The bond market and the equity market are pricing the same GPUs two different ways, and the spec sheet settles it.

Sunday night the industry’s biggest borrower said the quiet part. SpaceX closed $40 billion of debt for GPUs - the largest GPU-collateralized raise on record - and Elon Musk explained why it worked: “we’re investment grade and have Starlink & launch [revenue].” His read on everyone else: debt markets have “almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers,” and “200+ neoclouds are going to see guppy-level death and consolidation.” The largest issuer just described the collateral clock in one sentence: the loans go to whoever does not need the GPUs to pay them back.

The short version

• The replacement cost of one H100-equivalent of NVIDIA merchant compute halved roughly every 49 months across the merchant line: $38,058 in November 2020 (A100) to $17,019 in August 2025 (B300), per Epoch AI’s chip-cost estimates.

• A fleet financed at H100 prices ($25,000 per H100-equivalent) is replaceable today at $17,019 - a 32% collateral haircut at replacement parity, before any second-hand discount.

• CoreWeave trades at 1.2x the replacement cost of its installed fleet ($30.44B market cap vs $25.7B of campus capex). Nebius trades at 8.9x ($42.01B vs $4.7B).

• Same asset class, same clock. One market has priced it. The other has not said which it believes.

The fork

The clock side. Pull the four merchant releases and divide price by compute. The A100, November 2020: $12,000 for 0.32 H100-equivalents - $38,058 per unit. The H100, September 2022: $25,000 for 1.0. The B200, February 2025: $37,000 for 2.5 - $14,645. The B300, August 2025: $43,000 for 2.5 - $17,019. That is 55% off over 57 months, a halving every 49 months - a rate fitted to four merchant price points, one of which rose, not a measured constant. The line is lumpy - B300 priced higher per unit than B200, memory costs riding in the package - so one step rose while the four-year trend fell by half.

Two footnotes sharpen it. Google’s TPU v8i, August 2026, carries an internal cost near $3,429 per H100-equivalent by the same estimates - the best customer’s own arithmetic prices the unit at a fifth of the merchant rate, which tells you most of the merchant price is margin, and margin is what a lender repossesses. And the CoreWeave 9.625% bonds of 2032 traded at a 14.8% yield at the Oct 14 read - a B+ credit paying a distressed-grade coupon into 2032, while the stock sits 61% below its fifty-two-week high of $141.79.

The market side. Epoch attributes 755,954 installed H100-equivalents and $25.7B of capital cost to CoreWeave’s campuses - about $34,000 per unit of compute in the ground, shells and power included. CoreWeave’s market cap is $30.44B. Divide: the market pays $40,268 per installed H100-equivalent, 1.2 times what the whole campus cost to build.

Nebius: 105,375 H100-equivalents, $4.7B of capex, market cap $42.01B - about $399,000 per unit, 8.9 times. Nscale, still private, carries a $14.6B round mark against $1.4B of tracked capex - call it ten times. The private round and the newest public fleet both assume the compute holds value; the oldest, biggest fleet is priced barely above the cost of replacing it.

Two honesty notes on those ratios: they are equity-only - market cap over capex - and CoreWeave carries the debt this piece is about, so its enterprise-value multiple sits meaningfully above 1.2x while Nebius’s cash pile against its new facility cuts the other way; and replacement cost is not collateral value - a lender repossesses into resale prices, rental demand, and remaining useful life, each of which can sit below the spec-sheet floor.

There is a counter-case, and it deserves its name. Silicon Data’s residual-value indices put the four-year-old H100 at 58-61% of purchase price this month - stickier than any melting-point arithmetic - and the B200 at 1.5x purchase under tight supply (VIG Asset Management, Oct 6, crediting Silicon Data).

Used chips trade on rental demand and scarcity, not spec sheets. But note where the two curves sit today: a used H100 at ~60% of $25,000 is ~$15,000, already BELOW the $17,019 it costs to buy equivalent new compute. The residual curve is stickier than the fitted halving - and the used market still prices old silicon under replacement parity. Both curves converge from opposite sides; the second-hand print is where the lender learns which one owns the collateral.

Both can be right only if the new fleets earn their premium before their own clock runs. A B300 bought today depreciates on the same curve the A100 did. This week the debt side kept arriving anyway: Nebius secured a $3.5B senior secured facility from Citi to fund the buildout - the same collateral class, one more layer deep.

The napkin

$25,000 bought one H100-equivalent in 2022. $17,019 buys it today.

Same compute. Same workloads. 32% cheaper, three years later, from the only vendor whose price anyone can check - and the discount compounds at every generation, halving roughly every 49 months.

A 2023-vintage fleet of 100,000 H100s carries about $2.5B of chips at cost. Re-priced at B300 parity, the same compute is $1.7B of silicon - a $800M haircut with every GPU still running perfectly. The collateral shrinks because the spec sheet advances, not because anything breaks.

Why it matters

The compute buildout is financed on the assumption that the asset underneath the debt is durable. Bernstein put CoreWeave’s funding gap near $85B last week; the bonds already trade like a restructuring case; and the equity of the youngest landlords still prices their fleets at eight to ten times the oldest one’s replacement cost. The clock says the difference between those multiples is not quality - it is age, and age is the one thing every fleet accumulates at the same rate.

For an allocator the read is structural, not a trade: the older the fleet, the closer the market prices it to the melting point; the newer the fleet, the more of its value is a bet on earning back the premium before the next release date - Nebius at 8.9 times replacement is that bet, trading today. The refinancings of 2027 - H100-era collateral, B300-era replacement prices - are where the two markets have to agree.

The falsifier

The clock ticks in public. Rubin’s release is dated September 30 per Epoch - almost exactly one half-life after the H100 (September 2022 to September 2026). Its merchant price is the test: below ~$12,500 per H100-equivalent, half of the H100’s $25,000, and the clock holds.

The first arithmetic is already in, as a desk derivation from Epoch’s spec columns and Morgan Stanley’s estimated unit price: VR200 carries 17.5 petaFLOPS of FP8 in Epoch’s table, which divides by the same 1.98-petaFLOP H100 baseline Epoch’s own conversions use into ~8.8 H100-equivalents per package - and Morgan Stanley’s $55,000 planned unit pricelands at roughly $6,200 per H100-equivalent: half the threshold, 3.5 times the B300’s compute for 1.28 times its price.

