·E tracks bond coupons that have been rising; §06·F·B tracks $2.47T of backlog that has to be delivered. A firm that shortens useful life reports lower earnings against the same interest bill and the same delivery obligation — which is exactly the configuration in which a capex plan gets revisited. Microsoft already cut its calendar-2026 plan from ~$190B to ~$175B (§06·F·D) while everyone else raised. This section is where the next such decision would first become visible.

Provenance & the honesty line. [cited]: the four US hyperscalers purchased $433.9B of property and equipment in the four quarters through March 2026 against roughly $149B of reported depreciation; quarterly combined capex reached $129.8B in 1Q26, +80% YoY; 2026 plans total about $725B (Amazon ~$200B, Alphabet $175–185B, Meta $115–135B, Microsoft $120B+), roughly +77% on 2025. Useful-life history: Amazon, Alphabet and Microsoft all extended servers from a 3–4 year standard to 6 years across 2020–2023, collectively reducing annual depreciation by roughly $18B. The 2025 divergence: Amazon changed a subset of servers and networking equipment from six years to five, effective Jan 1 2025, explicitly citing the pace of AI development, and recognised a $920M accelerated-depreciation charge; Meta extended its estimate further in the same period; Microsoft and Alphabet remain at six, with Satya Nadella publicly saying he did not want to "get stuck with four or five years of depreciation on one generation." The critique: Michael Burry's estimate that shortening to a 2–3 year replacement cycle would cut reported earnings by more than $176B cumulatively over 2026–2028. The rebuttal: Nvidia's position that customers consistently use 4–6 year lives based on observed utilisation and longevity. What is modeled, and it is only one thing: the five bars are straight-line arithmetic on the cited $433.9B — purchases ÷ assumed life — and nothing more. They deliberately do not model reality, which includes assets not yet in service, salvage value, accelerated methods, the ~25–40 year schedules on buildings mixed into the same capex figure, and the fact that servers are only part of PP&E. Read the bars as "what one year of buying costs per year at each divisor", not as a forecast of any company's depreciation line. The teal reference line sits below all five for exactly that reason and is the honest anchor. Three further caveats: (1) the $149B is total depreciation, not server depreciation — it includes buildings and older fleet, so the true server-only lag is larger than the chart implies; (2) useful life is an estimate, not a policy — auditors sign off on it and it is revisited annually, so it can move without anyone acting in bad faith; (3) none of this is an allegation, and the section deliberately prints Nvidia's rebuttal at the same weight as the critique. Cross-refs: §06·F·D (the capex), §06·F·E (the debt), §06·F·B (the backlog), §06·D·M (the model cadence that drives the obsolescence argument), §02·D (kill-switch #3). Sources: company 10-K/10-Q filings and earnings calls, Silicon Analysts, Yardeni, Deep Quarry, Level-Headed Investing, CNBC. Not investment advice, and explicitly not an accounting opinion. Aug 17 2026 — the first live evidence on this question, and it points the other way [§06·D·C]. The A100 is the only generation old enough to test the argument above. Released May 2020, a six-year life has it fully depreciated as of May 2026. It is now past that date, renting at $1.76/hr — up 3.5% year-over-year — at near-full utilisation, and CoreWeave has contracted A100 capacity into 2029. On that evidence the four hyperscalers are over-depreciating, not under, and the Burry case is inverted rather than merely wrong. Three cautions carried in §06·D·C: the contract price is undisclosed, a rental rate is not a residual value, and A100 demand is specifically inference demand.

06·F·PPower Procurement — How the Committed Gigawatts Actually Arrive◷ queue + turbine + PPA data · 2026
06·F·P

Power Procurement

§06·F·C counts the gigawatts the labs have announced. §06·F·B's second chart infers ~51 GW from the contracted backlog. Neither asks the question this section is about: through which physical route does a gigawatt actually arrive, and in what year? There are only three routes, and all three are congested. The US interconnection queue holds 2,600 GW with a median five-year wait and data-centre requests waiting up to twelve; ERCOT alone is sitting on a 410 GW large-load queue that is 87% data centres. Large gas turbines are sold out — three firms make over 75% of them and an order placed today arrives 2028–2030. Nuclear is real and too late: 9.8 GW committed across thirteen hyperscaler PPAs, the biggest private nuclear procurement since the 1970s, almost none of it delivering before the 2030s. The contracts in §06·F·B assume delivery starting around 2027. The power does not.

The three routes to a gigawatt, and when each can actually deliver
Each bar is the window in which capacity ordered today can realistically energise, against the dashed line marking roughly when the AI contracts on this board assume delivery begins. Every route starts to the right of that line. The windows are drawn from cited lead times, not modeled: the grid figure is the queue's own median-to-commercial-operation, the gas window is what the three turbine makers are quoting, and the nuclear window is new-build reality rather than PPA signature date. A PPA signed in 2026 is a claim on electrons that mostly do not exist yet — which is why this section sits next to the backlog rather than inside it.
The demand side, from this board's own numbers
Three different claims about how much power AI has contracted for, at three very different scales — printed together because the spread between them is itself informative, and because the largest is a single grid operator's queue rather than anyone's commitment.
The 80% that never arrives, and why queue size is not supply
A 2,600 GW queue is not 2,600 GW of coming supply — roughly four fifths of it withdraws. Projects leave because the wait is unpredictable and because the grid-upgrade costs allocated to them arrive late and large. The same is true of ERCOT's 410 GW: it is a register of requests, many speculative, some the same project queued in several places. Read both numbers as evidence of congestion, not as an inventory. Which is what makes the gas-turbine number the hardest of the three. A queue position can be abandoned; a turbine order is a physical object with a factory slot. GE Vernova's 83 GW backlog stretching into 2029 is the most concrete constraint on this page, because it is the one where the supplier, the quantity and the delivery year are all disclosed and none of them can be wished forward. If you want one number for when the AI build-out is actually rate-limited, it is that one.

Provenance & the honesty line. [cited]: US interconnection queue backlog ~2,600 GW, median time to commercial operation approaching five years with data-centre requests facing up to twelve, and roughly 80% of queued projects withdrawing, mainly over delay and grid-upgrade cost; ERCOT managing a 410 GW large-load queue of which data centres are 87%. Turbines: GE Vernova, Siemens Energy and Mitsubishi Power supply >75% of large gas turbines and all three have backlogs pushing new availability to late 2028 at the earliest; Mitsubishi quotes 2028–2030 for orders placed now; GE Vernova's backlog is 83 GW with deliveries into 2029; lead times generally 3–5 years with premiums paid to jump. Nuclear: every major hyperscaler has signed at least one nuclear PPA by mid-2026, with ~9.8 GW committed across 13 disclosed projects — the largest private-sector nuclear procurement wave since the 1970s; Meta alone has 6.6 GW of firm-power agreements; NuScale is reported to be near a PPA tied to a 6 GW SMR programme with TVA. What is modeled: only the delivery windows drawn on the chart. The bars translate cited lead times into calendar windows and extend past 2032 where the route has no credible end date; they are a reading of the cited figures, not a schedule anyone published. The "contracts assume delivery from ~2027" line is this board's inference from §06·F·B's recognition profiles (CoreWeave's own disclosure puts ~41% of RPO inside 24 months) and should be read as an approximation of a range, not a date. Three caveats: (1) queues are requests, not projects — the 2,600 GW and 410 GW figures overstate real supply by a large and unknown factor, which the block above says plainly; (2) behind-the-meter generation is a real escape hatch this chart does not model — on-site gas, fuel cells and co-located generation bypass the interconnection queue entirely and are how several announced campuses intend to energise, which would move the grid bar left for those specific sites; (3) PPA signature ≠ new capacity — several hyperscaler nuclear agreements contract for output from existing plants or restarts, which delivers far sooner than new build but adds no net generation to the grid. Cross-refs: §06·F·C (GW committed), §06·F·B (the backlog and its implied megawatts), §06·F·L (whether there are enough electricians to install any of it), §05·F (the silicon those gigawatts power), §02·D (kill-switch #3). Sources: RMI, EnkiAI, interconnection.fyi, OilPrice, Forbes, Perkins Coie, NuScale filings, company announcements. Not investment advice. Aug 17 2026 — what a megawatt earns, now that this section covers what one costs to obtain [§06·F·W]. Neoclouds monetise a live megawatt at $9.4–10.4M a year, colocation at $3.5–4.4M, and new neocloud contracts are being written at $40–50M of total contract value per MW. Set against the delivery windows above, that is the number deciding how hard anyone will fight for a queue position. Aug 17 2026 — where that capacity physically sits is now mapped [§06·F·T]. US colocation inventory reached 29.0 GW in Q1 2026 at 1.2% vacancy, with Northern Virginia alone about 22% of the country. Siting friction is tracked there as a delivery risk: 225+ local moratoriums across 30 states, New York the first statewide pause, and roughly $8B blocked / $6B delayed since mid-2024.

06·F·TUS Data-Centre Capacity Map — Where the Megawatts Actually Are◷ Q1 2026 market inventory · Aug 2026
06·F·T

US Data-Centre Capacity Map

§06·F·P asks how a gigawatt gets delivered and §06·F·W asks what one earns. This asks where they physically are. US colocation inventory reached 29.0 GW in Q1 2026 — up 22% in a single quarter and 48% year-over-year — at a national vacancy rate of 1.2%. The concentration is extreme: Northern Virginia alone is roughly 22% of the entire country, and four markets are 55% of it. One scope caveat up front, because the request was for "all data centers": no public source publishes megawatts per site for every US facility. What exists is market-level inventory, and even that is only verifiable for four markets. Everywhere else is drawn as a hollow ring — present, unsized, and deliberately not estimated, because a bubble map invites the eye to compare areas and inventing an area would be inventing a fact.

Colocation inventory by market — bubble area proportional to megawatts
Radius scales as the square root of megawatts, so area is the honest visual quantity. The dashed red ring on Northern Virginia is the same market as sized by a second major brokerage — 4,039.6 MW against 6,485 MW, a 61% disagreement about the largest data-centre market on earth. Both are recent, both are reputable, and this board cannot tell you which is right. Hollow dashed rings are markets known to be significant with no published total; they are placed but not sized. Hover any market.
The four markets with a verifiable megawatt total
Together they are just over half the national figure. The remaining 45% is real capacity in markets this board could not size individually — it is shown in the table as an explicit residual rather than distributed across the map.
Two brokerages, one market, 61% apart
CBRE puts Northern Virginia at 4,039.6 MW of total inventory in Q1 2026. Avison Young puts it at 6,485 MW. That is a 2,445 MW gap — larger than the entire Austin market — in the most-studied data-centre market in the world, between two firms that both do this professionally. The likeliest explanation is scope: commissioned versus total-built, colocation-only versus including hyperscaler self-build, or utility-delivered versus IT load. None of the published summaries states its definition precisely enough to test that, which is why the map draws both and labels the difference rather than picking one. It matters for a bubble map specifically. Because radius scales as the square root of area, a 61% disagreement in megawatts is a 27% disagreement in drawn radius — visible, but far less visible than the underlying uncertainty. A reader eyeballing this map would under-estimate how unsure the largest bubble is. Hence the dashed ring.
Siting friction — a real constraint, reported without the politics
Where capacity can be built is now a live constraint on the build-out, so it belongs on this board as a risk factor. The documented position: at least 225 local moratoriums tracked across 30 states; New York became the first state to impose a statewide pause, by executive order in July 2026; and around $8B of projects have been blocked with a further $6B delayed since mid-2024. Statutory thresholds vary enormously — 100MW in Oklahoma, Minnesota, Vermont and Wisconsin; 50MW in South Dakota; 5MW in South Carolina; 1MW in Virginia. Two source conflicts: reports give New York's threshold as both 50MW and 20MW, and the state-level count as both "14 states considering or implementing" and "30 states with local moratoriums" — different objects being counted, neither reconciled. The one analytical observation worth making is mechanical rather than political: the states with the most local moratoriums — Ohio (35), Michigan (34), Georgia (24), North Carolina (19) — are among the states where construction is most actively proposed. Opposition concentrates where building concentrates, because you can only object to what is planned near you. That makes moratorium counts a poor proxy for regional sentiment and a decent proxy for where development pressure is highest. On the source that prompted this section: it is a political argument about which constituencies benefit from the build-out and contains no capacity data — no site list, no megawatt figures. It was not used as a data source and this board does not adopt or contest its framing. Siting is tracked here because permitting risk affects delivery schedules (§06·F·P), which is a question this board can actually test.

Provenance & the honesty line. [cited]: US colocation inventory 29.0 GW in Q1 2026, +22% QoQ and +48% YoY, with national vacancy at 1.2% and Q1 net absorption of 5.29 GW; market inventories of Northern Virginia 6,485 MW, Dallas–Fort Worth 3,704 MW, Atlanta 3,257 MW and Austin 2,558 MW; vacancy of 0.14% (Austin), 0.18% (Salt Lake City), 0.3% (Northern Virginia) and 1.0% (Atlanta); Q1 absorption of 1,909 MW (Austin), 912 MW (DFW), 538 MW (Chicago) and 480 MW (Northern Virginia); Atlanta with 2,076 MW under construction; Chicago displacing Phoenix for fourth place; the top-seven wholesale markets being Northern Virginia, Dallas–Fort Worth, Silicon Valley, Chicago, Phoenix, New York Tri-State and Atlanta; Reno and Abilene as tertiary markets whose pipelines may outpace some primaries. Separately cited: CBRE's 4,039.6 MW for Northern Virginia and primary-market supply of 9,432 MW (+36%) — a different scope from the 29.0 GW figure and not netted against it. Moratorium figures as set out above. What this map is not: (1) it is not "all data centers" — it is four sized markets and nine placed-but-unsized ones, covering 55% of national inventory by megawatt, with the other 45% shown as an explicit residual in the table. No public source lists per-site megawatts nationally, and this board will not synthesise one. (2) It is not a site map. Bubbles are metro markets placed at a representative city coordinate, not facilities; Northern Virginia's bubble covers a corridor tens of miles across. (3) The coastline is a simplified 50-point outline drawn for orientation only — it is not survey-accurate and no conclusion should rest on a bubble's exact position. (4) Inventory is not the same as AI capacity — colocation inventory includes long-standing enterprise and network workloads, so this map shows where data-centre capacity is, not where accelerators are. [derived]: all share-of-US percentages, the four-market total of 16,004 MW, the residual, and the observation that a 61% MW disagreement renders as a 27% radius difference. Cross-refs: §06·F·P (whether the power reaches these places), §06·F·W (what a megawatt earns), §06·F·S (the operators building them), §06·F·B (the contracts behind the construction), §06·F·L (the labour to build them). Not investment advice.

06·F·RNeocloud Financing — The GPU-Backed Debt Stack◿ filings + ratings · 2026
06·F·R

Neocloud Financing

§06·F·E tracks hyperscalers borrowing against balance sheets that already print tens of billions in cash flow. This section is about the other borrowers — the ones whose collateral is the GPUs. More than $20 billion of loans are now secured on Nvidia hardware, and the lenders are not venture funds but BlackRock, Blackstone, PIMCO, Carlyle, JPMorgan and Macquarie. The number this section is actually built around is smaller and stranger: CoreWeave closed one facility in March 2026 rated A3 — investment grade — and another in May rated Ba2, five notches lower and firmly high yield. Same borrower. Seven weeks apart. Which means neither rating is really a judgement about CoreWeave. The market is pricing the offtake contract behind each facility, so a neocloud's cost of capital is set less by its own creditworthiness than by whose name is on the customer side of the lease — and that is a materially different risk than a spread alone would suggest.

Two facilities, one borrower, seven weeks apart
Moody's scale, with the investment-grade boundary drawn. A five-notch spread on the same obligor inside a single quarter is unusual enough to be the finding rather than the background. The March facility was reported as the first HPC-infrastructure delayed-draw term loan to reach investment grade; the May facility was the first publicly syndicated one. The difference between them is what secures them — contracted revenue from investment-grade counterparties on one, and considerably less on the other. Read this as a chart about structured credit, not about a company. Hover either point for the detail.
What is actually disclosed, and what is not
These rows must not be added together, and they must not be compared against the >$20B headline. CoreWeave's figure is total debt; the market-wide figure counts only loans secured on GPUs. They are different measures, and the disclosed rows here in fact sum to more than the >$20B market number — which is a sign the two are not commensurable, not a sign that either is wrong. The "Others" row is left blank rather than solved for, because subtracting an approximate disclosed sum from an approximate and differently-defined total would produce a number carrying both errors and no information.
The collateral is the thing nobody can agree how to value
These loans are secured on GPUs. §06·F·N shows that the four largest owners of GPUs cannot agree how long a GPU lasts — Microsoft and Alphabet book six years, Meta extended, and Amazon went the other way, shortening to five and taking a charge for it. A lender writing a five-year loan against an asset whose owners' own depreciation schedules span five to six years has almost no margin between the loan tenor and the collateral's booked life, and the schedules themselves are estimates that have already been revised in both directions. This does not make the loans bad. The investment-grade facility is secured on contracted revenue rather than on hardware resale value, which is precisely the right structure for an asset with uncertain residuals. But it does locate the risk: the question is not whether the GPUs work, it is whether the counterparties keep paying for them for as long as the paper assumes. That question is answered in §06·F·B, not here — the backlog is the collateral behind the collateral.

Provenance & the honesty line. [cited]: more than $20B of loans outstanding collateralised by Nvidia GPUs across the neocloud sector; CoreWeave carrying above $21B of total debt in early 2026 at roughly 11% average interest; Crusoe with $425M of GPU-backed debt; Lambda having securitised $500M in what is described as the first GPU-backed asset-backed security. [cited — 8-K disclosed]: CoreWeave closed an $8.5B delayed-draw term loan on March 31 2026, reported as the first HPC-infrastructure DDTL to achieve investment-grade status at A3 / A (low); and a $3.1B delayed-draw term loan on May 18 2026, reported as the first publicly syndicated HPC-backed DDTL, rated Ba2 (Moody's) / BB+ (Fitch). Lenders named across the sector include BlackRock, Blackstone, PIMCO, Carlyle, JPMorgan and Macquarie. What is inference, not fact: the explanation that the five-notch gap reflects collateral structure rather than borrower credit is this board's reading. It is well supported — the facilities differ in security and syndication, and both ratings were published within one quarter on one obligor — but no rating agency has published a side-by-side rationale, and the agencies differ (Moody's rated one, Fitch and DBRS the others), so part of the apparent spread may be agency methodology rather than structure. Treat "5 notches" as exact and the explanation for it as argued. What is deliberately absent: (1) any residual estimate for undisclosed borrowers — the >$20B market figure minus the three disclosed names is not printed. Beyond the usual objection that both inputs are approximate, they are not the same measure: CoreWeave's $21B is total debt while the market figure counts only GPU-collateralised lending, so the disclosed rows sum above the market-wide total. That subtraction is not merely imprecise, it is invalid, and this section says so rather than quietly omitting the row; (2) any default-probability or loss-given-default estimate, which would require loan-level terms nobody has published; (3) Nvidia's own backstop arrangements, widely discussed as "circular financing", are named here but not quantified, because the disclosed terms are too fragmentary to size honestly. One thing to watch: if a neocloud refinances at a materially wider spread than these two prints, that is the first hard market signal on this board that the collateral is being repriced. Cross-refs: §06·F·E (hyperscaler debt, for the cost-of-capital contrast), §06·F·N (the depreciation assumption that values the collateral), §06·F·B (CoreWeave and Nebius RPO — the contracts behind the paper), §06·F·D (capex), §02·D (kill-switches). Sources: SEC 8-K filings, Moody's, Fitch, DBRS, company announcements, sector analyses. Not investment advice. Aug 17 2026 — the growth side of the same companies is now at §06·F·S, and it supplies an unexpected check on the numbers above. CoreWeave's Q2 residual below the operating line is $577M; the >$21B at ~11% cited here implies $578M of interest a quarter. A 0.1% match — which cannot all be interest, since the residual also carries tax, but is close enough to suggest the debt and rate on this page are about right.

06·F·SNeocloud Growth — The Fastest-Scaling Businesses Here, and the Least Profitable◷ Q2 2026 results · Aug 17 2026
06·F·S

Neocloud Growth

§06·F·R covers how the neoclouds are financed. This one covers how fast they are growing — and the two turn out to be the same story told from opposite ends. Q2 2026: Nebius +454%, CoreWeave +112% to $2.58B, Cerebras +74%. On a16z's reading it took CoreWeave about 25 quarters to reach the $2.6B AWS did not hit until its 40th. All three lost money doing it. The cleanest illustration is CoreWeave's own quarter: a 59% adjusted-EBITDA margin and a $626M net loss are the same three months. Everything in the $2.14B between those numbers is depreciation on the GPUs (§06·F·N) and interest on the debt that bought them (§06·F·R). The origin story matters too: most of these firms were crypto miners, and what AI repriced was not their chips but their power rights — the same pattern as railroad corridors becoming Sprint's fibre network.

CoreWeave Q2 2026 — from a 59% margin to a $626M loss in two steps
Solid bars are reported levels; dashed bars are the two deductions between them. This chart exists because "adjusted EBITDA" and "net income" describe the same quarter and disagree by $2.14 billion, and the disagreement is not noise — it is precisely the two things this board tracks in §06·F·N and §06·F·R. Hover the deductions for what sits inside each.
The three listed neoclouds, Q2 2026
Ordered by size. The pattern holds across all three regardless of scale or business model — and note that Cerebras, which sells its own wafer-scale silicon rather than renting Nvidia GPUs, has the worst net margin of the three by an order of magnitude.
Two cross-checks that close, and one that should not be over-read
(1) Dollars per megawatt, from two companies that did not coordinate. Nebius disclosed closing four billion-dollar-plus deals at a yield of $20–25M per megawatt. CoreWeave disclosed a $104B backlog and 4.2GW of contracted power — which divides to $24.8M per megawatt. Two independent disclosures, essentially the same number, and it lands inside the range §06·F·B uses to convert hyperscaler RPO into gigawatts. That conversion was the most fragile assumption in that section; it now has a direct anchor. (2) Implied contract duration. CoreWeave's annualised Q2 revenue over its 1.5GW of live capacity is about $6.9M per MW per year. Against $24.8M of backlog per MW, that implies a ~3.6-year average contract. §06·F·B records CoreWeave disclosing roughly 41% of RPO inside 24 months, which is back-weighted relative to a flat 3.6-year profile — so treat 3.6 years as a floor, not an estimate. (3) The one to be careful with. CoreWeave's residual below the operating line is $577M. §06·F·R's cited figures — over $21B of debt at roughly 11% — imply $578M of interest a quarter. That is a 0.1% match and it is genuinely striking. But the residual also contains tax and other items, so it cannot all be interest. The agreement is close enough to suggest §06·F·R's debt and rate are about right, and loose enough that it is not proof of anything. It is recorded as a coincidence worth watching, not as a reconciliation.
Where the sources disagree — including one publisher with itself
(a) a16z's post and a16z's article make different claims. The promotional post states that at 25 quarters CoreWeave out-earns what Azure, AWS or Google Cloud were making at 30. The article it links to states something else: that it took CoreWeave ~25 quarters to reach the $2.6B AWS reached by quarter 40. Quarter 30 against quarter 40, one comparison against three companies. This board uses the article's version, because it is the one with a named figure attached, and flags that the more-shared version is the stronger claim. The article also concedes the early hyperscaler quarters can "only be estimated". (b) Cerebras states two different revenues in one release. Its CEO describes "core revenue more than doubled to $210 million"; the same announcement reports GAAP total revenue of $180.1M, up 74%. The table above uses the GAAP figure and the growth rate that goes with it. A non-GAAP "core revenue" that exceeds GAAP total revenue is unusual enough to name. (c) Growth has not been rewarded. Despite the quarter, a16z notes CoreWeave is down about 16% over the past year and that only Nebius is near its previous high — the market is not simply extrapolating the revenue line, which is consistent with §06·F·R's finding that the paper is priced off the offtake rather than the borrower.

Provenance & the honesty line. [cited — Q2 2026 results, reported Aug 12–14 2026]: CoreWeave revenue $2.575B (from $1.212B), operating loss $49M (from $19M income), adjusted operating income $128M, net loss $626M (from $290M), adjusted EBITDA $1.51B (from $752M); Q2 capex $9.4B with FY26 guided $35–39B; active power up ~500MW in the quarter (300MW in June) to 1.5GW across 51 data centres, eight added in 2026, year-end target 1.85GW, contracted power 4.2GW and a 8GW by 2030 target; revenue backlog ~$104B at Jun 30 plus >$25B of net new commitments in early Q3; managed inference ARR $1M → >$100M, guided to ≥$250M exiting 2026. Nebius revenue $582.3M (+454%), AI-cloud revenue $575M (+514%), adjusted EBITDA +$236.2M (from −$21M), adjusted net loss $33.2M (from $91.5M), net income −$190.4M (prior-year +$584.4M was largely equity-revaluation gains), operating costs $216.3M → $758.2M; four deals above $1B each at $20–25M per MW; +1GW contracted in the quarter, 5GW targeted by end-2026, >1GW of compute deployed per year from 2027; production inference workloads more than tripled. Cerebras GAAP revenue $180.1M (+74%), cloud and other services $126.0M (+281%), gross profit $25.56M (down from $32.1M), net loss $450.4M (from +$309.5M), adjusted EBITDA −$53.1M; >600MW live and contracted for delivery by end-2027. [derived — this board's arithmetic]: the CoreWeave bridge (+$1,510M → −$1,559M → −$49M → −$577M → −$626M) is subtraction from the four reported figures, not a disclosed breakdown, so the $1,559M and $577M steps are residuals and each bundles several line items. Net margins (−24.3% / −32.7% / −250.1%), Cerebras' 14.2% gross margin, CoreWeave's 58.6% adjusted-EBITDA margin, $24.8M per MW, $6.9M per MW-year and the ~3.6-year implied duration are all computed here. Nebius' and Cerebras' prior-year revenues are back-solved from the disclosed growth rates rather than read from a filing. What is not attempted: (1) no forward projection — one quarter of triple-digit growth does not extrapolate, and this section deliberately has no 2027 or 2030 line; (2) no valuation view, though a16z's own note that sales multiples are "not the best choice for capital-intensive businesses" is worth carrying; (3) no attempt to net the three companies into a sector aggregate, since Cerebras sells its own silicon while the other two rent Nvidia's, and the models are not comparable. Source-chain note: the quarter-count comparison originates in an a16z newsletter chart whose underlying early-hyperscaler figures its author calls estimates; it is reproduced as commentary, not as a measurement. Cross-refs: §06·F·R (how the growth is financed), §06·F·N (the depreciation inside the bridge), §06·F·B (the RPO and megawatt conversion this validates), §06·F·P (whether the contracted gigawatts can actually be energised), §06·G (what an accelerator-hour rents for), §05·F (the silicon underneath). Not investment advice. Aug 17 2026 — the per-megawatt view is now its own section [§06·F·W], and it corrects a unit error in the widely-shared version of this comparison. Neoclouds earn $9.4–10.4M per active MW-year against colocation's $3.5–4.4M — that ratio holds. But SpaceX's much-quoted $30–50M is contract value, not annual revenue; divided by a three-to-four-year term it becomes $7.5–16.7M per MW-year, inside the neocloud band. On the contract basis where that number actually lives, Nebius' newest deals at $40–50M/MW already match or exceed it.

06·F·WRevenue per Megawatt — Neoclouds, Colocation, Hyperscalers◷ Q2 2026 disclosures · Aug 2026
06·F·W

Revenue per Megawatt

Power is the binding constraint (§06·F·P), so revenue per megawatt is the natural efficiency metric for everyone building into it. A widely-shared comparison puts the neoclouds at roughly $9–10M per active MW-year against colocation at $3.5–4.4M, and then adds SpaceX at $30–50M to argue the neoclouds have a long way still to climb. The first comparison is sound. The second is a unit error. The neocloud and colo figures are annual revenue over live capacity; SpaceX's is total contract value over contracted capacity. Divide it by a three-to-four-year term and it becomes $7.5–16.7M per MW-year — inside the neocloud band, not three to five times above it. On the contract basis where SpaceX's number actually lives, Nebius's newest deals at $40–50M/MW already match or exceed it. This section plots the two units on separate axes so they cannot be read across.

Two metrics that look alike and are not — plotted apart on purpose
Panel A is a flow: revenue earned in a year per megawatt actually running. Panel B is a stock: the whole value of a contract per megawatt committed, over a term nobody discloses. The red rule between them is not decoration — the axes are different and the rows must not be compared across it. The dashed bar on CoreWeave's row is the same company on the same quarter under three defensible denominator conventions. Hover any row.
Why the hyperscalers are not on this chart
They belong in the question and cannot be put on the axis, and the reason is itself the finding. Microsoft, Amazon, Google and Meta together run about 13GW of live IT capacity in North America — roughly 42% of the region's total, with another ~7GW building. But dividing cloud revenue by that number produces a figure that means nothing, for two reasons that push in opposite directions. First, the denominator is wrong: that capacity serves search, ads, video and internal workloads as well as cloud, and none of the four discloses the split. Second, the numerator is wrong: hyperscaler cloud revenue is substantially software, databases, networking and managed services — revenue that consumes almost no megawatts at all. A worked illustration of how wide it swings. Google Cloud ran at $24.8B in Q2 2026, roughly $99B annualised. If Alphabet's share of that 13GW were 3GW and all of it served Cloud, that is $33M per MW-year. If only half served Cloud, $66M. If a third, $110M. Every one of those is above SpaceX's contract figure and none of them is comparable to anything else on this page. The structural point survives the arithmetic: neoclouds sell watts, colocation sells space and cooling around watts, and hyperscalers sell software on top of watts. Revenue per megawatt ranks them by how far up the stack they sell, not by how efficiently they use power — which is worth knowing before treating a high number as an operational achievement.
The Nebius escalation, and the one claim worth watching
Nebius' contract value per megawatt has moved through three tiers inside one year: about $12M on its 2026 base contracts, $20–25M on Q2 deals, and $40–50M on short-term capacity signed heading into Q3 — more than tripling. Its contracted-power guidance has been raised at every quarterly report for a year: >1GW (Aug 2025) → >2.5GW (Nov) → >3GW (Feb) → >4GW (May) → 5GW now, alongside ARR of $3.0B at end-June, up 58% quarter-on-quarter. The line worth flagging is management's: that Nebius "could sell our entire 2027 capacity on these terms today" but is "deliberately not doing so", holding capacity back for higher-value immediate demand. That is an unfalsifiable statement about a counterfactual sale — it may well be true, and there is no way to test it. This board records it as management commentary, not as evidence of pricing power, and notes that the checkable version arrives when the 2027 book is eventually signed at a disclosed rate. One structural caution on the whole metric: Nebius earns its $9.4M per active MW on roughly 248MW while carrying 5,000MW under contract — a 20:1 ratio. Revenue per active megawatt systematically flatters any operator whose contracted book runs far ahead of what it has energised, because the denominator is the small number.

Provenance & the honesty line. [cited]: annualised revenue per active megawatt of IREN $10.4M, CoreWeave $9.8M, Nebius $9.4M, Equinix $4.4M, Coresite $3.6M, Digital Realty $3.5M; SpaceX quoted at $30–50M per MW on a contract basis. Nebius contract value per MW stepping from ~$12M (2026 base) to >$20M (Q2) to $40–50M (new, signed the week before the Q2 report); ARR $3.0B at end-June, +58% QoQ; four agreements above $1B TCV each; total contract value won the prior quarter; contracted-power guidance raised at five consecutive reports to 5GW. Hyperscalers: Microsoft, Amazon, Google and Meta controlling about 13GW of live IT capacity in North America (~42% of the region) with ~7GW under construction; Google Cloud revenue $24.8B in Q2 2026. [derived — and this is where the popular version goes wrong]: the neocloud-to-colo ratio of 2.6× is a like-for-like comparison and holds. The SpaceX comparison does not: $30–50M of contract value ÷ a 3–4 year term = $7.5–16.7M per MW-year, which overlaps the neocloud band rather than sitting 3–5× above it. The term is this board's assumption, taken from §06·F·S's implied ~3.6 years, because SpaceX's contract duration is not disclosed anywhere — if the true term were one year the original comparison would stand, and if it were five years the neoclouds would already be ahead. The conclusion is therefore about the unit mismatch, not about who is winning. CoreWeave's $6.87M / $9.80M / $10.30M range is computed from its own disclosures ($2.575B quarterly revenue annualised, against 1.0GW opening, ~1.05GW average and 1.5GW closing capacity); Nebius' ~248MW active and the 20:1 ratio are implied from $9.4M and $582.3M. What is not attempted: (1) no hyperscaler figure, for the two reasons set out above — the Google illustration is deliberately shown as a range across assumptions rather than a number; (2) no ranking across panels A and B, which is the entire point of separating them; (3) no forecast of where per-MW pricing settles. Source note: the per-MW figures come from an investor's public analysis rather than company filings, and the underlying active-megawatt denominators are not published by the companies — treat the levels as indicative and the ordering within panel A as the robust part. Cross-refs: §06·F·S (the Q2 results underneath these ratios), §06·F·B (the RPO-to-megawatt conversion this anchors), §06·F·P (why megawatts are the binding constraint), §06·F·R (the debt secured on them), §06·D·C (what a GPU-hour rents for), §06·G (rental pricing). Not investment advice. Aug 17 2026 — §06·F·T maps where these megawatts are, and the concentration matters for the ratios here: four markets hold 55% of US colocation inventory, at a national vacancy rate of 1.2%. Revenue per MW is being earned in a market with almost no slack.

06·F·OOptical Interconnect & Co-Packaged Optics — The Layer Between the GPUs◿ vendor + tracker data · 2026
06·F·O

Optical Interconnect & Co-Packaged Optics

A GPU cluster is not a pile of GPUs; it is a network with accelerators attached, and past a few thousand nodes the optics become the constraint rather than the plumbing. This section tracks that layer — and leads with a disagreement rather than a number. Two published forecasts for how many 1.6T optical modules ship in 2026 differ by roughly six times, one saying above 30 million and the other above 5 million, and they do not even agree on what 2025 was. Neither is averaged here. Both are drawn, both are attributed, and the gap between them is the honest state of knowledge about the fastest-growing component category in the build-out. Underneath that, the architectural shift is real and dated: co-packaged optics moved from conference slide to shipping product, with Nvidia's Quantum-X Photonics introduced in early 2026 and Broadcom's third-generation CPO announced the previous October, both at 200 Gb/s per lane, both keeping the lasers outside the package because lasers are still the part that fails.

1.6T optical module shipments — two forecasts, drawn side by side
The shaded wedge is not an uncertainty band this board calculated. It is the space between two independently published forecasts, one from supply-chain channel checks and one from module-market tracking, for the same product in the same year. They disagree by 6× in 2026 and by 50% in 2025 — a year that has already happened. When a component's near-term volume is this contested, forecasts built on top of it inherit the spread, which is why §06·F·B's backlog and §05·F's wafer model both stop short of pricing the network. Hover any year to see both numbers as published.
The co-packaged optics platforms that actually shipped
CPO moves the optical engine from a pluggable module at the faceplate onto the switch package itself, cutting the electrical distance the signal has to travel. That is where the power saving comes from — and where the serviceability problem comes from, since a failed optic is no longer a part you swap in the field.
Why the lasers stayed outside the package
The obvious version of co-packaged optics puts everything on the package, lasers included. Nobody shipped that. Both Nvidia and Broadcom kept the light sources in separate pluggable external-laser modules, and the reason is failure statistics: the laser is historically the least reliable element in an optical link, and a laser soldered next to the switch ASIC turns a field-replaceable part into a dead switch. Keeping them pluggable also allows a surviving laser's output to be turned up to cover a neighbour that has degraded. This is worth noting precisely because it cuts against the marketing. The power and density numbers quoted for CPO are real, but the architecture that reached customers is a compromise — the integration stopped exactly where the reliability data said it should. Early field results are reported as reassuring rather than proven; CPO's installed base is months old and its failure curve is not yet observable. Anyone modelling a 3.5× power saving into a total-cost-of-ownership case is extrapolating from a vendor figure and a short service history.

Provenance & the honesty line. [cited — and mutually contradictory]: one supply-chain source puts 1.6T module demand at 1.8M units in 2025 rising above 30M in 2026, with Nvidia over 60% of it and Google and Meta the remainder; a separate module-market forecast puts the same category at ~2.7M in 2025 and above 5M in 2026. This board cannot reconcile these and does not try — the likeliest explanations are different definitions of "1.6T" (module count versus 800G-equivalent lanes, or demand booked versus units shipped), but no source states its definition clearly enough to test that, so both are plotted as published. Market size, also from more than one basis: a data-centre optical module market of $22.8B in 2026 of which $14.6B is 800G and 1.6T, alongside an Ethernet-optics figure of $26B in 2026 against $16.5B in 2025 — overlapping but not identical scopes, printed rather than merged. Growth rates cited as +93% (2024), +82% (2025), +65% (2026f), with demand running roughly 30% above supply and shortage relief expected by end-2026. A second, smaller inconsistency, found while building this section and left visible rather than tidied: the same tracker's own totals of $16.5B → $26B imply +58%, not the +65% it forecasts elsewhere. The likeliest cause is that the totals and the growth rate come from different vintages of one forecast, but nothing published says so — so the stat card above prints the arithmetic the totals actually support and names the gap rather than picking a side. Hardware [cited, vendor-disclosed]: Nvidia Quantum-X Photonics introduced early 2026, 115 Tb/s per switch across 144 ports at 800 Gb/s, package built by TSMC combining 18 silicon photonics engines, 324 optical connections and 288 data links from 36 laser inputs, on the COUPE process that 3D-hybrid-bonds the electronic die onto the photonic die; Quantum X800-Q3450 CPO using 72 optical engines at 1.6 Tb/s; Spectrum-X Photonics for Ethernet; Broadcom's third-generation CPO announced October 2025 at the same 200 Gb/s per lane, Ethernet only. What is a vendor claim, not a measurement: the 3.5× power reduction versus pluggable transceivers is Nvidia's own figure at its own comparison point and has not been independently benchmarked on this board. Treat it as a directional claim. What this section does not contain: any estimate of optics as a share of cluster capex, because the unit forecasts disagree by 6× and multiplying an unstable volume by an unstable ASP produces a number with no information in it. One structural caution: a $100B figure for AI-cluster optics by 2030 circulates widely and originates as a question posed in a newsletter headline rather than as a published point forecast — it is noted here only so that readers who encounter it know its provenance. Cross-refs: §06·F·B (the backlog these networks serve), §06·F·D (the capex line they sit inside), §05·F (the accelerators being connected), §06·F·P (the power the saving is measured against), §03·F (the same "shortage clears next year" pattern in DRAM). Sources: DIGITIMES, LightCounting, Nvidia, Broadcom, TSMC, IDTechEx, SemiAnalysis, Lambda, The Register. Not investment advice.

06·F·LElectricians — The Labour Constraint on the Power Build-Out◷ JOLTS cited · shortfall estimates conflict
06·F·L

Electricians — The Labour Constraint

Follow this board's chain to its end and it stops at a person holding a conduit bender. §06·F·D counts $808B of hyperscaler capex; §06·F·C counts ~31.5 GW of committed compute; §06·D·P converts token demand into gigawatts of continuous load. None of that becomes energised capacity — and none of it buys a single HBM stack — until somebody physically wires it. Electricians are the last mile of the AI trade, and both Nvidia's and Microsoft's leadership have named the shortage as a binding constraint on data-centre expansion. This section tracks the openings data honestly, which means leading with an inconvenience: the official openings series does not show a boom.

The divergence — read this before the chart
The only official, continuous, seasonally-adjusted series for job openings in this space is BLS JOLTS for all construction — and it peaked in 2022 at ~399K and now sits ~38% below that. Taken alone, that reads as a cooling labour market. It is almost certainly a composition effect, not a demand signal: residential construction has been squeezed by rates while data-centre and industrial electrical work boomed, and the aggregate nets the two against each other. The electrician-specific evidence points the other way — BLS projects ~81,000 openings a year for the occupation, commercial apprenticeship applications jumped from ~70,000 to 120,000 between 2022 and 2024, IBEW Local 26 in the D.C. data-centre corridor doubled its membership since 2018 to ~14,700, and data-centre electricians are being quoted $150K+, with some packages at $240–280K. Wages and apprenticeship intake are screaming shortage while the aggregate openings count drifts down. This board shows both rather than picking the flattering one — and notes plainly that no public series measures electrician openings monthly, which is the actual data gap.
Construction job openings — annual average, 2015 → 2026
Bars are the annual average of the monthly BLS JOLTS series (seasonally adjusted, thousands); the vertical whisker on each bar is that year's monthly low-to-high range, which is wide — this series is noisy month to month. 2022 is the peak year; 2026 is marked * because it averages January–May only (the last published print). Note the recent turn: May 2026 printed 298K, the highest month since July 2025. This is all construction, not electricians — see the caveat above and the provenance below.
The electrician-specific picture MIXED BASES — DO NOT SUM
These come from different sources measuring different things — a stock of employed workers, a projected annual flow, a retirement rate, and three competing shortfall estimates. They are listed side by side without being combined.
Average pay by quarter — the price signal the openings count misses
Quarterly average of the BLS monthly series for average hourly earnings of all employees in construction, 2019-Q1 → 2026-Q2 (the last complete quarter). Wages are the cleanest available read on scarcity: they have risen every single quarter of this series without exception, including through the period when the openings count above was falling. Same caveat as the openings chart — this is all construction, not electricians alone; no official quarterly wage series exists for the occupation. Hover for the year-over-year rate.
IBEW membership — five published figures, five different definitions NOT A QUARTERLY SERIES
The IBEW does not publish membership quarterly. It reports once a year through its DOL LM-2 filing, and its public communications quote several different figures that are not on the same basis. Rather than interpolate a quarterly line that does not exist, this board shows the published numbers side by side — the 210,000-member spread between them is itself the finding, and it turns on whether you count active members, retirees, fee payers, or the union's headline claim.

What is consistently reported is the direction: record organising, with 24,000+ new members added in a single year — described by the union as its biggest one-year growth in more than 50 years — against a 1972 peak near 1,000,000. For quarterly-resolution membership you would need the LM-2 microdata at olmsapps.dol.gov, which is annual by construction.

Where the job boards actually help — and where they don't. The IBEW jobs board (ibew.org/jobs-board) is the most direct window into union electrical demand, but it is worth being precise about what it is: a directory of district and local-union referral pages, not an aggregated national listing. Each local runs its own "book" and posts its own calls, so the site returns where to look rather than how many — it publishes no national openings count and no historical series, which is why no number from it appears on the chart above. The same is broadly true of Indeed, ZipRecruiter and the specialist data-centre boards: they show a live snapshot of postings, and their historical archives are proprietary and not comparable across sites. So the honest construction is the one used here: an official, continuous, methodologically stable series for the aggregate (JOLTS), plus dated occupation-level figures from BLS and industry, plus a clear statement that the specific series a reader might want — monthly electrician openings — does not exist publicly.

Provenance & the honesty line. [cited — official, continuous]: the chart is BLS Job Openings and Labor Turnover Survey, Job Openings: Construction (FRED series JTS2300JOL), seasonally adjusted, monthly, level in thousands, series running Dec 2000 → May 2026 (last updated Jun 30 2026). Annual averages computed here from the full monthly record; 2026 = Jan–May average (5 months) and is flagged partial. Reference points from the same series: monthly peak 457K (Mar 2022), monthly trough of the recent cycle 135K (Dec 2024), latest print 298K (May 2026). [cited — occupation level, different bases]: 818,700 electricians employed (BLS Occupational Outlook Handbook, 2024); ~81,000 projected openings per year 2024–34 and +9% employment growth (BLS Employment Projections — note this is an annual flow including replacement demand, a fundamentally different quantity from a JOLTS point-in-time stock, and the two must not be plotted together); ~20,000 retirements a year with ~30% of union electricians near retirement age (Manufacturing Institute / Deloitte, 2024); commercial apprenticeship applications ~70,000 → 120,000, 2022–24; IBEW Local 26 membership roughly doubled since 2018 to ~14,700. [conflicting — shown, not reconciled]: the AI-attributable shortfall is variously ~300,000 (Fortune), 439,000 (ITIF) and 499,000 (iRecruit, data-centre construction specifically) — three different scopes and horizons; this board reports the range rather than choosing. Wages [cited, wide dispersion]: IBEW journeyman $56–78/hr all-in vs $42–62/hr non-union straight time; data-centre specialists $150K+ with travel and overtime, and quoted packages of $240–280K at the top end; Northern Virginia and the Bay Area lead, with Phoenix, Dallas and Columbus rising fastest. Four caveats: (1) The headline series is not the requested one. JOLTS measures all construction; no public monthly series isolates electricians, so the chart proxies the sector and says so. (2) Aggregate ≠ segment — residential weakness and data-centre strength offset inside the same number, which is why the trend is misleading on its own. (3) Job-board counts are not comparable to JOLTS: postings can be duplicated, stale, or aggregated across sites, and none of the major boards publish a consistent historical archive. (4) Shortfall estimates are advocacy-adjacent — they come from industry and think-tank sources with a stake in the answer, and their definitions differ; treat the range as a range. Cross-refs: §06·F·C (the gigawatts these workers must energise), §06·F·D (the capex funding it), §06·D·P (the power the tokens demand). Sources: U.S. Bureau of Labor Statistics (JOLTS via FRED; Occupational Outlook Handbook; Employment Projections), IBEW, Manufacturing Institute/Deloitte, ITIF, Fortune, iRecruit, Qmerit, BuildForce, Rinvio. Not investment advice.