That read is only as good as the 8.8x conversion, which is a desk derivation from Epoch’s FP8 columns, not a benchmarked ratio - a workload-weighted 6x would put Rubin near $9,200 per H100-equivalent, still under the threshold but by half the margin. The clock holding, pending Epoch’s own priced row and any market print. Watch: Rubin’s merchant price; any second-hand H100 market print; CoreWeave’s Nov 9 earnings and any refinancing terms on the 2032s; Nebius’s Nov 10 print under the new Citi facility. Tracked on the public scorecard, resolve by June 30, 2027.

The second test is Musk’s own. “200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many” is a dated, checkable claim against a set this desk tracks. Scored on the same ledger: consolidation, distress, or failure events across the neocloud universe by late 2027.

The collateral was never the buildings. It was a spec sheet with a release date.

Sources

• Epoch AI - AI data centers (chip costs, installed H100e, capex; vintage Oct 25) - A100/H100/B200/B300 per-H100e costs, TPU v8i internal cost, CoreWeave/Nebius/Nscale installed compute and capital cost (CC BY 4.0)

• Stock Analysis - CoreWeave (CRWV) - $30.44B market cap at ~$55.20, Oct 26 2026 close; 52-week range $53.50-$141.79; the Bernstein $85B funding-gap note in its Oct 23 news feed

• Stock Analysis - Nebius (NBIS) - $42.01B market cap at ~$153.27, Oct 26 2026 close; 274.10M shares out

• CNBC - CoreWeave stock and bond yields, Oct 14 2026 - the 9.625% 2032s at a 14.8% yield; S&P B+

• Nebius - $3.5B Citi senior secured facility, Oct 22 2026 - the week’s new GPU-collateralized layer; Reuters coverage

• Nscale - the $14.6B Series C mark - the private-side premium against $1.4B of tracked capex

• CNBC - SpaceX closes $40B debt deal, Oct 25 2026 - Musk’s X posts verbatim: the investment-grade criterion, the debt-market shutdown claim, the guppy quote; SiliconANGLE’s transcription

• Crusoe - Neuberger $700M power-infrastructure facility, Oct 22 2026 - the non-silicon collateral class

• VIG Asset Management / Silicon Data - GPU residual values, Oct 6 2026 - H100 at 58% of purchase after four years (a mid-October read puts it at 61%), A100 at ~a quarter after six, B200 at 1.5x purchase under tight supply

• Equity-print sensitivity: market-cap conventions vary across desks by a few billion dollars (share-count and timing); the ratios here re-derive on the cited stockanalysis prints and move within roughly 1.1-1.3x on any convention - the fork does not

Numbers derived from the SOMEN graph and the Oct 25 Epoch vintages at publication. The halving rate and all per-unit ratios are desk derivations from Epoch’s cost estimates - record the intervals, never flatten them. Per-unit costs are Epoch’s own per-H100-equivalent cost column at raw precision (A100 $38,057.69; B300 $17,019.40) - not division of the rounded package prices, which would give $37,500 and $17,200.

China has twenty AI satellites doing real work in orbit right now. The first twelve fly silicon no source we hold names; the eight that went up last week fly a chip almost no one has heard of.

The short version

• The only multi-satellite compute network flying - satellites operating as a constellation, not hosted payloads - is Chinese: twenty satellites across two launches - the first twelve carrying 5 peta-operations of combined compute and 30 terabytes of storage, joined by eight more on October 14, ten AI models “deployed and tasked” in orbit, in the state-media phrasing, and Alibaba’s Qwen3 running real-world workloads since January.

• The Western answer is filings: up to 1,000,000 satellites - verified at the FCC’s own acceptance notice, 88,000 - ICFS File No. SAT-LOA-20260202-00073, filed February 4, and 51,600 - with satellite designs the filings themselves describe as “currently being matured.”

• The silicon inside the first-group fleet remains undisclosed at every primary we hold - but the second group, launched October 14, named its chip at launch: a startup’s part, XFEON’s StarCore S1, picked through ADA’s own multi-round testing. A domestic MetaX GPU flew in July on a Three-Body ecosystem satellite. What has not appeared in any public compute-constellation manifest we hold: an Ascend or Cambricon flagship part. A Chinese field survey still holds for the champions. NVIDIA announced a space-grade AI chip in March.

• The money is different too: the two disclosed Chinese programs run on bank credit lines worth roughly RMB 15 billion, not venture rounds.

• Both stacks drew circles around the same narrow, sunlit band of orbits. Both must clear a launch-cost gate no current rocket closes.

This is a structural read, not a prediction. The dated tests that could break it are at the end.

The inversion

Recently this desk published a map of the two AI compute stacks and found they converge at a single chokepoint: every advanced-node path runs through one maker of the finest chip-printing machines (extreme-ultraviolet lithography), which depends on two German suppliers. The map ends at the cloud layer. It should have kept going up.

When it does, the stacks have swapped positions. On the ground, the American stack owns the silicon and the Chinese stack scrambles for it. In orbit, the Chinese stack owns the operating record and the American stack owns the filings.

Every row of that table is a reversal of the ground-level picture. On the ground the West ships silicon and China files reverse-engineering milestones. In orbit China ships constellations and the West files spectrum applications; the West’s flying hardware so far is a demonstration and two hosted nodes.

What is actually flying

Start with the fleet nobody covers. In May 2025, a Long March 2D lifted twelve computing satellites - ten of them built by ADA Space, a Chengdu company, stacked five to a cylinder on a payload adapter. State media put the fleet’s combined compute at 5 peta-operations per second - each peta a quadrillion calculations - with 30 terabytes of storage; the same report puts the finished constellation at 1,000 POPS. Nine months of in-orbit testing followed; by February, state media reported ten AI models deployed and tasked onboard, with laser links working between six satellites. On October 14 the fleet grew by two-thirds: a Zhuque-2E lifted eight more Star Compute satellites into dawn-dusk orbit - final-assembled on a Taizhou production line, LandSpace’s first compute-satellite mission, and carrying a named chip for the first time: XFEON’s StarCore S1 - and the payloads they will test read like a data-center spec sheet: 10-kilowatt-class liquid cooling, large deployable radiator structures, and the training of AI foundation models directly in orbit.

The workloads are not demos. A demonstration satellite trained a large language model in orbit in September 2024, fifteen months before the American claim to the same first. In November 2025, Qwen3 was uploaded to the constellation’s first-group compute center - the first general-purpose large model deployed from the ground to a running satellite, disclosed in January; end-to-end inference, question up and answer down, runs under two minutes. The constellation has been used to guide robots on Earth, offloading the compute to orbit and beaming instructions down.

The customers line up behind the fleet. In June, Tencent signed on for typhoon modeling over Shanghai. In July, SenseTime, one of China’s largest AI companies, contracted for dedicated computing satellites of its own - four flying before year-end, a thousand by 2030 - with compute the announcement describes, in the standard phrase of Chinese procurement, as “autonomous and controllable.” That phrase means domestic silicon. Which vendor’s silicon, the announcement does not say. Nothing does.

That is the finding. Not a chip we cannot identify - a chip the first fleet has never named, on the only multi-satellite compute network flying (satellites operating as a constellation, not hosted payloads).

A note on the sources behind that finding: the China-side record rests on state media and one specialist trade press service - china-in-space.com, an independent English-language newsletter on the Chinese space industry, formerly China Space Updates - with no independent benchmark or telemetry in the record we hold. That is not a caveat this desk can resolve from outside China; it is part of the read itself. The numbers above are what the record says, labeled by who vouched for them.

Behind the fleet, a whole field. Shanghai alone is running two programs. A Fudan-partnered “Star Pivot” project has already shown its first three-satellite verification unit - one compute node, one weather satellite, one remote-sensing satellite, laser-linked - on a phased path to a thousand.

Alongside it, a gigawatt-class “Pearl Constellation” would move part of the city’s computing off the grid entirely: verification of domestic compute payloads - and optical compute payloads - from 2027, with orbital inference sold by the token. Above the field sits the state. CNSA, China’s space agency, has opened a comprehensive feasibility study for a space-based intelligent computing constellation, the industry ministry is building a standards system for radiation-hardened chips and inter-satellite lasers, and the whole lane is a designated focus of China’s Five-Year Plan running through 2030. Many mid-scale constellations, one state umbrella.

What is actually filed

The Western side of the lane is a paperwork arms race. In late January, SpaceX filed the application its regulator’s acceptance notice describes as “a new non-geostationary orbit system of up to one million satellites” - a system it calls Starmind. An FCC “up to” is a ceiling, not a build plan - filings routinely request more than will ever launch, and this one is more than a hundred times the largest constellation anyone has flown. Read it as a reservation of orbital real estate under first-come rules, filed against rivals doing exactly the same. Starcloud - a venture-backed startup whose December orbit-trained LLM was widely billed as the first - followed on February 4 with 88,000 of its own.

In March, Jeff Bezos’s Blue Origin filed for up to 51,600 - sun-synchronous orbits, the polar paths that keep a satellite in near-constant daylight, between 500 and 1,800 kilometers, laser links routed through its own planned 5,408-satellite TeraWave communications constellation. All three filings are now verified - two at the FCC’s own notices, the third in the official filing comments of the Secure World Foundation, which name ICFS File No. SAT-LOA-20260202-00073.

The filings describe ambition more precisely than engineering. Blue Origin’s application requests waivers from the buildout-milestone and surety-bond rules - the deadlines for deploying half a constellation, and the cash penalties for missing them - and concedes the satellite design is “currently being matured.”

The crossfire around them is a family argument. Amazon had petitioned to deny SpaceX’s application, calling it “a speculative placeholder rather than a complete application.” SpaceX then filed Amazon’s own petition into Blue Origin’s record, asking the Commission to apply the same arguments there - “to facilitate equitable and consistent review.” GeekWire’s summary: “Amazon opposed SpaceX’s filing, and then Blue Origin (both effectively controlled by Jeff Bezos) filed the exact same thing as SpaceX.”

The execution gaps are public. Blue Origin has launched zero TeraWave satellites; by April its New Glenn rocket had flown three times, and in May a fourth vehicle exploded on its pad during an engine test - the December return-to-flight target is a rebuilt launch pad’s schedule wearing a rocket’s name. One filing-week assessment noted the two-flight record; the count has since grown, the gap has not closed.

The regulator’s own reviewers flagged what the filings cannot: Blue Origin “has not previously operated a satellite constellation at any scale”, and its communications backbone is itself still a pending application. The FCC has accepted the filing; an amendment remains pending review. Acceptance is not approval, on either stack - and the Chinese projects largely have not filed internationally at all; the Space.IDC unveiling came with no regulatory paperwork submitted, which cuts both ways: no spectrum fight, and no accountability milestone either.

The silicon the champions won’t name

Here is the layer the filings skip and the state media will not name. What computes in orbit, today?

On the Chinese side, the honest answer is a method, not a part - and the absence at its center is an absence of disclosure, not a proven absence of hardware. A field survey of the Chinese space-computing industry - published on the unveiling of a national Space Computing Innovation Center in Beijing - maps seven suppliers. The “heart supplier” makes radiation-hardened spacecraft processors in the classic mold: SPARC V8-architecture chips and rad-hard reprogrammable logic chips, the computers that keep satellites alive, not the ones that train models.

The compute path that actually flies is different. One commercial operator, Galaxy Space, has run industrial-grade high-compute chips wrapped in space-hardening engineering since 2020 - commodity silicon, armored by engineering. The survey’s most candid sentence concerns the absent: Huawei’s Ascend and Cambricon’s processors are “extremely strong” on the ground, but no large-scale, self-developed, long-duration flight-validated radiation-hardened compute chip from either has publicly debuted. Absence of public debut is what the record we hold supports; proving absence would take launch manifests and export records this desk has not audited. And the edge of that absence dissolved this year, from the bottom: a MetaX GPU payload flew in July on a separate Three-Body ecosystem satellite - not one of the twenty - and passed its in-orbit reliability check, and the October batch named XFEON’s StarCore S1 at launch. Startups disclose; the champions’ flagship parts still have none flying in the compute constellations.