06·F·BCloud Backlogs — The $2.5 Trillion RPO Wall◷ RPO anchored · early years sparse
06·F·B

Cloud Backlogs (RPO) — Four Hyperscalers + Two Neoclouds

The demand-side mirror of the capex spend: Remaining Performance Obligations (RPO) — contracted-but-not-yet-recognized cloud revenue, the most forward-looking disclosure these companies make. In 2026 the four big clouds — Microsoft, Oracle, Google, Amazon — collectively carry over $2.3 trillion in backlog, and adding the two listed neoclouds — CoreWeave ($103.7B) and Nebius (~$40B) — takes the six-firm total to $2.470 trillion, up from a few hundred billion a year earlier, as AI mega-deals (OpenAI's Stargate, Anthropic's multi-cloud pacts) landed in the books. This is the contracted demand that, in turn, drives the memory orders. The neoclouds are only ~6% of the total and are the most interesting part of it: CoreWeave's backlog grew 246% year-on-year and is now 20× its annual revenue, against roughly 2–3× for the hyperscalers — the same contracts, on a balance sheet with none of the cash flow behind it. Honesty note: RPO became a standardized disclosure only after the ASC 606 accounting standard (~2018), so a true 10-year history doesn't exist for the early years — those are shown as sparse / not-comparable, not invented.

Remaining Performance Obligations — stacked by quarter, 2021 → 2026
How to read it: each bar is one calendar quarter; the segments stack the six firms' contracted-but-unrecognised revenue into a combined total, on a linear scale. This replaces the earlier log-scale line chart, and the trade is deliberate: a log chart showed each company's trajectory well and its sum not at all; the stack shows the sum and flattens the small firms. Nebius at $40B is a sliver against Microsoft's $678B, and that is the honest visual relationship. Read the opacity, because it carries the honesty: a solid segment is a quarter that company actually disclosed; a faded segment is interpolated between two disclosures. In the old line chart interpolation was visible as a plain line between dots — in a stacked bar it would be invisible, so it is encoded as transparency instead. The tooltip says how many of the reporting firms disclosed each quarter. Early bars are mostly faded: Google and AWS only began breaking RPO out around 2022, CoreWeave from its 2025 IPO, Nebius only after its Sep-2025 Microsoft deal — a segment that is absent means no disclosure, not zero backlog. One structural caveat this chart creates and the old one did not. Stacking implies these numbers are simultaneous, and they are not: Microsoft's fiscal quarters end Jun/Sep/Dec/Mar, Oracle's fiscal year ends in May, and the other four are calendar. Each bar therefore mixes fiscal calendars offset by up to a quarter. That was harmless when the lines were independent; it is a real approximation now that they are summed. Values are labelled at the right
The two new segments are small on the chart and are not small as a risk
A dollar of Microsoft backlog and a dollar of CoreWeave backlog are not the same object. Microsoft's $678B sits on a company with ~$100B of annual operating cash flow; CoreWeave's $103.7B sits against $5.13B of 2025 revenue, and the capacity to deliver it has to be bought first, largely with debt. Its own disclosure is explicit that only ~41% of RPO is expected to be recognised within 24 months — the rest depends on building data centres that do not exist yet. Nebius is smaller and more concentrated still: its ~$40B is essentially two contracts — Meta at up to $27B and Microsoft at $17.4B (plus $2B optional) — so a single counterparty renegotiating is not a haircut, it is most of the book. And the circularity is worth naming. Microsoft and Meta appear on this chart as buyers of Nebius capacity while Microsoft simultaneously carries the largest backlog on it as a seller. The same demand can therefore be counted twice at two different companies — once in Nebius's backlog and once, downstream, in the hyperscaler contract it serves. The $2.47T total is a sum of disclosures, not a sum of distinct end-demand, and this board cannot size the overlap from public filings. Treat the combined figure as an upper bound. Cross-ref §06·F·E, where the same firms' debt is what funds the delivery.
The same backlog in megawatts — what $2.47T of contracts implies about power
Nobody reports backlog in megawatts. This converts the dollars above into implied contracted IT capacity, using three $/MW anchors that each come from a source disclosing both the money and the power — two of them companies on this very chart, which is why they bracket the answer rather than confirm it. The bars use the central rate; the dashed band is the full spread between the cheapest and dearest anchor, and it is wide on purpose: the honest content of this chart is the range, not the midpoint. Only the AI-identified portion of each backlog is converted. Microsoft's $678B is not all AI — it is Azure, M365 and everything else, and applying a GPU-cloud rate to the whole thing would be nonsense. So the hyperscalers contribute only their identified OpenAI + Anthropic dollars (a floor — "other" certainly hides more AI), while CoreWeave and Nebius contribute 100% because they sell nothing else. That is why the bars are roughly half the height a naïve conversion would give. The ramp before 2026 is modeled, not measured. The AI decomposition is a single May-2026 snapshot; there is no quarterly history of it. Each hyperscaler's AI share is therefore ramped linearly from zero at end-2024 — before the mega-deals were signed — to its identified share today. The early bars are a shape, not a measurement. The teal line is the reality check: §06·F·C's 31.5 GW of compute that OpenAI and Anthropic have announced themselves. This chart sits above it, which is what should happen — the labs' figure counts only their own announced deals, while the backlog counts every AI customer these six firms have.
The three anchors — and why two of them are on the chart above
Each row is a source that publishes both a dollar backlog and a power figure, so the ratio needs no assumption. CoreWeave and Nebius disagree by 2.1× — $28.0M/MW against $13.3M/MW — for the same product in the same quarter, which tells you how much contract duration, power definition (IT load vs facility) and deal mix move this number. The central rate is their geometric mean with the labs' cross-read; it is not a consensus, it is a midpoint between things that disagree.
A self-check the model half-passes, and the memory number it implies
Run the two disclosing companies back through the blended rate and see how far off it lands. CoreWeave's $103.7B at the central $23.2M/MW implies 4.5 GW against the 3.7 GW it actually reports contracted — 21% high. Nebius's $40B implies 1.7 GW against ~3.0 GW reported — 43% low. Neither is close, and they miss in opposite directions, which is exactly what a geometric mean of two disagreeing anchors must do. Treat any single company's bar as ±40%; treat the total as an order of magnitude with a defensible band. What it means for this dashboard. At §06·F·C's stated density — 1 GW ≈ 590K accelerators ≈ ~113 PB of HBM at 192 GB each — the central 51 GW implies roughly 30 million accelerators and on the order of 5.8 exabytes of HBM to be manufactured across the delivery window. That is the number this section exists to produce: the backlog is not an abstraction, it is a bill of materials, and §05·C·G's wafer-per-generation arithmetic is what has to absorb it. And the power itself is the harder constraint — 51 GW is roughly fifty large nuclear reactors of continuous draw, which is why §06·F·L tracks whether there are enough electricians to install it.
Backlog by customer — May 2026 decomposition (The Information)

Customer split is from The Information's May 21 2026 snapshot; Oracle's bar adds a hatched $85B segment for contracts landed after the snapshot, so bar lengths match each firm's latest reported total (Oracle $638B by May 31) rather than misleadingly drawing the current leader shortest.

OpenAI's spending commitment Anthropic's spending commitment Other revenue backlog Added after the May snapshot (Oracle +$85B)
RPO trajectory & AI concentration

Provenance & the honesty line. RPO/backlog [anchored — company 10-Q/10-K & earnings, quarterly]: Microsoft commercial RPO (FYE Jun): $141B (Jun'21) → $189B (Jun'22) → ~$224B (Jun'23) → ~$269B (Jun'24) → ~$315B (Mar'25) → $368B (Jun'25) → $398B (Sep'25) → $625B (Dec'25, +110% YoY) → $627B (Mar'26, +99%) → $678B (Jun'26, FQ4 reported Jul 29 2026, +84% YoY). The Q4 print is the most thesis-relevant disclosure in this section: Microsoft stated the entire ~$51B sequential increase came from customers other than the frontier-model companies, and that RPO still grew +25% excluding OpenAI — the first hard evidence that hyperscaler backlog growth is broadening beyond a handful of AI labs. Weighted-average duration is 2.3 years with ~30% recognizable within 12 months (up 37% YoY), so this backlog converts faster than Oracle's. Context from the same print: Q4 revenue $90B (+18%), Azure +43% and past $100B for FY26, quarterly capex and finance leases $41B (+69%) — but CY2026 capex guidance was cut from ~$190B to ~$175B, a datapoint that belongs to kill-switch #3 as much as to this section. Because the disclosed add was non-frontier, this board assigns the full +$51B to the other revenue backlog segment rather than to OpenAI or Anthropic (Microsoft 8-K/IR, FY26 Q4 earnings call, 10-K). Correction from a prior build: an earlier version showed Microsoft ~$112B for 2023 — that was the cloud-only RPO subset; total commercial RPO was ~$224B then. Now fixed. Oracle RPO (FYE May): ~$44B (FY21) → $65B (Dec'22) → $80B (Feb'24) → $98B (May'24) → $99B (Sep'24) → ~$138B (May'25) → $455B (Sep'25, +359% YoY) → $523B (Dec'25, +433%) → $553B (Feb'26) → $638B (May'26, +363% YoY, up $85B sequentially) — the $300B OpenAI/Stargate deal is the anchor (Oracle 8-K/IR, DCD, Cloud Wars). Google Cloud backlog (disclosed from ~2022): ~$52B ('23) → ~$92B (Mar'25) → "nearly doubling quarter on quarter to over $460 billion" ($468B, Alphabet Q1 2026 8-K) → $514B (Q2 2026, reported Jul 22 — up >$50B sequentially, with just over half to be recognized within 24 months; Q2 Cloud revenue $24.8B, +82% YoY, on Alphabet revenue $119.8B +24%). AWS RPO (cited from ~2022): ~$105B ('23) → $189B (Mar'25, +20%) → $364B (Mar'26, +93% YoY), excluding a new $100B OpenAI commitment on top of an existing $38B contract → $496B (Q2 2026, reported Jul 30 — up $132B / +36% sequentially, with triple-digit YoY growth). The same print showed AWS revenue $42.2B, +36.7% YoY — a fifth consecutive quarter of acceleration — with operating margin at 39% (+650bps), on Amazon revenue of $200.6B (+20%) and operating income $27.5B (+43%). The +$132B sequential add is not customer-decomposed, so this board books it to other revenue backlog rather than attributing it to OpenAI or Anthropic; that also means AWS's reported figure now exceeds the old commitment-inclusive $464B, retiring the two-basis flag this section previously carried (Amazon IR, CNBC, Global Data Center Hub, Cloud Wars). Current-snapshot decomposition [The Information, May 21 2026]: Microsoft $280B OpenAI + $30B Anthropic + $317B other ≈ $627B (May snapshot; the Jun'26 FQ4 print took the total to $678B and Microsoft attributed the whole +$51B to non-frontier customers, so it is added to other here — the OpenAI/Anthropic legs are unchanged from the May decomposition and are not re-derived); Oracle $300B OpenAI + $253B other ≈ $553B; Google $200B Anthropic + $268B other ≈ $468B (May snapshot; Google Cloud backlog has since risen to $514B at Q2 2026, reported Jul 22 — the +$46B QoQ add not yet customer-decomposed); Amazon $138B OpenAI + $100B Anthropic + $226B other ≈ $464B at the May snapshot; the Q2 print took the reported total to $496B and the +$132B add is not customer-decomposed, so it is carried in other here (per the reference chart; The Information / JPMAM). The honesty caveats: (1) RPO standardized disclosure dates to ASC 606 (~2018) — pre-2019 figures are sparse / not comparable and are drawn faint, not fabricated. (2) Fiscal calendars differ (Oracle FY ends May 31; Microsoft Jun 30; Amazon/Alphabet Dec 31) — the time axis is approximate calendar-aligned. (3) RPO is contracted revenue, not guaranteed cash — concentration in a few AI customers (OpenAI ~45% of MSFT's, $300B of Oracle's) is real counterparty risk, and skeptics question whether OpenAI can honor ~$60B/yr commitments. (4) The customer decomposition is reporting-derived (The Information), not a GAAP line item. Modeled: the connecting curves between anchored quarters. Sources: company 10-Q/10-K & earnings calls, The Information, JPMAM, DCD, Sherwood, Constellation, Cloud Wars, Global Data Center Hub. Not investment advice. [added Aug 11 2026 — CoreWeave & Nebius, and the chart form changed]: CoreWeave [cited] — revenue backlog ~$16B (Dec 2024), $25.9B (Mar 2025), $30.1B (Jun 2025), >$55B (Sep 2025), $66.8B (Dec 2025, 4× the start of the year), $99.4B (Mar 2026), and at Jun 30 2026 $103.7B of RPO / ~$104B backlog, +246% YoY, on Q2 revenue of $2.58B and FY2025 revenue of $5.13B; ~41% of RPO expected within 24 months; the figure excludes >$25B of commitments signed after the quarter closed. Definitional mismatch, disclosed: CoreWeave reports "revenue backlog" = RPO plus estimated future revenue under committed contracts, a broader measure than the RPO the four hyperscalers report; at Jun 2026 the two are within $0.3B of each other, but the earlier quarters in this series are the broader measure and are therefore not strictly like-for-like with the rest of the stack. Nebius [cited] — contracted backlog ~$40B at Q2 2026, built on Meta up to $27B and Microsoft $17.4B + $2B optional = $19.4B; ARR $3.0B at end-June 2026 (from $1.92B at end-March), Q2 revenue $582.3M (+454% YoY), FY2026 guidance $3–3.4B revenue and $7–9B exit ARR. Source conflict, not merged: the company's own Q2 figure is $40B while secondary coverage cites "nearing $50B" — the larger number sums the headline deal values including optional tranches ($27B + $19.4B = $46.4B). This chart plots the $40B contracted figure and prints both. Nebius's pre-2026 points are sparse and partly reconstructed — it did not quantify a backlog before the Sep-2025 Microsoft deal, so those bars are drawn faded and should not be read as measurements. Chart form: the log-scale line chart was replaced by a linear stacked-bar chart at request. Interpolated quarters, previously implicit in the line between dots, are now drawn at reduced opacity; the underlying per-company series and anchors are unchanged. The stack necessarily mixes fiscal calendars (Microsoft FY ends June, Oracle May, the rest December). Aug 17 2026 — the megawatt conversion now has a direct anchor [§06·F·S]. This section's RPO-to-gigawatt chart rested on an assumed $/MW, which was its weakest input. Two neoclouds have since disclosed it independently: Nebius closed four billion-dollar-plus deals at $20–25M per megawatt, and CoreWeave's $104B backlog over 4.2GW of contracted power implies $24.8M/MW. Two firms, one number, inside the range used here.

06·F·CCompute Commitments in Gigawatts — OpenAI vs Anthropic◷ point-in-time compile · memory bridge illustrative
06·F·C

Compute Commitments (GW)

The RPO section above shows the money; this shows the electricity — publicly disclosed compute commitments by the two frontier labs, in gigawatts. Power is the constraint upstream of everything in this dashboard: a gigawatt committed is land, transformers, accelerators, and — the reason it's here — a fixed ratio of HBM and DRAM that must be manufactured. As of the June 2026 compile, OpenAI has ~18.75 GW committed across Stargate/Oracle, AMD, AWS and Cerebras, and Anthropic ~12.8 GW across AWS, two Google structures and SpaceX — of which ~10.8 GW was added in April–June 2026 alone.

OpenAI ≈ 18.75 GW committed
Stargate (Oracle) 10
AMD MI450 6
AWS 2
Anthropic ≈ 12.8 GW committed · +10.8 GW added Apr–Jun 2026
Pre-Apr ~2
AWS 5
Google+Broadcom 3.5
Google 2
Stargate (Oracle) 10 GWAMD MI450 6 GWAWS 2 GW / 5 GWCerebras 0.75 GWAnthropic pre-April base ~2 GWGoogle + Broadcom 3.5 GWGoogle 2 GWSpaceX 0.3 GW
The memory bridge [illustrative]. At NVL72-class density (~120 kW per 72-GPU rack → ~1.7 kW per accelerator all-in, cooling and networking included), 1 GW ≈ ~590K accelerators ≈ ~113 PB of HBM at 192 GB each. The two labs' ~31.5 GW combined therefore implies on the order of ~3.6 exabytes of HBM-class memory to be manufactured across the deployment window — before counting DRAM, storage, or any other buyer. Stated assumptions, not a forecast: real deployments span years and mix silicon with very different memory content (AMD MI450, Trainium, TPU, Cerebras wafer-scale).

Provenance. Commitment figures are a point-in-time June 2026 compile (The Product Compass / TheWhiteBox Consulting / JPMAM) of publicly announced deals: OpenAI — Stargate with Oracle 10 GW, AMD MI450 6 GW, AWS 2 GW, Cerebras 0.75 GW (≈18.75 GW); Anthropic — AWS 5 GW, Google + Broadcom 3.5 GW, Google 2 GW, SpaceX 0.3 GW atop a ~2 GW pre-April base (≈12.8 GW, ~10.8 GW of it added April–June 2026). Caveats: commitments ≠ deployed draw — these are multi-year build-outs; announced GW definitions vary (IT load vs facility); the compile predates any July deals; and the memory bridge is labeled illustrative with its per-GPU wattage and 192 GB assumptions shown. Not investment advice. See also: §06·F·P, which asks through which physical route — grid queue, gas turbine or nuclear PPA — these committed gigawatts can actually be delivered, and by when.

06·F·MMicron RPO & Strategic Customer Agreements — the Memory-Side Backlog◷ FQ3 disclosures · first print of a new regime
06·F·M

Micron RPO & SCAs

The memory-side mirror of the hyperscaler backlogs above: with FQ3 (Jun 24), Micron began disclosing remaining performance obligations under a new regime of 16 take-or-pay Strategic Customer Agreements running 2026 through end-2030. Reported RPO at quarter-end was >$5B — but that counts only agreements executed by May 28; including the SCAs signed just after, RPO is ≈ $100B, with 14 of 16 agreements entering the FQ4 reporting. The construction is deliberately conservative: minimum committed volumes at floor pricing (Micron expects revenue to "well exceed" RPO), with price ceilings set at CQ2-2026 levels — which is also a cap worth remembering.

RPO · FQ3-executed
>$5B
agreements signed by May 28 only
RPO · incl. post-Q SCAs
≈$100B
14 of 16 report in FQ4; detail in the 10-K
Agreements
16 SCAs
take-or-pay · terms 2026 → end-2030
Capacity locked
~20% / ~30%
of DRAM / of NAND output under contract
Customer deposits
$22B
$18B cash + ~$4B letters of credit
Deposit timing
~$10B in FQ4
>$400M received in FQ3

The FCF trap, flagged: the $18B of cash deposits is not prepaid revenue — it flows through financing cash flow and is returned over the back half of each agreement. Adding it to FQ3's record $18.3B adjusted FCF double-counts; this dashboard doesn't. Near-term revenue from the FQ3-closed agreements is small (~$1.8B next 12 months — early automotive deals); the $100B substance arrives with the FQ4 10-K. Read against §06·F·B: cloud buyers now carry ~$2.2T of committed demand, and the memory supplier they depend on has, for the first time, ~$100B of committed supply — floor-priced, ceiling-capped, and deposit-backed. Calendar: Jul 10 2026 — SK Hynix ADR begins U.S. trading; Dec 9 2026 — CHIPS-grant two-year mark unlocks Micron's shift toward returning ~100% of excess cash.

Provenance. All figures from Micron's FQ3-2026 prepared remarks, 8-K press release, 10-Q and earnings call (Jun 24–25 2026): RPO >$5B as reported / ≈$100B including SCAs executed after quarter-end (14 of 16 in FQ4 reporting); 16 take-or-pay SCAs, generally 2026–2030; ~20% of DRAM and ~30% of NAND capacity committed (sources range 30–33% on NAND); $22B of deposits and financial commitments ($18B cash, ~$4B LCs), >$400M received in FQ3 and ~$10B expected in FQ4, returned in the back half of each agreement via financing cash flow; RPO built on minimum volumes at floor prices with ceilings at CQ2-2026 levels, revenue expected to "well exceed" RPO; next-12-month revenue on FQ3-executed agreements ≈$1.8B. Honesty notes: this is the first print of a brand-new disclosure — definitions will firm up in the 10-K; Micron's RPO (floor-priced minimum volumes) is not directly comparable to cloud RPO (contracted service revenue), so the §06·F·B mirror is a framing, not an equivalence. Not investment advice.

06·GAI Accelerator Rental Pricing
06·G

AI Accelerator Rental Pricing

● LIVE · 3RD-PARTY

What it costs to rent an AI accelerator by the hour is one of the cleanest real-time reads on accelerator supply/demand — and that same tightness is what pulls HBM and memory demand through the whole chain on this page. When GPUs are scarce, rental rates spike; when capacity floods in, they fall. Embedded below is Ornn's Compute Index, a live tracker of historical GPU rental price indices (H100, H200, B200, A100, RTX 5090).

Ornn Compute Index · historical GPU rental price indices Open live tracker ↗
🔌
The live tracker isn't displaying here

That's expected in two situations: inside the Claude artifact preview (external sites are blocked in the sandbox), or if Ornn doesn't permit its dashboard to be embedded in other pages. Open this file directly in your browser, or just visit the tracker:

Open Ornn Compute Index ↗

This panel embeds a third-party live website — Ornn's Compute Index (ornn.com) — not data compiled by this dashboard. Its figures, methodology, coverage and uptime are entirely outside our control, and the embed may not render in every context (the Claude artifact sandbox blocks external frames, and a provider can disallow being framed via response headers); a direct link is provided as a fallback. Why it matters here: GPU rental rates are a fast, market-set proxy for accelerator scarcity — sustained high rates signal tight supply (bullish for the memory makers feeding those accelerators), falling rates can foreshadow a demand cooloff. Rental pricing reflects many factors beyond memory (compute supply, financing, power, contract vs spot), so treat it as one signal among the others on this page, not a direct memory-price gauge. Source: Ornn (dashboard.ornnai.com). Not investment advice. Aug 17 2026 — this section embeds a live third-party dashboard but holds no figures of its own. §06·D·C now carries the numbers, including the argument that the best single demand gauge is the A100 rather than the H100, the near-3× premium between hyperscaler and commodity pricing for the identical H100, and the CME/Silicon Data compute futures launching Oct 5 2026.