The one listed company selling star-borne AI units - Orbita, in Zhuhai - claims 70 TOPS (trillion operations per second) for its compute-in-memory module, a chip that calculates where it stores; the company’s own module datasheet lists 20 TOPS peak. A fork, not a spec - and in August the company named the chip: Yulong810A, self-developed and aerospace-grade, beside fully-domestic 12-TOPS and 60-TOPS boards. The claim and the product line still do not meet.

The Western answer arrived in March, at NVIDIA’s developer conference: the Vera Rubin Space Module, an AI chip line purpose-built for the space industry. NVIDIA now places compute in both stacks’ orbital plans - a merchant module for any buyer, and a full 72-chip Vera Rubin rack system as the declared payload of SpaceX’s first-generation satellite. The merchant-silicon layer that dominates the ground stack has announced its orbital extension before a Western compute constellation flies.

So the inversion, precisely - and it is organizational, not silicon: the stack whose silicon is its strength on the ground has no compute constellation flying, though the one Western compute satellite up there runs an H100 and its chipmaker has announced a space module; the stack whose silicon is its scramble on the ground has a constellation flying - and an opacity that is dissolving from the edge inward, startup by startup, champion flagship by none.

The money and the physics

The financing difference is structural. No Chinese orbital-compute company has raised a Western-style venture round. ADA Space carries a RMB 5 billion intended lease-financing facility from the Bank of China’s Sichuan branch and its leasing arm - satellite leasing, signed October 2025, plus backing from China Construction Bank and the sovereign fund CIC; the Space.IDC project signed RMB 10 billion in credit with two more banks.

The napkin: about $250,000 of intended bank lease financing per planned satellite. Inputs: ADA Space’s RMB 5 billion facility - an intended leasing amount, not drawn cash; its 2,800 planned satellites. A blended version - the lane’s RMB 15 billion of disclosed credit against the same 2,800-satellite plan - reaches about $750,000, but no document ties the whole figure to one fleet, so the desk carries the narrower number. Either way it is a floor, not a cost - and the only per-unit capital figure this lane has put on paper.

The physics both stacks share is harsher than the financing. The largest announced satellite on each side lands at the same order of magnitude: SpaceX’s first-generation satellite, AI1, at 250 kilowatts peak; the G3 flagship class at 140 kilowatts of solar generation for the Chinese Space.IDC project. The compute those numbers imply is a generation beyond anything flying.

The napkin: the flying fleet is on the order of a thousand times smaller than the paper fleet, per satellite - a cross-unit proxy, not a benchmark. Inputs: the state-media figure of 5 peta-operations across the twelve first-group satellites is roughly 0.4 peta-operations per satellite (the eight newest satellites’ compute is undisclosed, so the flying figure can only lean on the first batch); the announced G3 class claims 460 to 1,000 petaFLOPS per craft - a different unit count, operations against FLOPS, treated here as an order-of-magnitude proxy rather than a benchmark; the gap is three orders of magnitude either way. The Western paper fleet is 1.14 million satellites strong - and that count is regulator-forced disclosure, where the Chinese plan count is voluntary press; the comparison leans on an asymmetry the systems themselves create. The Western fleet flying is one demonstration satellite and two Axiom compute nodes riding a relay constellation since January; a Western compute constellation flying is still zero. Both stacks are asking for a generation of machine that does not exist; only one stack is flying the previous one.

And the gate neither manifest closes: Google’s own feasibility study priced orbital compute as cost-effective only if launch falls to about $200 per kilogram, a cadence requiring Starship-scale flight rates early in the next decade.

The same study’s radiation testing carries a quieter warning for both stacks: when Google put its AI chips in a proton beam, the memory - the high-bandwidth stacked DRAM beside every accelerator - was the most sensitive subsystem, though no hard failures appeared at mission-relevant doses. The most contested component on the ground may be the most fragile one in orbit.

The Chinese projects book small trucks - the Space. IDC first batch rides Orienspace’s Gravity-1 and Galactic Energy’s Ceres-2. The announced architectures converge on the same orbits, too: sun-synchronous, 500 to 1,800 kilometers, the one narrow band where the sun never sets. Both stacks filed the same ring of sky. It is a finite ring.

What would break this read

Dated, checkable, and none of them far away:

• December 2026: New Glenn’s return to flight, carrying a Blue Ring spacecraft. A slip into 2027 pushes the Western launch-dependent timeline with it.

• The Sunrise amendment ruling - pending review at the FCC now. Grant, denial, or return; the treatment of the requested buildout waivers is the tell.

• Year-end 2026: SenseTime’s four satellites and Three-Body’s third group - the launches that would double the fleet again. The second group went up October 14, taking it from twelve to twenty; this clock has already started paying out.

• An Ascend or Cambricon part in a compute constellation. The chip-opacity fork is already resolving from the edge - XFEON’s StarCore S1 flew on the October batch, a MetaX GPU in July - so the fact this lane now turns on is whether a champion’s flagship silicon flies. Cambricon’s own investor answer (February: the company does not work on LEO satellite edge computing) is one side of that fork.

• 2027: first Space.IDC G1 launch, Pearl Constellation in-orbit verification, Starmind’s first launches, a first flight for NVIDIA’s space module.

References

• FCC public notice, Project Sunrise application - docs.fcc.gov

• FCC record, Sunrise amendment SAT-AMD-20260917-00370 - fcc.report

• GeekWire, the Bezos crossfire and TeraWave count - geekwire.com

• The Register, execution gap at filing - theregister.com

• Secure World Foundation comments (Blue Origin) - swfound.org

• Secure World Foundation comments (Starcloud, ICFS File No. SAT-LOA-20260202-00073) - swfound.org

• SpaceBytes, LC-36 rebuild and December target - spacebytes.net

• RedEye Eagle, FAA 12-to-50 launch-limit request - redeyeeagle.com

• CCTV English, ADA Space Star Compute and the state layer - english.cctv.com

• SCIO English (Xinhua wire), the first-batch founding figures - english.scio.gov.cn

• Science Daily, Qwen3 orbital deployment and end-to-end latency - stdaily.com

• Xinhua Finance, the ADA lease facility - cnfin.com

• Orbita module datasheet - myorbita.net

• china.org.cn, Three-Body in-orbit status - china.org.cn

• china-in-space, the quiet-lead survey - china-in-space.com

• china-in-space, the first-group launch quantified - china-in-space.com

• china-in-space, the SenseTime deal - china-in-space.com

• china-in-space, Space.IDC and the G1/G2/G3 classes - china-in-space.com

• DBC field survey, the seven suppliers - enet16.com

• Yicai, Pearl Constellation gigawatt class - yicaiglobal.com

• TAIBO relays, Space String and Star Hub - en.taibo.cn, en.taibo.cn

• Securities Daily, the WAIC field and phased plans - zqrb.cn

• PCMag, NVIDIA Vera Rubin Space Modules - pcmag.com

• Wikipedia, the space-based data-center field map (AI1 and Suncatcher figures as tabulated) - en.wikipedia.org

• Eastmoney retail post, the Orbita TOPS fork - caifuhao.eastmoney.com

Counts derived from the live graph at edit time: 656 companies; 8 Chinese orbital-compute entities tracked; 20 satellites flying across two launches. Evidence classes ride each table row; retail-grade items carry their dispute inline.

- The SOMEN Desk

The short version

• Eight school-tax agreements, executed: four phases of Terafab in each of two Grimes County school districts, maintenance-and-operations relief that sums to $1,660,400,000 through 2046 - Iola ISD $644.1M, Anderson-Shiro CISD $1,016.3M, per the Comptroller’s own recommendation table.

• The filings’ own jobs number: 617 direct permanent jobs at Anderson-Shiro and 383 at Iola. 1,000 total - repeated unchanged on every phase application, which either means 1,000 across all phases or 1,000 per phase; the filings do not say.

• The division nobody has run: $1.66M per direct permanent job on the filed numbers. On the announced 3,000 jobs, $553K. Every reading of the denominator lands above half a million dollars of school-district tax relief per job.

• The pressure that produced the number: SpaceX told school districts the project “could leave Texas entirely”without the breaks, and told the board meetings it “would likely have moved to a different site in a different state” - while the separate county agreement’s $5 billion minimum carries an unsigned SpaceX line and a walk-away-anytime clause.

• The receipts are public and linked below: the eight applications, the phase ladder as filed ($45.7B Iola + $73.5B Anderson - the $119B ceiling, digit for digit), and the recommendation letters.

The fork: which denominator

The filed side. Each district’s applications carry one employment figure - 617 and 383 - and repeat it on every phase application without stating whether it is cumulative or per-phase. If cumulative, Terafab is 1,000 direct permanent jobs. If per-phase, 4,000. The honest read is the ambiguity itself: the same filing that prices the abatement to the dollar does not pin the number the abatement is priced against. Even at the generous 4,000 reading, the relief is $415K per job.

The town side. The ~1,500 figure in local circulation has a traceable origin - KBTX’s own arithmetic: “The community of Iola has a population of just over 300, with the Terafab project bringing in roughly 5 times that number in new jobs.” Five times a 300-person town. The chamber briefing carried the same figure with entry wages above $70,000. (An 1,800 figure also circulated; this desk could not source it and drops it.)

The announcement side. The $16.8 billion first phase was announced with 3,000 jobs attached. That is the number most coverage carries; it is three times the filings’ cumulative figure.

The napkin

The per-job division, on the executed benefits:

Every reading above $415,000 per job. The district-side mechanics, as filed and reported: Iola’s Phase-1 benefit is $82,255,197, secured by a performance bond of $8,225,520 - exactly 10 percent; the recommendation is rescinded if the application changes. SpaceX will pay 92 percent of Iola ISD’s bond debt, per the board president. That is the finding: not a scandal, a price - stated in filings nobody divided.

Why it matters

What the districts sold: the largest investment in Texas history, per the coverage of the eight applications - a $119.2B four-phase build against $1.66B of school relief, roughly 1.4 percent. The boards judged it worth making, 5-2, twice, the same night. A town of 300 absorbing five times its population in jobs is its own arithmetic.

The first receipts cut both ways. Iola’s taxpayers: a debt-service rate that shrinks every year as Terafab’s taxable value grows, instead of the 30-cent jump the district braced for. Anderson-Shiro’s: taxes rising about 12 cents this year on 2024 bond payments - a benefiting district whose bills still went up while the paperwork sat unsigned at the county. The allocator’s read: the unit price of onshoring a fab in 2026, paid in school-district relief - now a filed, checkable number rather than a debate.

The receipts behind this piece sit on the desk’s public reference page: data.somen.ai/companies/terafab.

The falsifier

The 2029 hiring test resolves the denominator. Iola’s filing schedules all 383 jobs for 2029, its first operations year; Anderson-Shiro’s first phase reaches the tax rolls in 2029-30. If 4,000 direct jobs materialize, the relief was $415K each and the filings’ ambiguity was modesty. If 1,000 do, it was $1.66M each. Construction itself carries a nearer test: SpaceX told the boards it wants Phase 1 underway before the end of 2026. Tracked on the public scorecard, resolve by Dec 31, 2029.

The state priced the security to the dollar and never pinned the denominator. The 2029 payrolls will.

Sources

• KBTX - the eight-application recommendation table - the M&O benefits per phase ($1,660.4M combined), the Craven signature, the zone process, the $120K in fees (token-verified on-page)

• Texas Comptroller open data - J0035, the Anderson-Shiro Phase-1 application - 617 employment, the placement schedule, the $55-119B range

• Texas Comptroller open data - J0039, the Iola recommendation - 383 employment, the $82,255,197 benefit, the $8,225,520 performance bond, the rescission clause, the Parent Company field (full-text extracted)

• KBTX - the Iola 5-2 vote - the votes, the 5x-a-300-person-town line, the move-to-another-state statement, the end-2026 construction aim

• KBTX - the two-district budget contrast - Iola’s shrinking debt-service rate vs Anderson-Shiro’s 12-cent rise

• The Texan - the 92 percent bond-debt statement - Pointer at the public hearing (text via the search index; page bot-walled)

• KBTX - the records beat - the $5B unsigned minimum, the walk-away clause, the incomplete-first applications

• SpaceX updates - the announcement / Newsweek - the 3,000-jobs framing

• KBTX - the chamber briefing - the ~1,500 figure with entry wages above $70,000

Numbers derived from the SOMEN graph at publication.