06·D·CAI Compute Demand — Read Off the Price of Obsolete Silicon◷ rental + token indices · Aug 2026
06·D·C

AI Compute Demand

§06·G embeds someone else's live rental dashboard but carries no numbers of its own. This section carries the numbers, and it takes a specific and unobvious view of which one matters. The best single gauge of AI compute demand is not the H100 rate — it is the A100's. A chip released in May 2020, six years and three months old, still rented at near-full capacity and now used predominantly for inference. Its median on-demand rent rose from $1.70 to $1.76 over the year to August 2026. New silicon has kept arriving the whole time; the old silicon got more expensive. That is hard to explain with a training bubble and easy to explain with broad, price-insensitive inference demand. It also happens to be the cleanest live test of the argument in §06·F·N — because on the six-year schedule Microsoft and Alphabet book, the A100 finished depreciating in May 2026, and CoreWeave has just contracted A100 capacity out to 2029.

What an accelerator-hour costs — by generation, and by who is selling it
Bars are published price ranges; the thick tick is the index or median. The A100 sits at 70% of the H100 commodity index despite being two and a half years older — generation barely discounts the price. But the same H100 bought from a hyperscaler costs nearly three times the commodity index, in a band so tight (0.47% coefficient of variation over seven weeks) that it reads as an administered price rather than a market one. The dashed outline on the H100 row is a second published range that does not agree with the first — spot GPU pricing is thin enough that two surveys of the same chip in the same month differ by nearly 2× at the top end.
The A100 is now a live experiment on the depreciation question
§06·F·N sets out an argument this board could not settle: how long is a GPU's economic life? Microsoft and Alphabet book six years, Amazon cut to five and took a charge, and the bear case — Burry's — is that the real answer is two to three, implying $176B of overstated earnings across 2026–28. The A100 is the only generation old enough to actually test this, and the test has started returning results. Released May 2020, a six-year life has it fully depreciated as of May 2026. It is now three months past that point, renting at a higher price than a year ago, at near-full utilisation. And CoreWeave disclosed on its Q2 call that it recently signed an A100 contract extending into 2029 at what it called an attractive price — roughly nine years after launch, and about three years beyond the longest life any of the four hyperscalers books. (The end date is disclosed only as "into 2029", so the true figure is somewhere between 8.6 and 9.6 years — the conclusion holds across that whole range.) If that is representative, the four are over-depreciating, not under, and the Burry case is not merely wrong but inverted. Three reasons not to bank it yet: a contract price described as "attractive" but never disclosed could be attractive to either party; rental rate is not residual value, and a fully-depreciated asset earning anything at all flatters returns without saying what it would sell for; and A100 demand is specifically inference demand, which is exactly the workload most exposed if token prices fall (§06·D·R). This is one data point from one generation, and it is the most informative one available.
From October, compute has a forward curve
CME Group and Silicon Data announced on Aug 11 that compute futures launch on Oct 5 2026. Every demand signal on this board — token indices, rental rates, backlogs, wafer starts — is backward-looking or survey-derived. A futures curve would be the first forward-looking, market-priced statement about AI compute demand that exists, and the first one that can be wrong in public and settle. Two things it would let this board do that it currently cannot: price the expected path of GPU rents rather than extrapolating the observed one, and give §06·F·R's lenders a hedging instrument, which changes the credit calculus behind GPU-backed debt. One caution: a new contract's early prints are usually illiquid and unrepresentative, so this board will not read a curve from it until there is enough volume to mean something — and that is a judgement, not a date.

Provenance & the honesty line. [cited]: Silicon Data's stated view that the A100 rather than the H100 is the best single gauge of AI compute demand — a May-2020 chip "rented out at near-full capacity", "used predominantly for inference", with steady rates read as a signal of inference-demand strength (Aug 14 2026). A100 median on-demand $1.76/hr, up from $1.70 a year earlier, with a published range of $1.09–$5.07. H100 Silicon Data rental index $2.53/hr; published ranges $1.49–$6.98 and, from a different survey, $2.19–$11.06. H100 hyperscaler on-demand index traded $7.40–$7.52 from Mar 1 to Apr 20 2026 with a 0.47% coefficient of variation. LLM Token Expenditure Index at $4.20 per million tokens (frontier) and $0.85 (open-weight) as of Apr 22 2026. GPU rental market sized at ~$52B in 2026 rising to ~$199B by 2031. CME Group and Silicon Data announced compute futures for Oct 5 2026 (press release, Aug 11). CoreWeave's Q2 call: an A100 contract extending into 2029 at an "attractive price", with older Nvidia generations in solid demand. [derived — this board's arithmetic]: the +3.5% A100 year-over-year, the 70% A100-to-H100-index ratio, the 2.95× hyperscaler-to-commodity channel premium, the 4.9× frontier-to-open-weight token gap, chip ages (6.3 and 3.9 years), and the observation that a six-year life from May 2020 expires in May 2026 — placing the CoreWeave contract about three years beyond the longest booked life. What is deliberately not claimed: (1) no residual-value conclusion — a rental rate is a flow, a depreciation schedule is about a stock, and this section does not convert one into the other; (2) no forecast of A100 rates, because the whole argument is that this series is informative precisely when it is observed rather than modelled; (3) no aggregate "AI compute demand" number — the market-size figures above are third-party and are printed for scale only, not used in any calculation here. Two source problems, stated: the two H100 ranges are irreconcilable and both are drawn; and the token index is dated April, four months older than the rental data on the same chart, so the two are not contemporaneous and should not be read as a single snapshot. Cross-refs: §06·G (the live third-party rental dashboard), §06·F·N (the depreciation argument this tests), §06·F·S (CoreWeave's disclosures, including the A100 contract), §06·D (token demand), §06·D·R (usage share, and the inference exposure), §06·F·R (the lenders a futures market would serve), §02·D (kill-switch #3). Not investment advice.

06·HAWS Capacity Blocks — GPU Reservation Pricing◷ current anchored · history triangulated
06·H

AWS Capacity Blocks — GPU Reservation Pricing

A rare public, market-clearing price for the AI compute that drives memory demand. EC2 Capacity Blocks for ML let you reserve NVIDIA GPU clusters (H100/H200/B200/B300) for fixed windows, and AWS states the reservation fee is "updated regularly based on trends in supply and demand" — so it's effectively a dynamic spot price for guaranteed accelerator capacity. The current per-accelerator rates below are pulled from AWS's pricing page (current to its May 2026 update; AWS lists the next change for July 1, 2026). The 3-year history is the interesting part: on-demand GPU prices fell ~44% (H100 commoditizing) while reserved Capacity Blocks rose — a scarcity premium for guaranteed access. Source: aws.amazon.com/ec2/capacityblocks/pricing.

Effective hourly rate per GPU — 2023 → 2026
USD per accelerator-hour (instance rate ÷ GPUs), US regions. ● = an anchored data point (AWS page, dated examples, or cited reporting); lines between are interpolated. Two product lines move on different schedules: H100 on-demand held ~$6.88 then was cut ~45% in Jun 2025 (to ~$3.93, an AWS-page-dated figure); H100/H200 Capacity Blocks were cut separately (once in 2024, twice in 2025) but then raised +15% in Jan 2026 and again Jul 2026 — a scarcity premium. Newer Blackwell parts (B200/B300) price at a large premium.
Current Capacity Block rates (per accelerator-hour, US regions)

Provenance. Current rates [anchored — AWS pricing page, fetched, current to its May 13 2026 update]: per-accelerator effective hourly rate — P6e GB200 $10.582, P6-B300 $11.70, P6-B200 $10.296, P5 (H100) $4.326, P5e (H200) $4.975, P5en (H200) $5.721, P4d (A100) $1.475, Trn2 $2.235; per-instance examples e.g. p5.48xlarge (8×H100) $34.608/hr, p6-b200.48xlarge (8×B200) $82.368/hr, u-p6e-gb200x72 (72×B200) $761.904/hr. Upcoming [AWS page, effective Jul 1 2026]: P6-B300→$14.04, P6-B200→$12.355, P5→$5.191 (US), P5e→$5.97, P5en→$6.865 (US), P4de→$2.214. History [now with dated anchors, triangulated from reporting]: Capacity Blocks launched Oct 30 2023, P5 (H100) only in US East (Ohio), ~$60+/hr/instance (~$7.50/GPU-hr); Nov 2024 expanded to P4d/P5e/Trn1 with up-to-6-month durations; P6-B200 launched May 15 2025. Anchored mid-point: AWS's own pricing-page example dated Mar 20 2025 shows p5.48xlarge at $31.464/instance in US East Ohio = $3.933/GPU-hr. Two separate price tracks (a correction/clarification): the Jun 2025 ~45% cut applied to on-demand & Savings Plans, and Capacity Blocks were explicitly NOT included (DCD, CloudOptimo) — instead AWS cut Capacity Block prices on their own schedule ("once in 2024, and twice in 2025," per IT Pro/The Register), then raised them ~15% in Jan 2026 (p5e $34.61→$39.80/inst, p5en $36.18→$41.61, US West steeper $43.26→$49.75) and again Jul 1 2026. So the on-demand line and the Capacity Block line move on different cadences — the chart now distinguishes them. Modeled: the connecting curves between anchored dates, and the split between on-demand vs Capacity-Block series where a given month wasn't separately reported — daily rate histories aren't published. Caveats: Capacity Blocks are technically sold as a fixed total per block (min 1 day), shown here as the AWS-published effective hourly rate; rates vary 5–15% by region (US shown); on-demand and Capacity Block are different products (the chart distinguishes them). Source of record: aws.amazon.com/ec2/capacityblocks/pricing. Not investment advice. See also: §06·G, which tracks what an accelerator-hour actually rents for across providers — the price side of the capacity this section reserves.

07The Storage Layer◷ structural
07

The Storage Layer

● NAND & HDD

Below the chips, AI's data exhaust drives the other half of the supercycle: enterprise SSDs for hot data and a hard-drive oligopoly for the cold tier. Both are sold out.

NAND Flash — Revenue Share, Q4 2025
Samsung 28
Hynix 22
Kioxia 16
Micron 13
SanDisk 10
other

Source: TrendForce Q4 2025 (approx.) · SK Hynix incl. Solidigm

The flash field is a layer race. Kioxia and SanDisk — allied for 25+ years across the Yokkaichi and Kitakami fabs even after SanDisk's spinoff — co-developed BiCS FLASH and are pushing past 300 layers in 2026. Enterprise SSD demand is the swing factor: SK Hynix's NAND revenue jumped ~48% in a single quarter on it.

Hard Disk — The Three That Are Left
Seagate
HAMR · density
All-in on HAMR (Mozaic). Shipping 44TB; 100TB-class ahead. NVMe HDDs with Nvidia.
Western Digital
UltraSMR · volume
Nearline volume leader; ~89% cloud revenue. UltraSMR now, HAMR to 100TB by 2029.
Toshiba
MAMR · cautious
FC-MAMR 28–34TB; HAMR as test vehicles 2026–27. Step-by-step scaling.

From 200+ makers in the 1980s to three today. Nearline drives are ~90% of industry exabytes, and both WD and Seagate have sold out 2026 with agreements into 2028 — sending consumer drive prices up ~50% in five months. HDDs still win on cost-per-terabyte, so AI's cold data keeps the platters spinning.

07·BMemory at the Edge — Cars, Robots, AR/VR◷ model
07·B

Memory at the Edge — Cars, Robots, AR/VR

◷ model

The supercycle isn't only a datacenter story. AI is pushing memory into physical machines — and the standout is the self-driving car, which can carry more memory than a server. A robotaxi fuses lidar, radar and camera feeds in real time, runs neural nets on-board, logs everything for safety, and duplicates it all for redundancy. Robots and AR/VR add to the pull. None of this rivals datacenter volume yet, but it's the fastest-growing new demand — and, like phones, it's already feeling the squeeze (Meta raised Quest prices ~20% on the memory crunch).

🚗
Self-driving cars
Avg car: ~90GB → ~278GB (2023→26). L2 ~8GB; L4/L5 > 2TB. High-end hits 4TB by 2030. A robotaxi generates 16–20TB per hour of driving.
~3M fully autonomous + 15M+ L2+/L3 cars by 2030
🥽
AR / VR headsets
Quest 3 8GB, Vision Pro 16GB, Quest Pro 12GB. Heavy passthrough + on-device AI lift the floor each generation.
~9.6M units (2024) → ~37M by 2030 (≈23% CAGR)
🤖
Humanoid robots
Each needs FSD-class onboard compute + memory for real-time inference — tens of GB rising toward HBM-class. The softest forecast here by far.
~500 units (2025) → 0.1–1.2M by 2030 (estimates vary wildly)
Edge Memory Shipped — 2021 → 2030

Estimated DRAM + NAND shipped into these three categories per year, in exabytes. Automotive dominates — it's the only one at real scale today — but robotics is the steepest curve on a tiny base. Switch between viewing the stack by category or by memory type (DRAM vs NAND). Solid = historical estimate; hatched = projection.

Memory shipped into edge/physical-AI devices, exabytes/yr. Automotive = content-per-car × ~85M cars/yr; AR/VR = headset units × per-device memory; robotics = unit estimates × rising per-robot memory. By type, NAND (storage) is the bulk of capacity; DRAM is smaller in bits but higher-value. The 2026→ portion is modeled.

Why self-driving is the memory monster: content per vehicle by autonomy level

DRAM + NAND on board. The jump from driver-assist (L2) to full autonomy (L4/L5) is roughly 250×.

Edge memory is a modeled estimate (units × per-device memory content), and the three legs differ sharply in reliability. The DRAM/NAND split is itself an assumed ratio, not a separately-sourced series: by capacity (bits), storage dominates, so NAND is taken at ~86–93% of automotive and AR/VR content (rising slightly over time as autonomy/logging grows) and ~65–72% for robotics (which is more DRAM/HBM-weighted because real-time inference is bandwidth-bound). DRAM is far smaller in bits but higher-value per bit, so its revenue share is much larger than its capacity share shown here. A third type — NOR flash (boot code/firmware in ECUs) — is real in automotive but tiny by capacity and isn't charted separately. Automotive is the most solid leg: per-car content ~90GB (2023) → ~278GB (2026) rising toward 2–4TB on high-end/autonomous vehicles by 2030 (Micron); L1/L2 ≈ 8GB e-MMC, L3 ≈ 128–256GB, L5 > 2TB (TrendForce); ADAS+AD ≈ 40–45% of automotive memory revenue; ~3M fully-autonomous and 15M+ L2+/L3 vehicles expected by 2030, on a base of ~85M cars built/yr. AR/VR is moderate: shipments ~9.6M (2024) → ~37M by 2030 (TrendForce, ~23% CAGR), per-device memory 8–16GB (Quest 3 8GB, Vision Pro 16GB); already memory-constrained (Meta +~20% Quest pricing, Sony cut PSVR2). Humanoid robotics is the softest forecast in this entire dashboard — barely any units ship today (~500 in 2025; Tesla admits no Optimus does "useful work" yet), and 2030 estimates range from ~100K (Robotomated) to 1.2M (BofA) to Musk's "millions," with per-robot memory a guess; treat that layer as scenario, not forecast. Exabytes = billion GB. Sources: Micron, TrendForce, IDC, Verified/PS Market Research, BofA, company specs & disclosures. Not investment advice. See also: §07·B·R (humanoid robots), the newest and least-proven edge-memory demand pool on this board.

07·B·TTesla Shipments by Autonomy Level — the L2 Monolith (2020→2035)◷ deliveries anchored · L-split definitional · projection modeled
07·B·T

Tesla Shipments by Autonomy Level

The companion dataset to the "memory monster" ladder above (§07·B) — and the honest picture is starker than most coverage suggests: every consumer Tesla ever delivered is SAE Level 2. Autopilot and FSD (Supervised) require active driver supervision — Tesla's own filings say so verbatim — so owning the FSD package (1.28M cumulative holders) does not change a car's SAE level. L4 exists only in the unsupervised Robotaxi operation (live across the Austin metro plus two more Texas cities) and the Cybercab, which entered volume production in 1H 2026 with no steering wheel or pedals. Tesla holds no L3 certification anywhere and its stated architecture skips L3 entirely; L5 does not exist for any maker. The chart shows what that means in units: six solid years of L2, an L4 sliver appearing in 2026, and a modeled fan to 2035. Update (Jul 2): Q2 deliveries came in at 480,126 — a 25% YoY jump that smashed the ~406K consensus by ~74K and broke a two-year decline streak. Crucially, every one of those 480,126 is still L2 — a record quarter changes the bar height, not the autonomy split.

The memory bridge. This is why the split matters here: per §07·B, an L2+ vehicle already exceeds 300 GB of DRAM by 2026 (Micron), and L4 robotaxi content — redundant AI5-class computers, richer sensor buffering — is a multiple of that. Illustratively, the ~1.7M-car 2026 fleet-year at L2+ content alone ≈ ~0.5 exabytes of potential DRAM demand; the modeled 2035 mix (~3.7M L4-class units/yr) is the monster the section title refers to. Illustrative math — the per-level ladder lives in §07·B. The chart’s violet line uses fleet-average content (Tesla’s trim mix sits below the 300 GB flagship figure until late-decade), which is why its 2026 reading (~0.2 EB) sits beneath this all-at-flagship bound.
Deliveries by SAE level, 2020 → 2035 (millions of vehicles)
Solid bars = anchored actuals (company-reported; 2025 per the 10-K). 2026 = Q1 actual + consensus/model, drawn semi-solid. Hatched-outline bars beyond the divider = one modeled mid path (assumptions & the enormous bull/bear spread in the footnote). Blue = L2 (supervised — the entire consumer fleet). Red = L4 (unsupervised: Robotaxi fleet + Cybercab). The 2025 Robotaxi fleet (low thousands of converted Model Ys) is real but below this chart's resolution. The violet line (right axis) is the implied DRAM shipped per model-year — units × per-vehicle content — and it is wholly illustrative: the content curve is a stated assumption (footnote), not a disclosure. Its shape is the point: on this path 2035 units are ~9× 2020’s (and ~2.6× the 2023 peak) while memory shipped is ~900× — content per vehicle, not volume, is the story.
The zero rows, and why they're zero: L1 = 0 — base Autopilot (an L2 system) has been standard on every trim since 2019. L3 = 0 — Tesla holds no L3 certification in any jurisdiction and its stated approach skips L3, going supervised-L2 → unsupervised-L4 (contrast: Mercedes Drive Pilot is certified L3). L5 = 0 — geofence-free full autonomy does not exist for any maker, and no serious forecast puts it in production by 2035.

Provenance & the honesty line. Deliveries [anchored — company-reported]: 499,550 (2020) → 936,172 (2021) → 1,313,851 (2022) → 1,808,581 (2023) → 1,789,226 (2024, first annual decline) → 1,636,129 (2025, revised figure per the 10-K filed Jan 2026) → Q1 2026: 358,000+ delivered (+6.3% YoY; 408,000+ produced). Q2 2026: 480,126 delivered / 451,758 produced (Tesla 8-K, Jul 2) — up 25% YoY and 34% QoQ, ~74K above the 406,024 consensus (bullish estimates topped out ~418–420K), Tesla’s best-ever Q2 and first YoY growth since the 2023 peak; deliveries ran ~28K above production, working down Q1 inventory; Model 3/Y were 467,762 (97%), Other Models 12,364; 13.5 GWh energy storage deployed. H1 2026 = ~838K (358,023 + 480,126); full financials July 22. Note the split is unchanged by the beat — all 480,126 remain L2. FY26 tracking above the old ~1.65M consensus on this run-rate (Tesla 8-K/IR, SQ Magazine compile, Tesla Oracle, TechTimes). The L-split [definitional, not estimated]: an SAE level is a property of the system's design and certification, so the historical split is a matter of record rather than modeling: every consumer Tesla ships as L2 — Tesla's own investor materials footnote FSD (Supervised) with "Active driver supervision required; does not make the vehicle autonomous." The 1.28M cumulative FSD-package holders (Q1 2026) are therefore still L2 — owning the software ≠ a higher SAE level, a conflation this chart deliberately rejects. L4 [operational, small, partly modeled]: the unsupervised Robotaxi service is live (Austin metro + two additional Texas cities, priced under Waymo; paid Robotaxi miles roughly doubled sequentially in Q1 2026); the 2025 fleet was low-thousands of converted Model Ys (shown as ~0 at chart scale); Cybercab volume production commenced 1H 2026 per Tesla's Q4'25 and Q1'26 letters, but Tesla doesn't yet break Cybercab out of "Other models" (16,130 in Q1'26, which also holds S/X runout and Cybertruck) — so the ~30K L4 modeled for 2026 is an estimate, flagged. Projection [modeled — one mid path; the spread is enormous]: assumes an affordable-model-led L2 recovery, Cybercab becoming "the largest volume vehicle in the fleet over time" (Tesla's own Q1'26 language), and consumer unsupervised operation arriving late-decade in permitted regions only. Brackets: the Musk-scale case (2025 CEO award tranches: 20M vehicles, 1M Robotaxis in commercial operation, 10M FSD subscriptions) vs the analyst-flat case (~1.65M/yr persisting). Musk's autonomy-timing record — unsupervised driving promised repeatedly since 2016 — argues for discounting the top bracket, and this path sits well below it. If regulators force an L3-like intermediate for consumer unsupervised driving, some late-decade volume would route through L3; the chart follows Tesla's stated skip-L3 architecture. Memory bridge [illustrative]: per-level DRAM content and the Micron 300 GB+ L2+ anchor live in §07·B; the exabyte figure here is shipments × content arithmetic, not a demand forecast. The chart’s DRAM line uses these stated per-vehicle curves (GB, fleet-average new-build): L2 10 (2020) → 15 (’22) → 30 (HW4, ’23) → 60 (’25) → 120 (AI5 era begins, ’26) → 320 (’30) → 500 (’35), approaching Micron’s 300 GB+ L2+ figure late-decade as the mix upshifts; L4 ≈ 2× the consumer figure (600 → 1,100 GB) for redundant AI5-class computers and sensor buffering. Every point on that line is assumption × assumption — it is drawn to show shape (content growth dominating unit growth), not to forecast exabytes. Sources: Tesla 10-K, 8-K quarterly delivery releases, Q1-2026 shareholder letter, SQ Magazine, Tesla Oracle, TechTimes (Goldman/Barclays/RBC previews), notateslaapp. Not investment advice.