- The SOMEN Desk

The short version

• The installed base is Musk’s September 25 census: Colossus 1 at 150k H100 + 50k H200 + 30k GB200; Colossus 2 at 110k GB200 + 440k GB300 - about 780,000 accelerators, majority Blackwell. The year-end schedule takes Colossus 2 alone toward 1.21 million (the last of three 220k tranches is explicitly conditional); the fleet, if Colossus 1 stands, lands near 1.44 million.

• The efficiency number is from May: training MFU near 11%, confirmed by an internal memo xAI’s president called “embarrassingly low.” It has not been refreshed since.

• The comparison band - hyperscalers in the low-to-mid 40s - rides the same reporting wave, on smaller clusters, on possibly different arithmetic definitions. Treat it as an order, not a line on a chart.

• If the comparison were like-for-like, the gap is roughly 4x training throughput on the same silicon. Two known reasons it may not be like-for-like, below.

• And the leased object is not the May fleet: per the S-1, the lease covers capacity across both Colossus halls - terms the documents disagree about, a week apart (below). The hardware is not frozen at what produced the 11%.

The two states of the record

May. The Information reported, and an internal memo obtained by Business Insider confirmed, training MFU near 11% on a fleet then described as ~550k H100/H200 GPUs. One wire’s “60,000 GPUs actually used” gloss was wrong in the useful direction: MFU is not idleness.

MFU, in one line: of all the arithmetic a chip could theoretically do, MFU is the share your software actually wrings out. A cluster can be fully busy and still extract little.

The same wave carried the comparison set - Meta near 43%, Google near 46% - and the internal target of 50%. None of these are paired primary disclosures on matching hardware at matching scale, and MFU tends to fall as clusters grow; ByteDance has reported 55% on a 12k-GPU run, which is a different problem than a 500k one. There is also a definitional trap: against an FP8 peak the same workload prints roughly half the MFU it would against BF16. The May coverage did not say which denominator the 11% used.

October. The fleet is the census above - mostly GB200/GB300 now, with more landing in tranches through December. No public number updates the MFU on any of it. And the slack is monetized - on terms two documents, filed and posted a week apart, put in conflict. The filing side, per SpaceX’s own IPO paperwork: agreements for capacity across Colossus and Colossus II at $1.25B/month through May 2029, terminable by either party on 90 days’ notice. The other side, Musk’s clarification a week later: a 180-day lease, then mutual 90-day cancellation, short-term at SpaceX’s request - “has not committed to leasing Colossus for years.” (The ~325k-chip figure often attached is not the filing’s excerpt language, but neither is it mere relay: at least one outlet attributes it to what SpaceX said, outside the S-1.) Both statements belong in the record; the tension between them is itself the tenancy’s terms.

The book around it is a forward stack, not a billed one. Solid leg: Anthropic, S-1 language. Starting leg: Google’s $920M/month from this month. Conference leg: the ~$1.11B/month fourth tenant, billing from December 1 - a CFO remark, not a filing, and the same executive framed almost all the hosting deals as short, with a 90-day exit: stronger support than any number for reading this as an operating book, not a locked backlog. Soft leg: Reflection AI’s ~$150M/month, which lives in analyst posts and write-ups rather than the filing or the CFO record - count it and today’s billed book is ~$1.4B; without it, Anthropic stands alone at $1.25B. Run-rate if every leg lands: ~$3.4B/month. The filing’s own framing for the lease is the careful one: it monetizes unused capacity while keeping a reallocation right. (Coverage that says Colossus 1 was “inefficient for training, so rented out” is making a causal leap the record does not - the May MFU measured a Hopper training stack, not the Blackwell fleet, and is nobody’s stated reason for the lease.)

The napkin

IF May’s comparison were like-for-like: 11% against the low-40s is roughly a fourfold gap - an illustration of the order, not a measurement. The order is the finding; the multiple is arithmetic on two relayed numbers with no shared denominator.

Two reasons it may not be: the definition (BF16 vs FP8 peak - same work, half the printed MFU) and the scale (efficiency erodes as clusters grow; the comparison figures come from smaller ones).

The number that would settle it has not been printed since May. That is the finding.

Why it matters

The market translates fleet counts into capacity forecasts. If the largest fleets extract a fraction of their theoretical compute - and the only leaked measurement of the largest of them said 11% - then chip counts overstate effective training capacity by multiples, and part of the “compute shortage” is a software number wearing a hardware costume. The financing echo (this desk’s read, not The Information’s): clusters this size are borrowed against, and debt service does not care what fraction of the physics the stack extracts.

The refresh cadence is the tell to watch. xAI’s own target was 50% “in months.” If the next leak - on the Blackwell fleet, with a stated denominator - shows the ramp, the constraint story softens. If it stays quiet while the census grows, the silence is the datapoint.

The falsifier

A refreshed MFU disclosure above 25% on a like-for-like definition resolves the fork toward a software ramp; a number still near 11% on the 780k fleet confirms it. Tracked on the public scorecard, resolve by Jun 30, 2027.

Until it lands, the market is pricing a fleet it has never measured.

Sources

• Seeking Alpha’s relay of Musk’s Sep 25 census, verbatim - Colossus 1/2 breakdown and the year-end schedule (Bloomberg also reported the post; this is the captured carrier)

• The Information via our captured relay, with the BI-obtained memo - the May ~11% MFU, the “embarrassingly low” quote, the comparison figures, the Cursor rental. The primary objects are The Information’s report and the memo; this link is the relay we hold

• Musk’s May 28 clarification - the walk-back: restates the filing’s dollars and dates, adds the short-term framing; carries no chip census

• The hosting-book composite, via an aggregator - the four legs above; load-bearing claims graded leg-by-leg in the text

Numbers derived from the SOMEN graph at publication.

- The SOMEN Desk

The short version

• The announcement says “SpaceX and Tesla”: a $16.8 billion first-phase chip complex in Grimes County, Texas, 3,000 jobs, chips for robots, cars, and orbital data centers. The incentive filings say what that number is: two school-district eligible-property schedules that sum to $16.801B exactly - an abatement accounting slice, not all-in capex.

• The state registries say SpaceX. Four Terafab-side entities are on file: TERAFAB LLC (Delaware 3913340), TERAFAB TEXAS LLC (Texas 0806741990, at the Grimes County site), Terafab AI, LLC (Texas 0806598568, registered at 1 Rocket Rd, Starbase - SpaceX’s own address, and the entity behind the Texas incentive applications), and WIT Tech LLC (Texas 0804854926, Brownsville - the land company behind 7,700+ acres, nearly doubled in four days).

• The incentive filings themselves carry the same split. The narrative: “led by a consortium of affiliated advanced technology companies, including Tesla, Inc., SpaceX, and xAI Corp.” The consultant’s description: “a planned semiconductor fabrication plant by SpaceX.” The application’s own Parent Company field: Space Exploration Technologies Corp.

• The SEC-filed IPO documents carry no dollar figure for Terafab at all: 51 mentions describe “a general framework,” with timelines, milestones, and capital expenditures “not yet been determined.” The risk factor is SpaceX’s own dependence: Terafab “may not be successful, in which case we may not have other sources of sufficient AI chips.”

• The only Terafab number attached to anything resembling a signature is the smallest one - a $5 billion county-agreement minimum whose SpaceX signature line is unsigned, with a walk-away-anytime clause beside it.

The fork

What the announcement says. August 6: SpaceX and Tesla announce the Grimes County site and the $16.8 billion first phase. Coverage frames a joint venture, 3,000 jobs, the world’s largest building. The project’s own website presents the Tesla-SpaceX pairing. That is the entire public ownership record: names in press releases.

What the paperwork says. The state registry layer - file numbers, registered agents, addresses - tells a one-parent story. Terafab AI, LLC, the entity behind the eight Texas JETI incentive applications, registers at 1 Rocket Rd, Starbase: SpaceX headquarters. WIT Tech LLC, the land vehicle holding the 7,700+ Grimes acres, registers in Brownsville, the Starbase company town. TERAFAB TEXAS LLC sits at the Plantersville site itself. The Delaware parent shell, TERAFAB LLC, completes the set. Four entities, every address in the SpaceWorld, none in Tesla’s Austin. The incentive filings repeat the pattern: Tesla appears in one narrative sentence, while the applicant is Terafab AI, LLC on SpaceX letterhead, and the application’s own Parent Company field names SpaceX. The applications arrived incomplete, took weeks of corrections, and were greenlit as SpaceX told school districts the project “could leave Texas entirely” without the breaks.

What the IPO documents say. SpaceX went public in June. Its prospectus, grep-verified in full, mentions Terafab 51 times and prices it zero times: “a general framework for the future development of Terafab,” with specifics “subject to separate negotiations and agreements… and have not yet been determined.” The first 10-Q filed two days before the August announcement booked nothing. The only Terafab sentence with balance-sheet consequences is the risk factor - and it runs in SpaceX’s direction: no Terafab, no sufficient AI chips.

The napkin

The four-figure fork - same project, four numbers, paper strength inverted:

Spread check: $119B is 23.8x the $5B minimum; $16.8B is 3.4x it. The only number a counterparty could theoretically hold someone to is the smallest - and its signature line is blank. Every number above it is a press release, a schedule nobody has published, or a relay. The arithmetic cannot tell you what Terafab costs. The paperwork cannot either. That is the finding.

Why it matters

“Who owns it” is not trivia here; it decides whose equity absorbs a multibillion-dollar capex program and whose risk factor it already is. The record as it stands: the registries say SpaceX files, SpaceX buys the land, SpaceX signs (or, so far, does not sign) the county agreement; the incentive filings name Tesla in one narrative sentence and put SpaceX everywhere else. The prospectus says SpaceX depends on the outcome. Tesla’s own Austin semiconductor permits in the same window total $150,000.

The question was asked publicly within days of the announcement, and it has stayed open because the registry layer - four file numbers, four addresses, minutes to check - went unpublished. Every “joint venture” restatement between now and a signed document is a reading of a press release against the paperwork.

The receipts behind this piece sit on the desk’s public reference page: data.somen.ai/companies/terafab.

The falsifier

Any of the following resolves the fork, on the public scorecard (resolve by Apr 30, 2027): a SpaceX filing that books a Terafab capital commitment (the Q3 10-Q, ~November, is the first candidate); a Tesla 10-K or earnings call that books Terafab capex or names a Terafab entity; the executed county agreement with a signature on the $5B minimum line; or JETI documents naming applicant entities other than the four above. Until one lands, “joint venture” is press-release language and this fork stays open.

Sources

• Texas Comptroller open data - J0035, the Anderson-Shiro JETI application - the $55B-$119B four-phase range as filed, the “by SpaceX” description, the consortium narrative, the applicant entity + Hawthorne letterhead (full-text extracted)

• Texas Comptroller open data - J0039, the Iola file + recommendation letter - both Phase-1 eligible schedules (the $16.801B resolution), the letter addressed to SpaceX’s property-tax manager

• OpenCorporates - Terafab registry results - the four entities, file numbers, registered addresses (the receipt layer of this piece)

• SEC EDGAR - SpaceX 424B4, Jun 12 2026 - 51 Terafab mentions, zero dollar figures, the framework language, the AI-chips risk factor (full-text verified)

• KBTX - the Terafab records beat - the $5B minimum, the unsigned line, the walk-away clause, the incomplete JETI applications, the rail restriction, WIT Tech’s 7,700+ acres

• SpaceX updates - the August announcement - the $16.8B phase-one figure as announced

• FT - Grimes County 100% abatement - the $55B project framing (now resolved to the filed range floor)

• Statesman - Tesla Austin permits - the ~$150K Austin semiconductor permit trail

• Newsweek - the announcement coverage - the 3,000-jobs framing and groundbreaking

Numbers derived from the SOMEN graph at publication.

- The SOMEN Desk

The short version

• Tenstorrent Holdings filed a Form D on Sep 9, 2026: $1,286,829,662 sold to 134 investors of a $1,416,015,946 equity offering, first sale Aug 26.

• The filing discloses no valuation, no price per share, and no investor names. Form D never does.

• The circulating $3.2 billion is not a mark: it is a November 2025 talks figure - an $800M raise, Fidelity-led - that was never announced closed.

• The listing’s $108.51 (Oct 7) is an indicative model price, not a trade. A vendor can turn it into a valuation by assuming a share count. The filing cannot, because it publishes neither.

• Filed capital, unclosed talks, and an indicative price. The arithmetic cannot run from the record.