07·B·RHumanoid Robots — Units Shipped & the Coming Memory Wave◷ shipments anchored · forecasts diverge wildly
07·B·R

Humanoid Robots — Shipments & Memory Demand

A future memory-demand driver Micron is betting on directly. Each humanoid carries an on-board AI "brain" running vision-language-action models locally — and Micron's CEO says a humanoid will need ~10× the memory of a high-autonomy car (which itself exceeds 300GB DRAM by 2026). Today the market is tiny and pre-mass-production — ~13,000 units shipped worldwide in 2025 (Omdia) — but the maker field is crowded and forecasts run from Goldman's cautious 76k-by-2027 to Morgan Stanley's 1B+ by 2050. Honesty note: 2025 shipments are anchored; everything forward is a forecast, and the forecasts disagree by orders of magnitude. Per-robot memory is an estimate from the dominant compute platform, not a universal spec.

The memory math. The dominant humanoid "brain" — NVIDIA Jetson Thor (adopted by Figure, Agility, Boston Dynamics, Galbot) — carries 128GB of memory per unit. Micron's framing: humanoids ≈ 10× a L2+ vehicle's ~300GB at maturity, i.e. multi-hundred-GB per robot. Near-term reality check: 13,000 robots × ~100GB ≈ 1.3 petabytes — trivial next to a single AI data center. The story isn't 2026 volume; it's the multi-decade S-curve if humanoids reach the millions-to-billions the bulls project.
Humanoid shipments — history & forecast divergence
Annual unit shipments (log scale). ● = anchored actuals (2023–2025, Omdia/company). Beyond 2025 the lines fan out by forecaster: Goldman baseline (cautious), Morgan Stanley (China, aggressive), bull case (market-priced). The spread is the message — nobody knows the slope, only the direction.
The maker field (June 2026)

Provenance & the honesty line. Shipments [anchored]: ~13,000 humanoids shipped worldwide in 2025, Chinese firms in the top five, Figure 7th, Tesla 9th (Omdia via CNBC); Unitree the volume leader at ~5,500 units 2025; Tesla built only "hundreds" (missed its 5,000 target, public sales not before end-2027 per Musk); UBTech ~1.4B yuan in 2025 orders; Figure 03 at ~1 robot/hour at BotQ. Forecasts [cited — and they diverge hugely]: Goldman baseline 76,000 units by 2027 (vs ~500k some market participants price in) and $38B market / ~1.4M units by 2035; Morgan Stanley China shipments 50,000 (2026) → 446,000 (2030, $15B China market), and globally 1B+ humanoids by 2050 / $5T ecosystem; the spread between forecasters (Goldman conservative vs MS aggressive vs bull cases) is 4–10×. Memory per robot [anchored platform spec + CEO estimate]: NVIDIA Jetson Thor = 128GB memory, 2,070 FP4 TFLOPS, the dominant humanoid compute (Agility, Boston Dynamics, Figure, Galbot adopters; TrendForce/NVIDIA); Jetson T4000 = 64GB; Micron CEO Sanjay Mehrotra (Jun 2026): humanoids will need ~10× the memory of L2+ vehicles, which exceed 300GB DRAM/vehicle by 2026 — a "multi-decade demand cycle" (cryptobriefing/Micron). The honest caveats: (1) 2025 is the only solidly-anchored year; every forward point is a forecast and forecasters disagree by orders of magnitude — the chart shows three divergent scenarios rather than one false-precision line. (2) Near-term robot memory demand is negligible vs AI data centers — this is a long-duration option, not a 2026 catalyst. (3) Per-robot memory varies by tier (entry robots use far less than a Thor-class brain); the 128GB figure is the high-end platform, not an average. (4) The reference image's valuations/associations are point-in-time and some (e.g. 1X's $10bn target) are explicitly unverified. Modeled: the forecast curves between anchored points, and the scenario fan beyond 2025. Sources: Omdia (via CNBC), Goldman Sachs, Morgan Stanley, TrendForce, NVIDIA, Micron, The Product Compass / reference image, company disclosures. Not investment advice.

07·CHigh Bandwidth Flash (HBF) — and What It Is Not◷ spec + roadmap · Aug 1 2026
07·C

High Bandwidth Flash (HBF) — and What It Is Not

HBF is NAND flash, stacked and wired like HBM. Sixteen 3D NAND dies plus a base logic die, joined by through-silicon vias, in a package that deliberately matches HBM4's footprint, power profile and stack height — so it can sit on the same interposer next to a GPU. Generation 1 is 512 GB per stack at 1.6 TB/s read; the roadmap goes to 1 TB at >2 TB/s and 1.5 TB at 3.2 TB/s. The pitch, in Rambus' framing, is 8–16× the capacity of HBM at the same read bandwidth and roughly the same price point. SanDisk originated it, SK hynix joined in Aug 2025, and the two opened an Open Compute Project standardisation workstream on Feb 25 2026. First samples are due H2 CY2026, first inference devices early 2027.

Read this before the comparison — "HBF vs QLC" is a category error
The question this section was asked — how does HBF compare to MLC, TLC and QLC — has a precise answer, and the precise answer is that they are not on the same axis. SLC / MLC / TLC / QLC / PLC describe how many bits are stored inside a single flash cell. That is a decision made in the cell, about voltage levels on a floating gate, and it trades cost-per-bit against endurance. HBF describes how finished dies are stacked and connected to a processor. That is a decision made in the package, about through-silicon vias and interface width, and it trades packaging complexity against bandwidth and distance. You do not choose between them — every HBF stack is built from dies that are themselves TLC or QLC. HBF Gen1 stacks BiCS8 dies, the same family SanDisk ships as mainstream TLC SSDs and as UltraQLC capacity drives. So the useful question is not "HBF or QLC" but "which cell type goes inside the HBF stack, and what does that choice cost in endurance and write speed?" Both axes are laid out below: first where HBF sits against the memories it competes with, then the cell ladder it is built on.
Where HBF actually sits — capacity vs read bandwidth, per stack
Both axes are log scale, because the range spans three orders of magnitude. The interesting property of this chart is the empty space HBF is aiming at: HBM occupies the top-left (fast, small), enterprise SSDs the bottom-right (vast, slow, and off-package at the far end of a PCIe link), and almost nothing has occupied the top-right. HBF's claim is that it can reach HBM-class read bandwidth at roughly an order of magnitude more capacity, in the same physical envelope. What this chart cannot show is the catch — it plots read bandwidth only, and read is the direction flash is good at. Hover any point for the detail.
The other axis — bits per cell, and what each one costs
This is the ladder HBF is built on, not compared against. A flash cell stores charge; the controller reads back a voltage and decides which level it represents. One bit needs two distinguishable states. Two bits need four. Three need eight. Four need sixteen. The physical voltage window does not get any wider as you add bits — it just gets divided more finely, so the margins shrink, programming takes longer and more retries, reads need heavier error correction, and the cell tolerates fewer write cycles before the margins close entirely.
The two axes side by side
The constraint that decides what HBF can ever be used for
Flash reads can be made fast. Flash writes cannot. Writing to NAND means pushing charge through an oxide barrier and verifying it landed in the right voltage band — a physical process that no amount of interface engineering removes. As Objective Analysis' Jim Handy puts it, most of what flash gives up is in the write cycle; for reads, it can be faster. That single asymmetry determines the entire addressable market for HBF: it is aimed at inference, not training. Inference weights are static — loaded once, read billions of times, never modified. Training weights change on every step, which is exactly the workload flash is worst at. HBF therefore cannot replace HBM; the system still needs writable memory at speed, and SK hynix's own concept design says as much — its "H3" architecture pairs eight HBM3E stacks with eight HBF stacks on a Blackwell-class GPU, claiming up to 2.69× better performance per watt, rather than substituting one for the other. Endurance is the open question. Nobody has published an HBF endurance rating; Siemens EDA's Jongsin Yun put NAND generally "in the level of thousands" of writes. If the stack uses TLC that implies roughly 1,000–3,000 cycles, and QLC would be lower still — tolerable for weights written once per model deployment, disqualifying for anything that churns. Why it matters to this board: if HBF works, it attacks the memory wall in §06·D·W from the capacity side rather than the bandwidth side, and it makes NAND — not just DRAM — a direct beneficiary of inference scaling (§07, §06·D). If it slips, HBM stays the only game and the DRAM squeeze in §05·B tightens further.

Provenance, a corrected unit, and what is genuinely unknown. [cited — vendor spec]: SanDisk/SK hynix HBF Gen1 — 16 dies + base logic die, 256 Gb per die, 512 GB per stack, 1.6 TB/s read, matching HBM4 footprint, power profile and stack height; Gen2 >2 TB/s and up to 1 TB; Gen3 3.2 TB/s and up to 1.5 TB; samples H2 CY2026, inference devices early 2027 (SanDisk press releases Aug 6 2025 and Feb 25 2026; SK hynix newsroom; Tom's Hardware; TechPowerUp). OCP standardisation workstream announced Feb 25 2026. The 8–16× capacity claim is Rambus' Steven Woo, via Semiconductor Engineering, and this board checked it: 512 GB against a 64 GB HBM4 stack is 8×, against a 36 GB HBM3E 12-high stack is 14× — the claim is arithmetically consistent. A unit error in one source, corrected here: Semiconductor Engineering states "HBF capacity is 256 GB per die, which gives 512 GB per 16-high stack." Those two figures contradict each other — 256 GB × 16 is 4 TB, not 512 GB. The die is 256 gigaBIT (32 GB), which multiplies correctly to 512 GB. SanDisk's own materials are self-consistent; this board uses the bit figure and flags the discrepancy rather than silently picking one. Tom's Hardware's "4 TB of VRAM" headline is a GPU-level total — eight 512 GB stacks — and is consistent. [modeled / illustrative]: the comparison points on the scatter (HBM3E 24 GB and 36 GB at ~1.2 TB/s, HBM4 at 64 GB / ~2 TB/s, DDR5 DIMM 96 GB / 51 GB/s, enterprise NVMe 30 TB / 14 GB/s) are representative configurations for scale, not quotes from a single price list — per-stack HBM capacity varies with die count and vendor. Cell endurance figures are ranges, not specifications: rated P/E cycles vary by process node, controller, over-provisioning and whether the drive runs pseudo-SLC caching, and published ranges disagree — SLC 50–100K, MLC 3–10K, TLC 1–3K, QLC 0.1–1K are the commonly cited bands and are shown as bands for that reason. What is genuinely unknown: (1) HBF's endurance rating has never been published — no number here is a specification, and the TLC/QLC inference is this board's, from the BiCS8 die family, not a vendor statement. (2) Nor has its write bandwidth, which is the figure that would actually bound its use. (3) No silicon has shipped. Every performance number above is a target on a roadmap, from a company with an obvious interest in the number being large; the first samples are still months away at the time of writing. Treat this section as an explanation of an architecture, not as evidence that it will work. Cross-refs: §01·B (memory type field guide), §05·B (HBM's bite of DRAM supply), §06·D·W (the memory wall), §07 (the storage layer). Sources: SanDisk newsroom, SK hynix newsroom, Open Compute Project, Semiconductor Engineering, Tom's Hardware, TechPowerUp, VideoCardz, TrendForce, Kingston and Lexar technical notes on NAND cell types. Not investment advice.

08Demand vs. Supply◷ Jun 2026
08

Demand vs. Supply

Where HBM Demand Comes From · 2026
AI / ML55%
55%AI / ML training & inference
25%High-performance computing
12%Graphics
8%Emerging (auto, edge AI, 6G)

Source: PatSnap / JEDEC-tracked analysis, 2026

The Gap That Won't Close
HBM demand growth
+70%
Year-on-year, 2026 (after +130% in 2025)
DRAM supply growth
+16%
Below the 20–30% historical norm (IDC)

Supply can't keep pace; NAND growth is similarly capped near 17%. Downstream: Nvidia has reportedly cut gaming GPU output 30–40%, smartphone prices face 8–10% rises, and consumer hard drives are up ~50% — the AI buildout reaching into every device.

When Does Relief Arrive?
2025
Demand inflects. HBM up ~130% YoY; the supercycle becomes operational reality.
2026
Peak scarcity. Memory and HDD capacity sold out; prices spike across the stack.
2027
First new fabs begin to come online — but too late to ease 2026.
2028
Capacity ramps; HBM TAM projected near $100B. Possible rebalancing.
08·BThe Apple Mac Symptom◷ live feed · compiled to Aug 17 2026
08·B

The Apple Mac Symptom

● LIVE

The clearest place the memory shortage shows up for ordinary buyers: Apple Mac delivery wait times. Because Macs use unified memory soldered on at build time, a DRAM squeeze hits them directly — and the wait scales with how much RAM you configure. This is the abstract supercycle made tangible. Through July this section said high-memory Macs waited months "while the base model still ships same-day" — as of Aug 2 that is no longer true, with the MacBook Air itself in a reported "major" shortage at 2–6 weeks despite a $200 price rise. The queue has spread down the range, not just up the configuration ladder. As of mid-2026 the symptom has mutated — Apple has begun deleting the highest-memory configurations rather than quoting a wait for them, and on Jun 25 passed the cost to buyers with the largest mid-cycle price rise in recent memory. Read the wait-time chart together with the withdrawal ladder and the price rises below; on its own it now understates the squeeze.

Through 2026, Apple's online-store shipping estimates blew out specifically on high-memory configurations, while base models stayed quick. Reporting (MacRumors, Tom's Hardware, TrendForce) tied it directly to the AI memory crunch this dashboard tracks: Samsung and SK Hynix shifted capacity toward high-margin HBM/server memory, starving consumer DRAM. Apple removed the Mac Studio's 512 GB option entirely in March 2026, and demand from people building local-AI Mac clusters (the "OpenClaw" ordering frenzy) made it worse. Wait time has effectively become a real-time memory-scarcity gauge.

96 GB only
Mac Studio M3 Ultra — 256 GB & 512 GB both withdrawn (Mar 6 / May 5)
+15–20%
Mac price rises, Jun 25 · iPads +15–25%
Memory & storage cost in 3 quarters (Counterpoint)
−6.1%
AAPL on the hike, Jun 25 — $263B erased, worst day since Apr 2025
−120bp
FQ3 gross margin ex-tariff (49.3%→48.1%) — Apple: >100% of it is memory. ≈$1.3B in one quarter
2–6 wk
MacBook Air delivery, Aug 2 — "more constrained than they can ever recall" after a $200 rise
Jun–Jul 2026 — the symptom changed form: Apple stopped queueing and started repricing
Through spring the shortage showed up as wait time. It now shows up two other ways, and both are worse signals. (1) Configurations are being withdrawn, not queued. Apple pulled the Mac Studio's 512 GB option on Mar 6 (raising the 256 GB upgrade $1,600→$2,000), then pulled 256 GB on May 596 GB is now the only M3 Ultra memory configuration that exists, and the Mac mini lost its 256 GB storage tier. A withdrawn SKU has no lead time, so the wait-time gauge mechanically improves as scarcity worsens — the survivorship trap flagged on the chart below. (2) The cost moved to the sticker. On Jun 25 Apple raised prices mid-cycle across Macs, iPads, Home and Vision Pro — $100–$300 per device, ~+15–20% on Macs and +15–25% on iPads: MacBook Air $1,099→$1,299, MacBook Pro $1,699→$1,999, iPad Air $599→$749, after the Mac mini's base went $599→$799 (May 1) and the M4 Pro mini rose $200 to $1,599 (Apr). Apple's own words: "We have never seen a component price increase this much, this quickly." Tim Cook called it a "hundred-year flood" he hadn't seen in 40 years (WSJ, Jun 17), with memory and storage up ~4× in three quarters (Counterpoint). The market read it as margin risk: AAPL fell 6.12% to $275.15 on Jun 25 — its worst session since April 2025, ~$263B of market cap. This is kill-switch #5 (demand destruction) arriving at the most consumer-visible company on the board — the first time this cycle's memory cost has been passed to end buyers at scale.
Jul 30 – Aug 2 2026 — the price rise did not clear the queue, and this section's own framing is now wrong
Two things happened in four days, and together they break the reading this section adopted in June. First, Apple put a number on it. FQ3 (Jul 30) was a record June quarter — revenue $109.4B, +16%, iPhone +22%, net income $29.8B — and the gross margin fell anyway. Reported 50.1% flattered by roughly 2 points of tariff refunds; ex-tariff 48.1%, down 120bp from March's 49.3%. Apple's own attribution on the call: "more than 100% of the gross margin decline can be explained by the memory cost change" — memory cost more than the entire decline, and cheaper non-memory components absorbed part of it. On $109.4B of revenue, 120bp is roughly $1.3 billion of gross profit in one quarter. Guidance for September is 47–48%, memory again the primary driver, with Apple expecting to pay more, cushioned only by carry-in inventory. Cook, on the call: "we're in what I would characterize as a 100-year flood on the memory pricing, with exponential increases." Second — and this is the part that falsifies this section's June framing — the shortage reached the cheapest mainstream Mac. This section previously said a buyer configuring high memory waits months "while the base model still ships same-day." That is no longer true. On Aug 2 Bloomberg's Gurman reported the MacBook Air in a "major" shortage, with retail sources saying shipments are "more constrained than they can ever recall" and a 2–6 week US delivery estimate — higher-RAM configurations still worst, but the base machine is now in the queue too. Apple raised the Air $200 in June and still cannot supply it in August. So the June conclusion — that Apple had "stopped queueing and started repricing" — was half right and is now superseded. It is doing both at once, and the queue has spread down the range rather than staying at the top of the configuration ladder. Price did not clear demand. That matters for kill-switch #5: demand destruction requires buyers to actually stop buying, and a sold-out product at a 18% higher price is the opposite signal. The thing to watch is no longer the wait on a 256 GB Studio — it is whether the Air's queue persists into the September quarter, when Apple has already told us memory costs rise again.
Apple gross margin — where the memory cost actually lands
The solid line is gross margin on a comparable ex-tariff basis; the hollow ring is the 50.1% Apple reported, which included about two points of tariff refunds that have nothing to do with operations. Plotting the reported number would have shown the margin going up in the quarter Apple said memory was crushing it. The dashed segment is guidance, drawn as a band because Apple gave a range. Hover any point for Apple's own wording.
One number in the guidance that cannot be read cleanly
Apple guided September gross margin to 47–48%. Against what? Measured against the reported 50.1%, that is a fall of 210–310bp and reads as a collapse. Measured against the ex-tariff 48.1%, it is 10–110bp and reads as a mild continuation. The two readings differ by about 200 basis points, and Apple did not state which basis the guide is on — specifically, whether it assumes tariff refunds repeat. This board uses the ex-tariff comparison on the chart above, because it is the only one that isolates the thing this section is about. But the ambiguity is real and is not resolved by anything published, so the more dramatic framing — "Apple guides margins down 300bp on memory" — is available to anyone who wants it and is not obviously wrong. It is simply not the comparison this board can defend.
In stock / days ~2–4 weeks ~5–10 weeks 12+ weeks pulled from sale
Average Mac Wait Time by Quarter — past 2 years

Modeled average delivery estimate across a representative basket of Mac configurations (base through high-memory), in weeks. The shape is the story: waits were normal (days) right through 2024 and most of 2025 — then the AI memory shortage hit in early 2026 and they spiked. This isn't a chronic Apple problem; it's a 2026 memory-supply event.

Current quarter · Q2 2026 · avg ~7 wk across the basket
High-memory configs at 16–18 weeks; base models still ship in days. Checks for the latest estimates on load.
Blended average across base + mid + high-memory configs, in weeks. Orange dots mark quarters anchored to specific shipping-estimate reports; the current quarter updates live when run inside Claude. Read Q3 2026 with care — the dip is survivorship, not relief: the configurations that carried the longest waits (M3 Ultra 256 GB, 512 GB) were withdrawn from sale, so they leave the basket entirely and the average falls while supply gets tighter. Apple also shifted the adjustment from queue to price on Jun 25 (+15–20% on Macs). The Q2 2026 anchor remains the cleanest RAM gradient on record (MacRumors, Apr 6; store checks through Jun): Mac mini M4 16 GB ≈ 1 month · mini M4 Pro 48 GB ≈ 10–12 wk · 64 GB ≈ 16–18 wk · Mac Studio M3 Ultra 256 GB ≈ 4–5 months (in-store pickup pushed to September) — and the 512 GB RAM option was removed from sale entirely, the closest thing to an infinite wait time. Wait scales with GB: that is this chart’s thesis, printed on Apple’s own store. Apple’s response was price, not queue: the mini M4 Pro start rose +$200 ($1,399→$1,599, Apr), then $100–500 across the Mac line in late June (§08·C) — kill-switch #5’s mechanism, again. Interpretation caveat, disclosed: the Jan MacBook Pro waits (M4 Max 36–128 GB at ~3–7 wk) were ambiguous — shortage vs. the imminent M5 Pro/Max refresh, which shipped late Feb — but the Apr mini/Studio waits carry no refresh excuse and track RAM capacity monotonically. Bloomberg adds a roadmap casualty: Apple accelerated an entry M6 while canceling M6 Pro/Max outright (next Pro silicon jumps to M7).
Normal (≤1 wk) Mild (1–3 wk) Elevated (3–6 wk) Severe (6+ wk) Anchored to a report

Wait times are configuration-specific snapshots from dated press reports, not a live per-model feed — Apple doesn't publish lead-time data, and its store estimates change daily and vary by region and channel. Anchors (US online store unless noted): Mac mini M4 Pro/64 GB = 16–18 weeks; base M4/16 GB ≈ 4 weeks; 512 GB M4 backordered into June (MacRumors, TrendForce, Apr–May 2026). Mac Studio M3 Ultra/256 GB = 16–20 weeks (4–5 months); 512 GB option removed from sale Mar 2026; in-store pickup pushed to September. MacBook Pro M4 Max/128 GB ≈ 5 weeks (early-2026 reporting), 36–48 GB a few weeks, base configs days. MacBook Air / iMac base ≈ same-day. Bars show the approximate midpoint of each reported window, scaled to ~20 weeks. The quarterly history is a modeled basket average, not a published series: 2024–2025 quarters reflect Apple's normal few-day lead times (no shortage reporting existed then); Q1 2026 is anchored to the Feb 2026 Mac Studio "14→54 day" reports (Tom's Hardware), and Q2 2026 to the Apr–May 2026 Mac mini/Studio blowout (MacRumors/TrendForce). The current quarter rechecks for the latest estimates on load when run inside Claude; opened as a file it shows the compiled value. The drivers — AI capacity diverting DRAM, plus local-AI Mac-cluster demand — are well documented; exact weeks for any given config today will differ. Aug 17 2026 update [cited]: Apple FQ3 2026 (reported Jul 30): revenue $109.4B, +16% YoY, a June-quarter record; net income $29.8B; iPhone revenue +22%, also a record; gross margin 50.1% reported, including approximately 2 percentage points of tariff refunds, giving 48.1% ex-tariffdown 120bp sequentially from March's 49.3%. Apple stated on the call that "more than 100% of the gross margin decline can be explained by the memory cost change", and Cook described "a 100-year flood on the memory pricing, with exponential increases." September-quarter gross margin guided to 47–48%, memory named the primary driver, with memory costs expected to rise further and be partly offset by carry-in inventory and cheaper non-memory components. Shares fell after hours despite the beat. [derived — this board's arithmetic]: 120bp on $109.4B of revenue is ≈$1.31B of gross profit in a single quarter; the two-quarter ex-tariff path 49.3% → 48.1% → 47.5% (guide midpoint) is −180bp. Basis ambiguity, flagged not resolved: Apple did not state whether the 47–48% guide assumes tariff refunds repeat, so the September fall is 210–310bp against the reported 50.1% or 10–110bp against ex-tariff 48.1% — a ~200bp difference in interpretation. The chart uses ex-tariff and says so. MacBook Air shortage (Aug 2, Bloomberg/Gurman via MacRumors): a "major" shortage with retail sources saying shipments are "more constrained than they can ever recall"; US online-store delivery 2–6 weeks, higher-RAM configurations delayed longest; the Air is now affected alongside Mac Studio and Mac mini. Apple raised prices on 14 products in June, the Air among them ($1,099→$1,299). Framing correction carried into the section body: this section previously asserted that base models "still ship same-day". That was accurate when written and is not now; the claim has been retracted in the summary rather than quietly deleted, and the reason is dated. Context [cited]: Gurman also reports Apple preparing to refresh every Mac it sells over two years — 11 new models — with Mac sales projected to rise for a third consecutive year, partly on demand from users running agent-based AI workloads locally; the MacBook Neo, launched as "the $599 Mac", is now $699. Not carried: no August wait-time figures for Mac Studio or Mac mini have been published since the July checks, so the chart's Q3 bar is not updated for the Air's queue — it would move the opposite way to the survivorship effect already flagged, and this board will not model the netting of two unquantified opposing effects. Jun–Jul 2026 update [cited]: Config withdrawals — Mac Studio M3 Ultra 512 GB pulled Mar 6 2026 with the 256 GB upgrade raised $1,600→$2,000 (Tom's Hardware, MacDailyNews); 256 GB pulled May 5 2026, leaving 96 GB as the only M3 Ultra memory option (MacRumors, 9to5Mac, VideoCardz), with delivery reported into October (Macworld); Mac mini 256 GB storage tier discontinued and base price $599→$799 on May 1 (MacRumors). Price rises — Apr: Mac mini M4 Pro +$200 to $1,599; Jun 25: mid-cycle increases across Mac, iPad, Home and Vision Pro of $100–$300 per device — Macs ~+15–20%, iPads ~+15–25% (MacBook Air $1,099→$1,299; MacBook Pro $1,699→$1,999; iPad Air $599→$749), Apple citing component costs it had \u201cnever seen \u2026 increase this much, this quickly\u201d; Tim Cook's \u201chundred-year flood\u201d framing (WSJ, Jun 17); memory and storage up ~4\u00d7 in three quarters (Counterpoint Research). Market reaction — AAPL \u22126.12% to $275.15 on Jun 25, ~$263B of market value, its worst session since April 2025 (CNBC, Forbes, TIKR). Honesty note on the Q3 2026 bar: it is lower than Q2 and that is largely survivorship — the longest-waiting SKUs were delisted, so they exit the basket; a withdrawn configuration has no lead time to average in. Treat Q3 as a partial quarter (to Jul 24) and read it alongside the withdrawal ladder and the price rises, which move the opposite way. Sources: MacRumors, Tom's Hardware, Tom's Guide, TrendForce, BigGo, Cybernews, 9to5Mac, VideoCardz, Macworld, MacDailyNews, CNBC, Forbes, Counterpoint Research, WSJ. Not investment advice.

08·CMemory Inside Apple Products◷ model
08·C

Memory Inside Apple Products

◷ model

A different lens on the same demand: how much memory ships inside Apple devices sold each year, by type, 2010 → 2030. Apple alone consumes a meaningful slice of world memory, and the total has exploded — not mainly because Apple sells more units (iPhone volumes plateaued years ago), but because each device carries far more memory: storage has gone from 16 GB to 256 GB–1 TB, and RAM from 512 MB to 8–12 GB as on-device AI (Apple Intelligence) pushes it up. NAND storage dwarfs DRAM, so it's shown on its own axis context in the tooltip. Update (Jun–Jul 2026): the model's demand-side thesis just went from inference to confirmed fact — Tim Cook told the WSJ that memory costs have become unsustainable, called the squeeze a “hundred-year flood,” and confirmed iPhone price increases; Apple already raised Mac prices $100–500 (and iPad/Apple TV/HomePod) in late June. And in the middle of it: Cook is outgoing — hardware chief John Ternus becomes CEO September 1, a week before the iPhone 18 launch.

Estimated memory shipped inside Apple products per year, in exabytes (EB = billion GB). Solid = historical estimate; hatched = projection. Stacked by type: NAND (storage) and DRAM (working memory). Hover any year for the split and the unit/-capacity assumptions.

This is a two-layer model, not reported data — Apple discloses neither unit shipments (it stopped in 2018) nor memory content. It multiplies estimated annual unit shipments per product line (iPhone, iPad, Mac, Watch, others) by an estimated capacity-weighted average of NAND storage and DRAM per device, each rising over time. Anchors: iPhone ~232M units (2024), record ~247M (2025, IDC); Macs ~22–25M/yr (~6.2M in Q2 2025, IDC/Canalys); iPad ~50M/yr. Capacity assumptions: iPhone average storage ~16 GB (2010) → ~256 GB (2026E), DRAM 0.5 GB → ~8 GB; Macs/iPads scaled higher. NAND outweighs DRAM by ~20–40× per device, so it dominates the stack. Projection to 2030 assumes low-single-digit unit growth and continued capacity step-ups (on-device AI lifting the DRAM floor — Apple Intelligence needs ≥8 GB) — IDC notes 2026 smartphone units actually dip ~1% on the memory shortage itself, even as ASPs hit records. Jul 2026 refresh — the cost shock, quantified [cited]: the TechInsights model (via WSJ) has the same 12 GB DRAM package going from about $39 in iPhone 17 Pro to about $145 in iPhone 18 Pro (+272%), with the 256 GB base NAND tier from ~$13 to a projected $51; estimated build cost rises from ~$582 to ~$726, implying ~$1,371 to hold margin on a $1,099 phone — i.e. roughly +$270 on the 18 Pro to preserve margins. Base-model hikes of $50–150 and up to $200 on Pro are the projected range; DRAM/NAND contract prices jumped 90–95% QoQ in Q1 2026. Procurement color: Samsung reportedly planned +60% on LPDDR5X, opened at +100% — and Apple accepted immediately in emergency procurement meetings. The DRAM-per-device floor is rising exactly as modeled: the iPhone 18 line carries 12 GB to power on-device Apple Intelligence, and DigiTimes reports even the base iPhone 18 moves 8→12 GB. On duration, the disagreement is surfaced, not resolved: Cook expects memory pricing to “return to reasonable levels,” while IDC calls the shift a “permanent reallocation” of capacity toward AI, and analyst timelines for relief run from Q4 2027 (Counterpoint) to 2028 (Intel’s Tan) to 2030 (Kearney); Cook even said “everything needs to be on the table” when asked about sourcing from Chinese memory suppliers. This section’s cross-links write themselves: Apple raising retail prices on memory costs is kill-switch #5’s demand-destruction mechanism operating in the wild (§02·D), and the affordability ceiling TrendForce cited for the 3Q26 contract-price deceleration (§03) is the same phenomenon from the supplier side. Treat the totals as order-of-magnitude with a wide band; the trend (memory-per-device, not unit count, drives the growth) is the robust takeaway. Sources: IDC, Canalys, Counterpoint, Apple teardowns (iFixit/TechInsights), company history. Not investment advice.

08·DAI App-Layer Revenue — Run-Rate: Anthropic × OpenAI × Facebook-at-the-same-age◷ disclosed run-rates · derived monthly · + BEA wage share
08·D

AI App-Layer Revenue — Run-Rate

The demand under the demand. All the memory in this dashboard is ultimately paid for by the revenue the AI application layer throws off — so here is that layer's income, built from every annualized run-rate the two leading labs have disclosed, with Facebook's first revenue years overlaid by company age as a benchmark. This section previously said the two labs had diverged, with OpenAI "flattened" near $24–25B. That was wrong, and the correction is the story. On Aug 13 2026 Bloomberg reported OpenAI's run rate above $40B — roughly double its end-2025 level, +20% month-over-month in July alone, driven by Codex, subscriptions and a new advertising business. It did not flatten; the prior build read a two-month pause as a structural ceiling and projected it forward. Logged in §10·B. Anthropic ran too: its last official figure is still $47B (Series H, May 28), but third-party trackers put it near $70B by late July. So both accelerated, and the gap between them narrowed — from roughly 2.0× at the old anchors to about 1.75× now. Against Facebook at the same age the comparison stays absurd: two-plus years into earning revenue Facebook made ~$48M/year, and these two are three orders of magnitude larger.

Implied monthly revenue = run-rate ÷ 12. Solid dots are reported/actual figures; between them, values follow exponential (constant-growth) interpolation; the dashed tip projects the current month. Facebook is time-shifted 20 years — the chart's 2024 is anchored to Facebook's 2004 (its founding year and first revenue, $382K), so 2025≙2005, 2026≙2006.
Four provenance tiers now, because three were not enough to stop the error above
● filled dot = the company said it. Anthropic's $47B (May 28), OpenAI's ~$2B/month (June). ◌ dashed ring = somebody else measured it. New in this build: OpenAI's $40B is Bloomberg citing people familiar, with OpenAI declining to comment; Anthropic's ~$70B is Yipit and TickerTrends, not Anthropic. faint line = interpolated between two of the above. dashed line = this board projected it — and that last tier is the one that just failed, so it now appears nowhere on either lab's curve. Why the distinction earned its own marker. The prior build drew a projection and a disclosure in visually similar ways, and a flat projection quietly became a claim about the world. A third-party estimate is not a company disclosure — trackers have historically run hot, which is why Anthropic sits at $70B here while its own last stated number is $47B, a 49% gap between the two most authoritative figures available. Both are printed; neither is blended. One internal inconsistency, disclosed: Bloomberg gives both "+20% month-over-month in July" and "tops $40B" in August. Applied to June's $24B, the first implies ~$28.8B for July; the curve here interpolates ~$31B between the June and August anchors. The two cited facts do not perfectly reconcile, and the ~8% gap is left visible rather than smoothed.
Run-rate detail — disclosed points & sources
AI app-layer revenue as a share of the US wage bill WHOLE LLM INDUSTRY · BEA DENOMINATOR
The run-rate chart above answers "how fast is this growing?" This one answers "how big is it, really?" — by measuring AI app-layer revenue against the thing it is most often claimed to displace: wages. The denominator is the BEA's Compensation of Employees: Wage and Salary Disbursements (FRED A576RC1) — the total US wage-and-salary bill, published monthly at a seasonally adjusted annual rate, which makes it directly comparable to an annualized revenue run-rate. Both numbers are annual-rate dollars, so the ratio is apples-to-apples in units. Aug 1 update — the numerator is now the WHOLE LLM industry, not two firms. The prior build measured only Anthropic + OpenAI and flagged that as a floor. It now counts every company whose product is a large language model: those two plus Google’s Gemini, xAI, Mistral, Cohere, Meta, ByteDance/Doubao, Moonshot, DeepSeek and the rest of the Chinese field — $94B of model-layer run-rate, itemised in the table below. The result is the interesting part: it barely moves the answer. The share goes from 0.55% to 0.70% of the US wage bill, because Anthropic and OpenAI alone are 79% of the entire model layer. The amber band on the chart is every other lab on earth. Read it as a scale reference, not a displacement rate — the caveats below are load-bearing, and the biggest one is that the numerator is global company revenue while the denominator is US-only wages. The dashed blue line is a deliberate analogy: e-commerce as a share of US retail sales (Census/FRED ECOMPCTSA), age-aligned so that AI's January 2024 sits on e-commerce's Q1 2000 — the last comparable case of a new channel eating an established aggregate. It is a different numerator over a different denominator; it is here for shape and pace, not arithmetic comparison.
WHOLE LLM industry run-rate ÷ US wage billAnthropic + OpenAI only (79% of the layer)● both firms' figures disclosed that month— — e-commerce % of US retail sales, age-aligned from 2000 (Census/FRED)wage denominator extrapolated (BEA last print May 2026)
The numerator, itemised — every LLM lab with measurable model revenue
Scope: revenue earned by firms whose product is a large language model, from that model — subscriptions, API and enterprise contracts, annualized run-rate as of Jul 2026. Each row is tagged by how solid the figure is: disclosed (the company said it), reported (credible third party), estimate (a published range), modeled (this board’s construction from indirect evidence — treat with the most caution). Google is the weakest link and the largest of the uncertain rows: Alphabet publishes no Gemini revenue line at all, so its $12B is modeled from its 21% enterprise-API usage share, its disclosed $1.2B of 2025 subscriptions and ~800% YoY growth in genAI-model products — a range of $8–16B is defensible and the total moves ±$4B with it.
Deliberately excluded — the resale channel, and why summing press releases inflates this market
These are real revenues. They are also, largely, the same dollars counted a second time at a different point in the chain — so adding them to the roster above would not measure a bigger industry, it would measure the same industry twice.
The two readings, both true. Small: after the fastest revenue ramp in software history, the entire LLM industry — every lab, worldwide — collects about seven tenths of one percent of the US wage bill: roughly $6.98 for every $1,000 of wages paid. Anyone arguing that AI has already replaced labour at scale has to explain that number, and it is no longer answerable by saying the sample was too narrow. Fast: the same ratio was $0.21 per $1,000 in January 2024 — a ~33× rise in 31 months — and the industry has captured roughly 6% of the increase in the US wage bill over that span. A metric can be trivially small in level and violently fast in slope at the same time; this chart is deliberately drawn on a linear axis so you see both, rather than a log axis that would flatter the growth.

Provenance & the honesty line — read before quoting this ratio. Denominator [cited — official]: U.S. Bureau of Economic Analysis, Compensation of Employees, Received: Wage and Salary Disbursements (A576RC1), retrieved from FRED, Federal Reserve Bank of St. Louis — billions of dollars, seasonally adjusted annual rate, monthly; series runs Jan 1959→May 2026, last updated Jun 25 2026. Values used run Jan 2024 ($12,062.7B) → May 2026 ($13,388.8B). Jun and Jul 2026 wages are extrapolated at +$48.2B/month (the trailing six-month average increment) because BEA had not published them at build time — those two months are shaded on the chart and flagged in the tooltip. Numerator [rebuilt Aug 1 2026 — whole industry]: two series are drawn. The amber-brown line is Anthropic + OpenAI from the disclosed anchors driving the chart above; the amber line adds every other LLM lab, itemised in the roster table with a per-row confidence tag. Both are interpolated between anchors at constant growth (exponential, not linear) and projected at the tip; ● marks months where both of the two leaders had a disclosed figure. The rest-of-industry series is the softest thing on this chart and is labelled so: its terminal point is pinned to the roster sum ($20.2B), but its history back to Jan 2024 is modeled, not compiled from monthly disclosures — most of these firms have never published a run-rate time series, and several (DeepSeek, Zhipu, MiniMax) publish nothing at all. Sanity checks used: Menlo Ventures put enterprise LLM API spend at $3.5B (late 2024) → $8.4B (mid-2025) and total enterprise genAI spend at $37B for 2025 with $18B of that infrastructure; those are different scopes than this roster but bracket it plausibly. Do not quote an intermediate month of the industry line as a cited figure — quote the endpoint, and quote it as ~$94B ±$5B. Four caveats that limit what this ratio can mean: (1) Geography mismatch — the big one. Anthropic and OpenAI revenue is worldwide; A576RC1 is US wages only. The ratio therefore overstates any true "share of US wages," and is best read as a scale reference against a familiar macro aggregate rather than a share of a matched market. (2) Industry coverage — this caveat was addressed on Aug 1 2026, and the answer was undramatic. Earlier builds counted only Anthropic + OpenAI and warned the true figure was "meaningfully higher." The numerator now covers the whole model layer (roster above): the share rose from 0.55% to 0.70% — real, but not the order-of-magnitude gap the old caveat implied, because the two firms are 79% of the layer. What remains excluded, and deliberately: (a) the resale channel — Microsoft’s $37B AI run-rate, Perplexity, AWS Bedrock, Google Vertex — because those dollars are already counted in the labs’ books or are the labs’ own models passed through; (b) vertical AI applications built on top of models (Cursor, Harvey, Abridge and the rest), which are a genuinely additional layer this chart does not attempt — Menlo Ventures sized total enterprise genAI spend at $37B for 2025, of which applications were $19B, so an application-inclusive numerator would be a different and larger measurement; (c) ad-monetized model use, chiefly Meta’s, which earns real money from Llama without ever selling a token. (3) Run-rate ≠ realized revenue. A run-rate annualizes an instantaneous rate; the BEA series is an annualized flow of actually-paid wages. The units match, the epistemics don't quite. (4) Revenue is not displacement. AI subscriptions are bought by employers and employees, often alongside labour rather than instead of it; nothing here identifies substitution. A rising ratio is consistent with augmentation, substitution, or simply a new category of software spend. The e-commerce overlay [cited — added Jul 30 2026]: U.S. Census Bureau, E-Commerce Retail Sales as a Percent of Total Sales (ECOMPCTSA), retrieved from FRED — quarterly, seasonally adjusted, series begins 1999-Q4. Alignment: AI month 0 (Jan 2024) is placed on e-commerce 2000-Q1, so both curves are read at the same elapsed months since commercial take-off; quarterly prints fall on every third month and the months between are linear interpolation. What it shows: 31 months in, e-commerce was 1.5% of US retail sales while the whole LLM industry is 0.70% of the US wage bill — the analogue was ~2.1× larger at the same age (it was ~2.7× against the older two-firm numerator), which is a useful corrective to \u201cAI adoption is unprecedented\u201d framing. But AI grew ~32\u00d7 over the window against e-commerce's ~1.9\u00d7, roughly 17\u00d7 the pace. And the long tail is the real lesson: e-commerce took 26 years to go from 0.8% to 16.9% (2026-Q1) \u2014 S-curves of this kind are measured in decades, not quarters. Comparison caveat, stated plainly: the two lines share an axis but not a unit \u2014 retail sales and wages are different aggregates, and e-commerce substitutes for a channel while AI spend may or may not substitute for labour. Treat the overlay as a pace and shape reference only; do not read a crossing as AI \u201covertaking\u201d e-commerce in any economic sense. Why it's on-thesis: §08·E tracks tokens per employee and §06·D the token economics; this puts the same question in dollars — if AI spend is to justify the memory build-out in §05–§06, it eventually has to become a visibly larger share of the compensation it is meant to leverage. Today the entire industry is about seven tenths of one percent — and that is now the complete count, not a floor. Sources: U.S. Bureau of Economic Analysis via FRED (A576RC1); U.S. Census Bureau via FRED (ECOMPCTSA); Menlo Ventures 2025 State of Generative AI in the Enterprise and 2025 Mid-Year LLM Market Update; Microsoft FY26 Q3 disclosure; Alphabet Q2 2026 results; CNBC (Cohere investor memo); Sacra; Bloomberg and FutureSearch (xAI); company disclosures for Mistral, Perplexity and Moonshot; plus the run-rate anchors cited above. Not investment advice. [refreshed Aug 14 2026]: OpenAIBloomberg, Aug 13 2026: annualized run rate tops $40B, roughly double end-2025, attributed to Codex/AI-coding, subscription sales and a new advertising business; Greg Brockman cited internally that the run rate rose more than 20% month-over-month in July; enterprise is >40% of revenue; Q1 2026 revenue was ~$6B. OpenAI declined to comment, so this is reporting on anonymous sources — a dashed-ring tier, not a disclosure. Both labs have filed confidentially to go public, with Anthropic expected to reach the market first, as soon as this autumn. Anthropic — last official run-rate remains $47B (Series H, May 28 2026, at a $965B valuation); third-party trackers put late-July ARR at $69B (Yipit) to $74.1B (TickerTrends), plotted here at $70B with the explicit caveat that these trackers run hot and that Bloomberg notes the two companies may not measure run-rate the same way — the same gross-vs-net dispute this section already documented. Roster changes: xAI revised down $3.0B → $2.0B — published estimates now span $0.5B standalone to $3.5B including X, a 7× spread, and the prior build sat at the top of it; Mistral nudged to $0.8B after its Jun 2026 $3.5B raise at a $20B valuation; Microsoft's $37B AI run-rate is unchanged because Microsoft did not restate it at Q4 FY26 — it is now the stalest figure in the double-count block, though Q4 did disclose Copilot past 30M paid seats (from 20M), Copilot revenue +60% QoQ, and Azure crossing $100B annual at +43%. Google Gemini remains the weakest line on the roster at a modeled $12B, unchanged and now the oldest estimate here — Google still publishes no Gemini revenue line. Rest-of-industry dips slightly versus the prior build solely because of the xAI revision, not because any company's revenue fell.

Provenance & the honesty line. The metric: neither lab reports monthly revenue — both periodically disclose an annualized run-rate (latest month × 12). "Implied monthly" here is simply run-rate ÷ 12; it's a forward annualization, not trailing recognized revenue, so full-year GAAP totals run lower during fast growth. Anthropic [disclosed]: $87M run-rate (Jan 2024) → $1B (Dec 2024) → $4B (Jun 2025) → $9B (Dec 2025) → $14B (Series G, Feb 12 2026) → $19B (Mar) → $30B (Apr 7 company post) → $47B (Series H, mid-May, $965B valuation); ~$48.5–50B projected for Jun–Jul (VentureBeat revenue-walk, CNBC Series H, company disclosures). OpenAI [reported/blended]: ~$2B run-rate exit-2023 → $5.5B exit-2024 → $10B (Jun 2025, Reuters) → $20B (CFO, Jan 2026) → ~$25B (Feb–Mar) → ~$24B (Jun, "≈$2B/month" by its own account), apparently plateaued mid-2026 (Reuters, OpenAI funding post, TechnologyChecker Jul 3). Facebook overlay [actual, historical]: aligned by company age, not calendar — chart-2024 = FB-2004 (founded Feb 4 2004, first revenue $382K per its IPO S-1); plotted values are actual annual revenue ÷ 12 ($382K → $9M → $48M → $153M across 2004–07), no projection (TechCrunch/S-1). Filling gaps: dots are disclosed/actual; between two dots, exponential (constant-growth) interpolation; the dashed tip projects the labs' current month — Anthropic forward from mid-May $47B (nudged for reports of cooling token spend), OpenAI held near its stated ~$2B/month. Confidence & the dispute [surfaced, not buried]: Anthropic's figures are largely official fundraising disclosures; OpenAI's are blended and vary by method — and in April 2026 OpenAI publicly argued Anthropic's ~$30B figure was overstated by ~$8B, contending Anthropic books gross end-customer spend through AWS/Google/Azure while the net figure is closer to ~$22B; some trackers (Sacra) instead put OpenAI higher, near ~$33B. Treat every run-rate headline as directional, not audited. Why it's on-thesis: this app-layer revenue is what funds the compute commitments (§06·F·C), the hyperscaler backlogs (§06·F·B), and ultimately every bit of memory demand in this dashboard — the tokens that hit the memory wall (§06·D·W) are the same tokens being monetized here. Adapted from a user-supplied run-rate tracker; figures re-verified Jul 2026. Sources: CNBC, VentureBeat, Reuters, Epoch AI, Sacra, FutureSearch, TechCrunch/Facebook S-1, OpenAI & Anthropic disclosures. Run-rates are annualized public disclosures, not audited financials — not investment advice.

08·EToken Usage per Employee — The Leaderboard◷ reported + derived
08·E

Token Usage per Employee — Leaderboard

The demand side at maximum resolution: how many tokens does one employee burn per day? 2026 made this answerable — companies built literal internal leaderboards ("Claudeonomics" at Meta, "AI God" ranks at Sendbird), then the bills arrived and the same companies started dismantling them. The spread below runs five orders of magnitude, from ≥100M tokens/day power users to the median firm’s ~$11/employee/month — token demand is a power law, exactly like the app-layer revenue (§08·D) and capex (§06·F) it feeds.

#WhoTokens / employee / day$ signalLabel
1Sendbird "AI God" engineers (5–10% of eng staff)≥100Mprizes, planned extra vacation daysreported
2Heavy agentic developers, industry-wide (incl. Microsoft E+D pre-cancellation)~20–80M$500–$2,000/mo billsderived
3Sendbird CEO (rank: "Custom Tool Builder")30M70% via Claude Codereported
4Meta — company-wide average, all ~85K employees~29M"approaching billions" in 2026derived: 73.7T ÷ 85K ÷ 30d
5Typical enterprise Claude Code developer~3–6M (active day)$13/active day · $150–250/mo · 90% under $30/dayderived from reported $
6Uber engineers (95% monthly adoption, ~70% of committed code AI-generated)capped$1,500/mo/employee limit after budget gone in 4 monthsreported
7Aerospace & defense mfr power userscap burned in 4 days$250/mo limit · Opus + fast-mode disabledreported
8Top-1% firm on Ramp (per-employee median)$7,450/employee/moreported
9Top-10% firm on Ramp$611/employee/moreported
10Median firm on Ramp$11.38/employee/moreported
11US economy-wide (Atlanta Fed, anticipated 2026)$2,068/employee/year (+50% YoY)reported

Sources, derivations & the honesty line. The two marquee numbers [cited]: Meta employees consumed 73.7 trillion tokens in roughly 30 days, tracked on an internal leaderboard called "Claudeonomics"; costs are approaching billions for 2026, and Meta is dismantling the board for a centralized "AI Gateway" with formal token budgets from 2027 — across all 85,000 ranked employees that derives to ~29M tokens per employee per day, an average that matches an AI-startup CEO’s personal burn. Sendbird ranks staff from "Beginner" to "AI God" — 100M+ tokens/day — with 5–10% of engineering at god tier, ~70% of it through Claude Code, and prizes for usage. The middle of the curve [cited]: enterprise Claude Code averages ~$13/developer/active day and $150–250/month, with 90% of users under $30 on any active day; agentic tasks consume 400K–2M tokens each, power users reach $500–$2,000/month, and per-developer consumption rose ~18.6× in nine months; Microsoft’s Experiences+Devices division ordered engineers off Claude Code by June 30 after ~$2,000/engineer/month exhausted the annual budget — moving to $39/seat Copilot, a 51× CFO math. The long tail [cited]: Ramp per-employee medians: $11.38 typical firm, $611 top decile, $7,450 top percentile — average company $140,842/mo vs median $2,246, a power-law distribution — and the Atlanta Fed anticipates $2,068 per employee for 2026, up 50%. The cultural whiplash, six months end to end: Meta, Amazon and Microsoft set internal token-usage targets in early 2026; by mid-year Amazon shut its leaderboard down ("Please don’t use AI just for the sake of using AI") and replaced counts with shipped-code metrics, Uber capped at $1,500/mo after burning its annual budget in four months, and a defense manufacturer’s $250 caps were burned through in four days; Meta’s CTO put the epitaph on tokenmaxxing: "All motion is not progress and token usage alone is not a measure of impact of any kind". Honesty notes: units are deliberately mixed — some rows disclose tokens, some only dollars, and $→token conversion varies ~100× with model choice and caching, so dollar rows are left in dollars rather than fake-converted; "derived" rows show their math; rankings compare unlike populations (one company’s elite vs another’s average). Why it’s on-thesis: these are the tokens that hit the memory wall (§06·D·W) and get monetized in §08·D — and the budget caps arriving mid-2026 are a live demand-quality read for kill-switch #5 (§02·D): Goldman still projects 24× enterprise token growth by 2030 (120 quadrillion/month), but the era of burning tokens as a KPI is visibly ending. Reported + derived snapshot, Jul 2026 — not investment advice.

08·FRetail Attention — Google Trends◷ Trends CSV · 261 weeks · Aug 1 2026
08·F

Retail Attention — Google Trends

Every other section on this board measures what the industry is doing: prices, wafers, capex, backlogs. This one measures what ordinary people are typing into a search box — five years of worldwide search interest for nvidia, micron, sk hynix, sandisk and ethereum. Rebuilt Aug 1 2026 from Google Trends’ own CSV export — 261 weekly points, replacing the earlier screen transcription of a panel whose first term was misspelled. Correcting the spelling changed the chart more than it changed the world. Nvidia now holds the 100 index point, so every other line sits on a materially lower scale than before — the denominator moved, not the data. Nvidia peaked in the week of Jan 26 2025, which is the DeepSeek week: search interest topped out on the crash, not the rally. Across all 261 weeks no memory name ever outranks nvidia, and its five-year mean of 28.5 is 2.7× ethereum’s 10.5. The memory-over-crypto crossing is real and survives the correction, but it is later than this section first claimed: micron passed ethereum for good in the week of May 3 2026 and sk hynix on Jun 21 2026 — not February. That error is logged in §10·B.

What changed when the spelling was fixed — and why every number here is smaller
The original capture searched nividiai and v transposed. Google Trends does not autocorrect; it returned the misspelling as its own term, which sat at 1 for the whole window. The real cost was not that one flat line: it was that Nvidia was absent from the index. Trends scales every series against the single highest point in the panel, so with Nvidia missing, ethereum’s Aug-2025 spike became the 100 and all four remaining terms were measured against it. With the spelling corrected, nvidia takes the 100 and ethereum’s same peak now reads 49. Every previously-plotted value is therefore roughly half what the old chart showed — nothing about search behaviour changed; the yardstick did. This is the single most misread property of Google Trends, and it is worth stating plainly: these numbers are only ever meaningful relative to the other terms in the same query. Add a bigger term and everything shrinks; remove it and everything grows. The substantive finding survived the correction — memory names really did overtake crypto in search interest this year — but the timing did not, and the magnitudes were roughly doubled. Both are corrected here and logged.
Worldwide search interest, weekly — Aug 1 2021 → Jul 26 2026 SOURCE CSV · 261 WEEKS
How to read the scale, because it is the most misread thing about Google Trends: the numbers are not search volumes. They are relative, and indexed across the whole panel at once — 100 is the single highest weekly value any of the five terms reached, and everything else is a percentage of that. Here nvidia’s week of Jan 26 2025 is the 100, so every other reading is expressed as a fraction of Nvidia’s loudest week. These are the source file’s own weekly integers — no smoothing, no monthly averaging, no interpolation. Trends reports sub-1 weeks as “<1”; those are carried as 0.5. The shaded region marks the period after memory overtook crypto and stayed there. Hover any fortnight for all five values.
What retail attention is worth as a signal — and which direction it points
Search interest is a coincident-to-lagging indicator of retail participation, not a leading one. People search a ticker after it has already moved. So the honest reading of this chart is not "retail is discovering memory, therefore up" — it is retail has now arrived, and this board has a very recent, very expensive illustration of what that looks like at the extreme. §03·C and §04·F document the Korean single-stock leveraged-ETF episode: ₩14T of net retail buying since May 27, assets that kept rising as prices halved, and then ~₩34T of margin equity destroyed with 320,000–460,000 accounts liquidated. SK hynix search interest going from a flat zero to 39 in seven months is the same phenomenon measured with a different instrument. Two readings are available and this section does not adjudicate between them: the constructive one is that broadening awareness of a genuine supply shortage is what an early-innings re-rating looks like from the outside (§02·C). The cautionary one is that a term which sat at zero for 53 of 60 months and then went vertical is the textbook signature of late-cycle retail chasing — and the last time this panel printed a 100, it belonged to a crypto asset that is now 79% below it. The one thing the chart establishes cleanly is the rotation itself: the marginal speculative dollar of attention has moved from crypto to memory. Whether that dollar is early or late is not a question search data can answer.

Provenance & method — upgraded Aug 1 2026 from transcription to source file. Source: Google Trends, worldwide, past 5 years, Web Search, terms nvidia, micron, sk hynix, sandisk, ethereumthe corrected query. [cited — extracted]: the series here are Google Trends’ own multiTimeline.csv export, retrieved Aug 1 2026: 261 weekly rows, 2021-08-01 → 2026-07-26, five columns, values as published. This replaces the earlier build, which read 60 modelled monthly points off a screen capture. Nothing on this chart is interpolated, fitted or smoothed — each plotted point is a row in that file. Weeks Trends reports as “<1” are carried as 0.5, the only transformation applied. What the correction changed, stated precisely: (1) Scale. With Nvidia absent from the panel, ethereum’s Aug-2025 peak held the 100; with Nvidia present it reads 49. All four previously-plotted series are therefore about half their former printed values. This is a property of Trends’ indexing, not a revision to reality. (2) Timing. The earlier build said micron overtook ethereum for good in Feb 2026; the source file says the week of May 3 2026 — a three-month error, logged in the Correction Log (§10·B). The sk hynix call (Jun 2026 → actual Jun 21 2026) was right. (3) Resolution. Weekly, not monthly — a 5-year Trends window is delivered weekly, and aggregating to months would have smoothed away the peak that defines the index. Two limits inherent to the source, unchanged: (1) Trends measures the string, not the subject — “micron” is also a unit of length and “sandisk” is a consumer-storage brand bought by people who are not investors, so those lines carry non-investment traffic; “sk hynix”, whose name has no other meaning, is the cleanest of the four. (2) Values re-index if the term list changes — this panel is only comparable to itself, which is exactly the lesson the misspelling taught. On the earlier error: the previous build was labelled a transcription and carried an explicit warning not to quote intermediate months; it was fitted to reproduce both anchors the capture printed, and it did. It was still wrong about the crossover date, because a fitted curve can satisfy its endpoints and miss its middle. That is the argument for source files over screen reads, and it is why this section now uses one. Cross-refs: §03·C and §04·F (the Korean retail-leverage episode), §04·E (the DRAM ETF), §04 (the equities), §02·C (innings). Not investment advice.

08·GThe PC Industry Symptom — Priced Up, Stripped Down, Shipped Less◷ IDC Q2 2026 actual · Jun forecast
08·G

The PC Industry Symptom

§08·B watches one vendor because Apple is the most legible. This is the same instrument pointed at the whole computer industry — every Windows OEM, the channel, and the person buying a stick of RAM. The symptom takes the same three forms Apple's did, in a different order. Apple queued, then deleted configurations, then repriced. The PC industry repriced first (+15–20% across Dell, HP, Lenovo, Acer and ASUS), then began stripping configurations — some builders now ship PCs with no memory at all — and is now losing the volume. IDC has cut its 2026 forecast three times, from −2.4% to −11.3%: 32.17 million fewer PCs, more than Apple ships Macs in a year. And the number that makes this section worth reading: the market's value goes up anyway+1.6% to $274B on 11% fewer machines. That is kill-switch #5 with a receipt: demand destruction, quantified, in the largest consumer market memory has.

The mechanism is the one this whole board tracks. Samsung, SK hynix and Micron moved wafer starts to HBM and server DRAM (§05·D, §05·C·G), which starved the consumer channel. Dell's COO Jeff Clarke put the pass-through in one sentence on the February earnings call: "the spot market for a gigabit of DRAM over the last six months is up nearly five and a half times," with NAND up roughly 4× alongside it — "the cost basis is going up across all products." HP's CFO gave the balance-sheet version: memory and storage went from 15–18% of PC component cost to about 35% in a single quarter, a net $0.30 FY2026 EPS headwind after mitigations. When a third of your bill of materials reprices in a quarter, you have three levers — charge more, ship less content, or ship fewer units. The industry is pulling all three at once, and the chart below is what that looks like from the outside.

The same three symptoms as §08·B, in the opposite order — and one Apple never had to use
(1) Price moved first, and in lockstep. Dell went first, raising 15–20% from mid-December 2025; Lenovo followed in January, expiring every outstanding quote on Jan 1; HP, Acer and ASUS confirmed the same 15–20% band and reset channel contracts. Apple did not reprice until Jun 25 — six months later. (2) Then the configurations thinned. IDC's phrasing is that vendors "struggle to maintain full product portfolios" — the industry-wide version of Apple deleting the Mac Studio's 512 GB option (§08·B). (3) And then a symptom Apple has no equivalent for: some system builders now sell PCs with no RAM in them. Paradox Customs added a no-memory option to its configurator so buyers can supply their own — the seller removing the scarce component from the product and making it the customer's problem. On a vertically integrated Mac with soldered unified memory that is not even possible; on a Windows desktop it is a checkbox. It is the clearest admission on this board that the component, not the computer, is the constraint. The result is a market that is smaller and more expensive at the same time. Q1 2026 actually grew 3% — but IDC is explicit that this was borrowed: buyers "accelerated purchases ahead of anticipated price hikes." Q2 gave it back, −4.9% to 68.2M, ending nine consecutive quarters of growth. Jitesh Ubrani's summary is the line to keep: "The era of bargain-priced PCs and tablets is behind us for now."
2026 unit shipments — what happened, and what the full-year forecast still requires
Year-on-year change in worldwide PC unit shipments. Solid bars are IDC actuals, mid-tone bars are derived on this board, dashed bars are IDC forecasts. Read the two shaded bars in the middle together, because they are the point: the first half of 2026 was only −1.2% — barely a scratch — but hitting IDC's full-year −11.3% requires the second half to fall −20.5%. That is a 19-point step, not a gradual deterioration, and it has to begin in the quarter now underway. IDC's own stated Q4 figure is −20%, which means Q3 has to be roughly as bad as Q4 rather than a way-station to it. This is the most testable claim on the page: Q3 2026 shipment data lands in October and will either confirm the step or force another revision.
The same year, forecast three times in four months
IDC's 2026 PC forecast, revised as the memory bill came in. The forecast has moved further than the market has — and IDC attached its own caveat to the last one: the Middle East escalation was not yet in the numbers, "adding another source of risk," so the next revision has a known direction. Ryan Reith's framing: "the complete uncertainty around when these pressures will subside" is what turned this "from a million-dollar question into a trillion-dollar question."
Q2 2026 by vendor — and the one company growing
Worldwide PC shipments, calendar Q2 2026, against the same quarter a year earlier. Lenovo, HP and Dell all shrank; ASUS was flat at +0.2%. Only Apple grew materially — +10.1%, taking share from 8.5% to 9.9% — meaning the vendor §08·B documents as deleting its highest-memory configurations and raising prices 15–20% is the one gaining ground. Both facts are true and the tension between them is the most interesting thing here; the reconciliation is below the table.
Why the most memory-constrained vendor on this board is the one gaining share
§08·B and this section look contradictory and are not. (ASUS also edged up, by 0.2% — inside anyone’s rounding, and not the same phenomenon.) Apple is the most visibly damaged — soldered unified memory means it cannot ship a stripped configuration, so scarcity shows up as deleted SKUs and a 6.1% one-day share-price fall. But the same integration is why it is winning volume: Apple buys memory on its own long-term terms, controls its silicon roadmap, and can absorb or pass cost as one decision rather than negotiating a channel. Meanwhile the Windows OEMs are competing for the same scarce DRAM against each other and against the hyperscalers whose capex this board tracks in §06·F·D — and losing, because a server DIMM earns the supplier several times what a consumer module does (§03·F's contract tape shows the spread directly). Two caveats on the "Apple is winning" reading. First, the growth measures units, not margin — §08·B's price rises and Q3 withdrawal ladder say the cost is landing regardless. Second, the sources disagree on the size of Apple's win: IDC has Mac shipments +10.1% in Q2 2026; Omdia has +16% against an overall market it puts at −4% rather than IDC's −4.9%. Both are reported here rather than averaged. And a scale check before anyone reads too much into share: Apple's 6.7M quarter is under a tenth of the market. The 32.17M units the industry loses this year is 1.26× Apple's entire annual Mac business — the damage is happening somewhere Apple's growth cannot offset.

Provenance — what is IDC's, what is derived here, and three places the numbers do not agree. [cited — IDC]: Q1 2026 +3% at 65.6M units, attributed to pull-forward ahead of price rises; Q2 2026 −4.9% to 68.2M from 71.7M, ending nine consecutive quarters of growth — Lenovo 16.6M / 24.4% / −2.1%, HP 13.0M / 19.1% / −9.0%, Dell 9.3M / 13.6% / −5.0%, Apple 6.7M / 9.9% / +10.1% (share up from 8.5%), ASUS 5.0M / 7.4% / +0.2%. Full-year 2026 forecast (Mar 13 2026, reaffirmed Jun 2): shipments −11.3%, 284.7M → 252.53M, a fall of 32.17M units; Q4 2026 −20% YoY; market value +1.6% to $274B; ASP growth +17%; tablets −7.6% (151.9M → 140.36M) with value +3.9% to $66.8B; no relief before end of 2027, some easing from 2028, and pricing "unlikely to return to the levels seen in 2025." Forecast revisions: −2.4% (Nov 2025) → −8.9% (Jan 2026) → −11.3% (Mar 2026). [cited — vendors]: Dell +15–20% from mid-December 2025, Lenovo from January 2026 with quotes expiring Jan 1, HP / Acer / ASUS confirming the same band with channel-contract resets; Dell COO Jeff Clarke on DRAM spot "up nearly five and a half times" over six months and NAND ~4×; HP CFO on memory and storage moving from 15–18% to ~35% of PC component cost in one quarter and a $0.30 net FY2026 EPS headwind; HP's PC operating margin ~5%. [cited — retail]: a 32GB DDR5-6000 kit at roughly $80 in mid-2025 against $380–$589 on Aug 9 2026 ($11.88–$18.41/GB); system builder Paradox Customs offering prebuilt PCs with no memory installed. [derived on this board — the two shaded bars]: IDC publishes Q1 and Q2 actuals and a full-year forecast but not the intervening halves. Taking Q1 2025 as 63.7M (implied by Q1 2026's 65.6M at +3%) and Q2 2025 as the cited 71.7M gives H1 2025 = 135.4M against H1 2026 = 133.8M, or −1.2%; the residual needed to reach 252.53M is H2 2026 = 118.7M against H2 2025 = 149.3M, or −20.5%. Both derived figures inherit the assumption that IDC's full-year forecast has not been rebased on the H1 actuals — if it has been quietly revised, the required step is smaller. Stated so it can be checked in October. Three disagreements, surfaced rather than merged. (1) ASP. IDC states +17% ASP growth for 2026, but its own value and unit forecasts imply +14.5% ($269.7B/284.7M = $947 → $274B/252.53M = $1,085) — a 2.5-point gap between two figures from the same house, probably different vintages. Both are printed above; neither is adjusted. (2) Apple's Q2 growth. IDC +10.1% vs Omdia +16%, on market totals of −4.9% and −4% respectively. (3) Q1 2026 growth. IDC +3% (one IDC write-up says +2.5%), Counterpoint +3.2%, Gartner +4% — the trackers use different definitions of "PC," and the chart uses IDC's throughout so the series is internally consistent even where it is not the highest or lowest available. Two structural caveats. Units are not revenue and neither is profit — the market shrinking 11.3% while growing 1.6% in value is the whole point, and a third measure, vendor margin, is going the other way again (HP's PC margin ~5%). And the forecast is the weakest object here: three revisions in four months, with the publisher itself flagging that a geopolitical escalation is not yet included. Cross-refs: §08·B (the Apple version of this section), §08·C (memory inside Apple products), §03·E (what a gigabyte used to cost, for scale), §06·D·D (what is actually driving the demand), §03·F (the contract and spot tape doing this to them), §05·D (the wafer allocation that caused it), §02·D kill-switch #5 (demand destruction), §09 (Chinese memory entering Western PC BOMs). Sources: IDC, Jun 2 2026, Tom's Hardware on the IDC forecast cut, TrendForce on Dell and Lenovo, IDC Q2 2026 tracker, Omdia, Counterpoint, Gartner, TechPowerUp, VideoCardz, 9to5Mac, CNBC. Not investment advice.

09The China Front◷ Jun 2026
09

The China Front

As Korean and U.S. makers chase the high-margin top of the stack, China's CXMT and YMTC scale commodity memory into the gap — and race, under export controls, to build a domestic HBM supply chain for Huawei and other local AI players.

+130%
CXMT 2025 revenue growth (to ~$8B)
~12%
CXMT share of global DRAM wafer capacity (→15% by 2028)
~50%
Reported yield on CXMT's initial HBM3 stacks
~13%
YMTC NAND shipment share, nearing Micron's

CXMT nearly tripled monthly DRAM wafer output to ~290k in two years and is migrating from DDR4 to DDR5. It filed a Shanghai STAR Market IPO (~$4.1B) and shipped HBM3 samples to Huawei, targeting mass production by end-2026 — though analysts widely see competitive HBM volume as a 2027–2028 prospect.

YMTC, China's NAND leader, is pushing into DRAM and HBM; a new Wuhan fab ramping in H2 2026 could make it the world's #3 NAND maker. Domestic tool-maker NAURA is building the equipment base behind both.

The read: China meaningfully relieves the commodity DRAM/NAND squeeze — for Chinese and non-U.S. devices, where HP, Dell, Acer and Asus qualify CXMT parts behind a compliance firewall. On HBM it sits ~two generations behind, constrained by U.S. equipment controls. Near-term risk to incumbents is margin compression on lower-end HBM3/3E after 2026 — not the HBM4 top end.

China's Catch-Up Meter

Commodity DRAM & NANDcompetitive at volume
High-bandwidth memory (HBM)~2 generations behind

Volume advantage where it counts least for AI (commodity bits); a deep technology and equipment gap where it counts most (HBM). Counterpoint expects Korean firms to retain HBM dominance through the forecast window.

10·ALive News Stream◷ Jul 12 2026
10·A

Live News Stream

● STREAMING

A live wire across the whole memory and storage stack — every story timestamped with when it was published, newest first. The feed auto-refreshes every 2 minutes; you can also pull manually. Filter by category, or switch between a continuous stream and a grouped-by-day view.

Last update just now · next poll in 2:00
stories from 10+ sources
No updates in this category.
10·BCorrection Log — Every Error Caught, Disclosed Rather Than Patched◷ appended every build
10·B

Correction Log

The honesty contract with its receipts. Every material error found during a build is written down here rather than quietly patched — what was wrong, when it was caught, and what changed as a result. Validation before every release has caught real bugs in nearly every session, which is the argument for running it: the errors below were not found by readers, they were found by the checks. A board with no correction log is not a board that made no mistakes. This was previously an untitled table inside §10·A; it is now its own section because three other sections already cited it as "§10·B" and that reference pointed at nothing.

The honesty contract at work: every material error found during builds, disclosed rather than silently patched. Validation before every release has caught real bugs in nearly every session.

FoundWhereWhat was wrongFix
Aug 17 '26§06·F·W Revenue per MWThe new section was first numbered §06·F·M — which §06·F·M Micron RPO & SCAs already had. A second duplicate section number, introduced within an hour of fixing the first one (§06·D·M).Renumbered to §06·F·W. Worth recording because of how it was caught: the uniqueness check added to the audit after the §06·D·M fix flagged it on the very next build, before deploy. The check paid for itself immediately — the previous duplicate had gone unnoticed long enough for thirteen cross-references to accumulate against it.
Aug 17 '26§06·D·M / §06·D·ETwo different sections carried the same number. "AI Model Size Over Time" and "Memory Demand by End-Use" were both numbered §06·D·M, which made all thirteen cross-references to that number ambiguous — a reader following one could land on either section. Found by an audit that had never checked section numbers for uniqueness, only section IDs.Model Size keeps §06·D·M, since eleven of the thirteen references unambiguously mean it (model cadence, parameter counts, release timelines). Memory Demand by End-Use becomes §06·D·E. Two references were genuinely ambiguous — one in §06·D·W about converting token demand into "memory demand", one in a §06·D·D table row about where GB land — and both were repointed to §06·D·E as a judgement call, disclosed here rather than made silently. The duplicate-number check is now part of the standing audit.
Aug 17 '26§05·E DRAM WSPM RequiredThe model's whole-industry capacity line is below the sum of four of its own components. Vendor-level data puts Samsung, SK hynix and Micron at 3,137k wafers/month by end-2030, with CXMT a further 600–950k3,737–4,087k from four named producers against the 3,396k this section models for the entire industry. Re-running the section's own 2030 requirement of 3,775k gives 92–101% demand/supply, not the 111% reported: the 2030 shortfall narrows to ~38k or flips to a 312k surplus. This is the strongest evidence yet against one of the board's most bullish sections, and it was found by adding data the user supplied rather than by re-reading the model.Recorded in §05·E itself, not only in the new §05·D·V, so a reader of the bull case meets the counter-evidence in place. Neither series is overwritten: three specific reasons the gap may be definitional rather than real are given (installed vs effective capacity, CXMT wafer starts vs usable output, and this model's 2029–30 tail already being flagged as the most extended-bull of four houses). The harness recomputes the 111% from §05·E's own arrays and the 92–101% from the vendor figures, so neither number can drift from its inputs.
Aug 17 '26§08·B Apple Mac SymptomThe section's headline claim was falsified by events and had to be retracted, not extended. Since March it read: high-memory Macs wait months "while the base model still ships same-day." On Aug 2 Bloomberg reported the MacBook Air — the mainstream base machine — in a "major" shortage at 2–6 weeks, with retail sources calling shipments "more constrained than they can ever recall". Worse for the June framing: this board had concluded Apple "stopped queueing and started repricing". It is doing both, and the Air was sold out at a price $200 higher. Price did not clear demand.Summary claim retracted in place with the date and the reason, rather than deleted. The June "repricing replaced queueing" reading is explicitly marked superseded. The harness now asserts the phrase appears only as a quotation inside its own retraction and never as a live assertion.
Aug 17 '26Site-wide as-of dateThe board was carrying four different values for "today" at once: the header and §10·D freshness denominator said Aug 10, the footer stamp Aug 14, the catalyst calendar Aug 15 — against a real date of Aug 17. §10·D measures every section's age against that denominator, so every freshness figure on the board was understated by a week, and the "stalest sections" ranking was computed off the wrong base.All four aligned to Aug 17 2026. The harness now reads the header, the freshness denominator, baseline.asof, the footer stamp and CAT_TODAY and fails unless all five agree — so the board can no longer hold more than one opinion about what day it is.
Aug 15 '26§02·E Catalyst CalendarThe calendar's internal "today" was hard-pinned to 2026-07-30. Every countdown rendered 16 days short, and six events that had already occurred still displayed under "Ahead" — including the Aug 1 Korea trade release whose result §03·B was already reporting elsewhere on the board. A second sweep found two rows marked completed with no outcome at all (the Jul 21 Korea flash, and Tesla's Jul 22 Q2 carrying an empty outcome string)."Today" advanced to Aug 15. The six past events are marked completed; the Jul 21 Korea flash got its outcome ($54.9B, −11.3% vs June), and the six with no captured result now render an explicit "outcome not yet captured on this board" line rather than a silent blank. The harness now fails on any past-but-upcoming row, any future-but-done row, and any completed row with neither an outcome nor that flag.
Aug 15 '26§02·E Catalyst CalendarThis board applied a weekend shift that the evidence contradicts. The nine new Korea rows were dated by pushing any weekend release to the next business day — which moved MOTIE's Nov 1 full-month print to Nov 2. But MOTIE demonstrably publishes on the 1st at weekends: the completed Aug 1 2026 row in this very table was a Saturday. The shift rule was assumed, not observed.Nov 1 restored. MOTIE dates now stay on the 1st regardless of weekday, with the reason stated in the row. KCS weekend behaviour is not evidenced either way, so the one KCS date landing on a Sunday (Oct 11 → Oct 12) is labelled an assumption and named as the least reliable date in the set. The harness checks the two agencies against different rules.
Aug 15 '26§05·F AI Wafer ConsumptionAn independent source validates this section's total and undermines its composition. A Morgan Stanley SKU-level build puts 2027 accelerator units at 21.14M against §05·F's 21.30M — 0.7% apart, from unrelated inputs. But per company the two disagree by up to 3.1× (Meta 1.70M vs 0.55M), with Microsoft 1.8×, AMD 1.5× and Google 1.5×, in both directions. The section's per-company bars were drawn with the same visual confidence as its total; they do not carry the same support.§05·F now states that its aggregate is well-supported and its company split is one plausible allocation among several. Recorded in §05·C·H as well, and no per-vendor HBM share is inferred anywhere from either source.
Aug 15 '26§05·C·H (new)An earlier draft of the new section claimed the source's gigabyte and gigabit totals matched "exactly". They do not: 6,077,256 kGB × 8 ÷ 1,000 = 48,618.048 mn Gb against a stated 48,618. The gap is 0.0001% — immaterial to every conclusion, but "exactly" was the wrong word in a section whose whole argument is that the source reproduces cell for cell.Reworded to "consistent to the source's own rounding", with the arithmetic shown. The harness now asserts both that the two agree within rounding and that they are not bit-identical, so the stronger claim cannot creep back.
Aug 15 '26§04·B Analyst RatingsThe list was billed as covering "the trailing twelve months" and carried a "● live · 12 mo" tag, but its two oldest entries — Jun 25 2025 and Aug 6 2025 — were 13.7 months old. The window claim had quietly drifted past the data as the section aged. Separately, this board initially dated Samsung Securities' SK hynix cut to August; re-reading the source's sentence structure places it in the Jul 30 group alongside Shinhan, NH and Daishin, with only that firm's Samsung Electronics cut described as "this month".Window restated from the data as "since June 2025" in both the heading and the live tag, with the reason shown inline. Samsung Securities' SK hynix action re-dated to late July in both the chart data and the list. The harness now derives the oldest entry date and fails if any entry predates the stated window, so the label cannot drift again.
Aug 15 '26§04·B Analyst RatingsA source headline that understates its own reporting. The Aug 11 piece is titled "SK hynix Targets Diverge by 2 Million Won" and gives a range of ₩2.7M–₩4.7M — but reports Kiwoom's ₩2.1M target four paragraphs earlier in the same article. On the article's own figures the spread is ₩2.6M, not ₩2.0M.This board plots the full ₩2.1M–₩4.7M range and states the discrepancy rather than adopting either the headline range or silently correcting it. The harness recomputes the spread from the plotted targets and requires the prose to match that arithmetic.
Aug 15 '26§03·C Korean LeverageOn Aug 1 this section concluded that "the appetite has now broken too." Two weeks of data show that was too strong a claim from too narrow a measure. Domestic single-stock leveraged turnover did collapse — −91.0% in a month, from 33% of all KOSPI trading to 4.8%. But Korean retail bought $662.85M of a US 3× semiconductor fund on Aug 12–13 alone, holding $6.559B of it — their fourth-largest US position, up from tenth in April — while domestic margin loans rose seven straight sessions to ₩30.93T. The instrument was switched off; the appetite was not. The error was treating one venue's volume as a proxy for demand.New Aug 15 block and a two-panel chart showing the domestic channel closing and the US channel opening over the same fortnight, deliberately not netted against each other. The section now states the narrower conclusion: a regulator can close a venue quickly, and doing so relocates leverage rather than removing it. The harness asserts the section does not claim vindication.
Aug 15 '26§03·C Korean LeverageThree figures in the new reporting that do not reconcile with this board's own anchors, and one that does not reconcile with itself. A cited −38.6% maximum drawdown against this board's −40.40% from its own dated peak and trough; a cited "+5.8% this month" against +24.75% from this board's last recorded close, which would require an unreported +17.9% single session on Jul 31; a peak dated June 22 where this board carries June 19 and an earlier translation gave "July 19". Separately, the Aug 6 source prints two different turnover figures for Aug 4 (₩1.2556T and ₩919.8B) and labels two consecutive days "Aug 4".All four disclosed in a dedicated block, with the arithmetic shown. No series was adjusted to fit. Both Aug 4 figures are drawn, the second marked as a same-day second print; the Aug 14 bar is marked as a different fund basis. June 19 is retained as the peak date on corroboration grounds and the disagreement logged.
Aug 14 '26§06·F·R Neocloud FinancingThe section described CoreWeave's two 2026 term-loan facilities as closing "ten weeks apart" — in the summary, the chart title, the chart annotation, the side panel and a stat card. The two disclosed closing dates are Mar 31 and May 18 2026, which is 6.9 weeks. The section's own data object already carried the correct weeks:7; the prose had been written before the dates were pinned and never reconciled to them.All five occurrences changed to "seven weeks". The validation harness now computes the interval from the two closing dates and fails if the page states anything more than a week away from it, and separately asserts the phrase "ten weeks" does not appear inside that section.
Aug 14 '26§06·F·R Neocloud FinancingA category error in a table this board built. The GPU-backed debt table listed CoreWeave's $21B of total debt alongside Crusoe's and Lambda's GPU-collateralised facilities, under a headline figure of ">$20B of GPU-backed loans". Those are different measures, and the disclosed rows therefore sum to $21.9B — more than the market-wide total they sat beside. A reader could reasonably have read the rows as components of the headline. They are not.Table caption and provenance now state explicitly that the rows must not be summed or compared to the headline, name the total-debt-versus-GPU-collateralised distinction, and call the residual subtraction invalid rather than merely imprecise. The harness asserts the disclosed sum exceeds the market figure and that the page says so.
Aug 14 '26§05·G Memory Form FactorsThe 3D DRAM "volume, if it works" bar runs to 2032 on a chart whose axis ends at 2031. It was silently clipped at the plot edge, so the least certain claim on the page rendered as though it had a firm end date one year earlier than the data said.Bars extending past the axis now draw a ▸ continuation marker and a "runs past 2031" label. The harness enumerates every bar exceeding the axis and requires each to be marked.
Aug 14 '26Site-wide metadataThe section count is written in five places, one of which is inside a URL-encoded SVG in the og:image tag (%3E79 sections). A word-boundary sweep that updated the other four skipped that one, because E79 has no boundary before the digits — so the social-preview card would have advertised a stale count indefinitely.og:image corrected. The harness now sweeps every "N section(s)" string in the file with no word-boundary assumption, reports the distinct values it finds, and additionally pins the og:description, og:image and search-placeholder sites by name.
Aug 14 '26§08·D App-Layer RevenueThe section asserted the two labs had diverged, describing OpenAI as "flattened near a $24–25B run-rate" with a user base "plateaued ~900M weekly", and projected that flat line forward through Jul 2026. Bloomberg reported on Aug 13 2026 that OpenAI's run rate tops $40B — up ~67% from the figure this board called a plateau, and +20% month-over-month in July alone. Two consecutive prints at ~$24B were read as a structural ceiling; they were a pause.OpenAI series re-anchored to the Bloomberg figure and the "divergence" framing removed. A fourth provenance tier (third-party tracker, dashed ring) was added so measured-by-someone-else points can no longer be conflated with this board's own projections — which is the class of error that produced this one.
Aug '26§08·F Google TrendsSection was built from a screen capture of a panel whose first term was misspelled “nividia”. Two consequences: Nvidia was absent from the index, so every plotted value was ~2× too high (ethereum’s peak held the 100; it is really 49); and the transcribed curve put micron’s permanent crossover above ethereum in Feb 2026 when the source file says the week of May 3 2026 — a three-month error. The fitted series reproduced both cited anchors and still missed the middle.Re-ran the corrected query and rebuilt the section from Google Trends’ own multiTimeline.csv — 261 weekly rows, no interpolation. Crossover dates corrected; scale note added explaining that the index moved, not the behaviour.
Jun '26§06·F·B BacklogsMicrosoft 2023 shown as $112B — that was the cloud-only RPO subset, understating total commercial RPO by ~halfCorrected to ~$224B with SEC-anchored quarterly series
Jun '26§04·F Leveraged ETFsMUU AUM shown ~$400M — a stale Nov-'25 snapshot; real figure ~$5.9B (~15× error)Rebuilt all four funds on dense monthly data
Jun '26§04·E ETF RaceIBIT milestones mixed trading vs calendar days; axis too short for the ~$99B peakRecomputed on calendar days; axis extended to 900d
Jun '26§06·H AWS BlocksJun-'25 ~45% cut wrongly implied to include Capacity Blocks — AWS explicitly excluded them (separate cut cadence, then hikes)Two-track pricing rebuilt with dated anchors
Jun '26§04·D·S Short InterestGarbled "debt cut $4.4B to $5.7B" figure in the FQ3 backdropRemoved; replaced with clean sourced figures
Jun '26§06·F·B (code)Duplicate const lg declaration — would have broken the entire dashboard's JSCaught by pre-release validation; renamed
Jul 30 '26§10·A Signals (code)Two different functions were both named renderScorecard() — one rendering the §10·A signals scorecard from scorecardData, the other the kill-switch strip. JavaScript hoisting means the later declaration silently overrode the earlier, so #scorecard was never populated and that panel rendered empty. No error was thrown (unlike a duplicate const), which is why it survived several validation passes.Renamed the signals renderer to renderSignalScorecard() and updated its call site; both panels now render
Jul 30 '26§04·D Price TargetsSK hynix's "current price" was entered as ₩2.9M in the Jul 24 refresh — that figure was the stock's Jun 25 all-time high, misread from a source that quoted the ATH, not the live quote. Real price was roughly ₩1.8M at the time and ₩1,322,000 by Jul 30, so the card understated the drawdown and materially overstated implied upside.Corrected to ₩1,322,000 (Samsung likewise refreshed ₩248.5K→₩207,000); provenance now names the error explicitly
Jul 30 '26§03·C Korean LeverageThe Jul 24 build called the week of Jul 17–23 "the first stabilization" and led with it. Within three sessions the KOSPI fell 16.17% over Jul 28–29 — the worst two days in its history — erasing the rebound entirely. The call was premature: falling margin balances and returning foreigners were read as the end of the unwind when they marked a pause in it.Headline reframed to "then the worst two days in KOSPI history"; the stabilization is retained as a dated, bounded episode rather than a turn, and the forced-selling note now flags that "the peak has passed" was also claimed on Jul 21
Jul '26§06·F·B BacklogsOracle's snapshot bar drew shorter than Microsoft's despite Oracle leading ($638B vs $627B) — segments summed to the May-vintage $553BHatched +$85B "post-snapshot" segment added
Jul '26build toolingA Python \\n escaping bug injected three nav links into a JS string (join), breaking the whole file's scriptCaught by validation before release; reversed byte-exact, re-applied correctly
Jul '26§04·D TargetsRating-only firm entries added as pt:0 — the renderer draws proportional bars, so they would have shown $0-width barsCaught in validation before release; entries removed, folded into prose
Jul '26freshness system§04·B chip briefly set to a static date, misrepresenting a live-feed sectionSelf-caught same turn; reverted to 'live feed'
Jul '26§06·D·W per-MWTable said GB300 ~4.1 PB/s while its own footnote math (490×8) gives 3.9; Rubin GPUs/MW 330 vs 340Harmonized before release; table and footnote now agree
Jul '26build wiringCatalyst-calendar and provenance-index nav links + render hooks shipped without their sections — two dangling anchors (interrupted batch)Sections built Jul 6; renderer verified against its DOM target
Jul 14 '26§03 · §02·D Price basisKill-switch #1 recorded 2Q26 DRAM as ~+low-90s% — that was 1Q26’s figure; TrendForce’s 2Q26 conventional-DRAM rise is +58–63% (the §03 +61% bar is right)Re-verified vs TrendForce; §02·D corrected to the 1Q→2Q→3Q ladder (~+90–95% → ~+58–63% → +13–18%); §03 conflict flag resolved
Jul 14 '26§05·C·G HBM4Board credited only SK hynix & Micron with HBM4 first parts; Samsung had begun industry-first commercial HBM4 mass production Feb 12 2026 and was Nvidia-certified for Vera Rubin Jun 5Samsung’s commercial HBM4 added; all three now shown in HBM4 mass production
Jul 14 '26§05·B Vendor shareThe “Micron overtakes Samsung” HBM headline was circulating as a 2026 event; it is Q2 2025 Counterpoint data — Samsung rebounded to ~35% in Q3 2025Added a flagged vendor-share block to §05·B showing the firm/quarter divergence instead of one number
Jul 16 '26§03·C Leverage introA same-day edit said KODEX SK hynix 2× fell 45% “despite the underlying rising since debut” — SK hynix is down since May 27, just far less than the fundCorrected same session per the Bloomberg/Yahoo framing (fund −45%, >60% off peak, stock's decline much smaller); disclosed here rather than silently fixed
Jul 17 '26§04·F · §03·C legendsTwo chart legends and a §03·C note labeled SNXX as a "Samsung 2×" fund — SNXX is Tradr's 2× SanDisk (SNDK); §04·F's own table and provenance always said SanDiskLegends and note corrected; the series data itself (SNXX AUM) was right throughout
Jul 17 '26§02·B Supply PipelineChart titled “combined DRAM capex,” but its values track each maker's total memory capex (Micron 8→13.8 = FY24→FY25 total, not DRAM-only)Caption clarified to “memory capex (DRAM-led),” basis stated; values refreshed to latest guidance (Micron FY26 ≈$25B +81% cited; SK hynix/Samsung ramp; +2027 projection)
Jul 17 '26§06·D·R OpenRouterThe per-week tooltip listed invented per-model token amounts under real model names (fixed illustrative shares × weekly total) — contradicting the section's own cited leaderboard (implied GPT-5.6 Luna ≈16T/wk vs its printed 273B/wk) and showing models before their launchTooltip rebuilt cited-only: the snapshot week lists the leaderboard's exact figures with a coverage note; other weeks state the split isn't public; colour bands de-named (visual stack only)
Jul 17 '26§06·C · §03·D wiringTwo renderers were orphans — never called: the quarterly-units-by-vendor chart's initial draw (its only call sat in a legend click-handler, but the legend is built by the renderer — a deadlock, so the chart was permanently blank) and §03·D's nine-metric grid (defined, wired nowhere)Found via a programmatic all-renderer orphan sweep (the sweep owed since v10); both wired into their sections' on-open renders and verified by invoked-render tests
Jul 17 '26\u00a703\u00b7F Spot tapeAnnotated a \u201c$1.63 Jan-2025 trough\u201d from a secondary source; archived TrendForce captures show DDR4 8Gb at $1.50 on 2025-03-20 \u2014 lower, and later. The label overstated the bottomAnnotation now reads \u201c$1.50 \u2014 lowest capture\u201d; the $1.63 point retained as a dashed-ring superseded anchor rather than deleted
10·CProvenance Index — Every Section’s Sourcing, One Table⚡ auto
10·C

Provenance Index

The honesty contract in one view: this table is built at open, live, from every section’s own provenance footer — nothing here is written twice, so it cannot drift from the sections it indexes. Sections without a footer are visual/structural and say so.

Method. Assembled at runtime from each section’s final provenance footer (first ~170 characters, linked to the section). Structural sections without footers are listed as such rather than omitted. Not investment advice.

10·DData Freshness & Changelog — What's Current, What's Aging, What's New⚡ auto
10·D

Data Freshness & Changelog

The honesty contract, turned inward. This board mixes live-feed, dated-snapshot and structural sections; this panel makes each section's age visible — built at open from the same dataAsOf map that stamps every section's freshness pill, measured against the board's own as-of date (Aug 17, 2026). Below it, a running changelog of what's been added or refreshed — the positive complement to the Correction Log.

Section freshness — sorted stalest-first (click a chip to jump)

How age is bucketed: fresh ≤14d · aging 15–45d · stale >45d · live refreshes in-app · structural not time-sensitive. Dates are parsed from each section's own provenance stamp; "stale" means due a look, not wrong. Not investment advice.

Changelog — additions & refreshes