The fork

The talks side. The number that circulates is $3.2 billion - but the coverage it rides on says, in its own body, that Tenstorrent was “in talks to raise $800m at a $3.2bn pre-money valuation led by Fidelity Management in November 2025.” No closed round was ever announced at that number. The same coverage reports early-stage takeover conversations with Intel and Qualcomm. The last closed round on record is the $693M Series D, December 2024 - a $2B pre-money, about $2.6B post, and itself part note: $593M of equity plus a $100M convertible from 2023 that converted in. Everything between that post and today is talk.

The filed side. The $1.29 billion notice is the largest capital event in the company’s filed history - the $1.416 billion offering is 21.5x the fully-sold $65.8M July-2025 notice, and the amount sold is roughly 2x the $693M December-2024 Series D. Rule 506(b), equity securities, $129.2M still on offer. The filing says capital arrived. It does not say at what price.

The indicative side. The listing carries a daily price series - 122 days of it, latest $108.51 - built from blended transaction and quote data, not from trades you could audit. Two independent vendors run cents apart on the same days (Stock Analysis: $108.48-51, and our own capture ticked $108.35 to $108.51 across the week), which tells you the series is real but the instrument is a model. And the model becomes a number only when someone adds a share count: Stock Analysis prints an implied $5.76 billion, dated Aug 26 - the day the filing’s first sale closed - using an assumed share count the filing never discloses. A vendor can cross that bridge. The record cannot.

The napkin

IF the unclosed $3.2B talks figure held as pre-money:
$1,286,829,662 primary would buy 1,286.8 / (3,200 + 1,286.8) = 28.7% of the company.

No disclosed investor group claims a stake near that - though Form D would not show it either way, and a lead, a club, or a converting instrument can absorb that much quietly (the last round itself closed a $100M note into equity). The implausibility is this desk’s prior, not a record fact.

Run the same conditional off the CONFIRMED December post instead - $2.6B pre - and $1,286.8M buys 1,286.8 / (2,600 + 1,286.8) = 33.1%: dilution larger, price still unpublished.

So the number moved before the money did, or the filing is not one primary round. The napkin stays conditional. That is the finding.

Why it matters

For calibration, the clean peer priced its paper: Etched’s $700M Series D printed a $21B valuation in round coverage. Groq - no longer a challenger so much as an NVIDIA-systems operator since its licensing deal - still printed $3.5B on its August round. (“Post” at both is the standard reading of “valued at,” not a term either company printed.) Tenstorrent moved roughly double their capital and printed none.

The practical point for an allocator: every “at a $X billion valuation” restatement of this round you read between now and a priced disclosure is invention - a talks figure repeated as a mark, or a vendor’s assumed share count presented as a valuation. The company raised real money from 134 real buyers. What they bought remains the only number nobody printed.

The checkable mechanics, if you want them: EDGAR full-text search on “Tenstorrent” surfaces the notices as they land; Fidelity’s public monthly holdings listings carry Tenstorrent Series D1/D2/C1 at value level (the Sep 9 listing: $18.4M for D-1) - institutional marks are public monthly, even when per-share terms are not; and any named-round coverage would carry the post-money that resolves the fork. None of that requires this desk. That is the point of a falsifier.

The falsifier

Any disclosed post-money for the Aug-2026 raise - named round coverage, fund portfolio disclosures, or an eventual S-1 - resolves this fork. Tracked on the public scorecard, resolve by Apr 30, 2027. The near-term watchpoints are already in the wild: two Fidelity portfolios valued the same Series D-1 at $75.63 (May 31) and $111.61 (June 30) - 47.6% apart - and ARK Venture marked it $81.39 then $89.42. Registered funds are the only independent pricing evidence on record, and they disagree with each other.

Sources

• SEC EDGAR - Tenstorrent Holdings Form D, Sep 9 2026 (CIK 2031944) - amounts sold/offered, investor count, first-sale date, prior-round comparisons

• The Next Web - Intel/Qualcomm coverage - the talks figure ($800M at $3.2B pre, Nov 2025), the takeover conversations, the Dec-2024 Series D baseline

• Clarity (Hiive) - Tenstorrent listing - the $108.51 indicative price, Oct 7 2026 capture (series moves daily)

• Stock Analysis - Tenstorrent - the independent price series (cents apart) and the $5.76B implied valuation dated Aug 26 2026

• Fidelity - Select Technology Portfolio holdings, Sep 9 2026 - Tenstorrent Series D1/D2/C1 at value level

• AltStreet - the Form D, both notices, and the fund marks - the 47.6% Fidelity fork ($75.63 vs $111.61) and ARK’s marks

• BetaKit - Series D close - $2B pre / $2.6B post, $593M equity + $100M converted note

Numbers derived from the SOMEN graph at publication.

- The SOMEN Desk

The short version. NVIDIA’s own quarterly filing carries $366 billion of total future commitments and, separately, $108.5 billion of maximum guarantee exposure. One of the largest single coherent GPU fleets ran at about 11 percent model-FLOPs utilization - the share of the math a chip could theoretically do that training work actually uses, a software-efficiency result and not idle silicon. Different paper on different sides of this market: the $366 billion mixes instruments, and only its supply-purchase lines may be canceled, rescheduled, or adjusted before firm orders; guarantees pay only on a trigger; take-or-pay is the tenants’ form, signed faster than the compute is being delivered, mostly in paper that is not filed anywhere. The whole ledger, one row per number:

Every number in this piece sits in that table, in the references at the end, or in a napkin. The prose carries only the ones the argument collapses without.

In August, NVIDIA filed a 10-Q carrying $366 billion of total future commitments - supply agreements, cloud capacity, leases, equity - and the instruments do not share one set of walk rights: only the supply-purchase lines may be canceled, rescheduled, or adjusted before firm orders. Separately, and mostly not yet effective, sits the guarantee stack: a maximum gross exposure of $108.5 billion, of which $105 billion is capped at a single campus, SB Energy’s PORTS in Pike County, Ohio, occupied by an OpenAI affiliate.

The instrument is named in NVIDIA’s own August filing: residual value guaranties, a form borrowed from commercial real estate. The chipmaker guarantees the long-lived layer - land, power, and shell - and sells the short-lived one: the campus hosts NVIDIA compute exclusively, hardware that refreshes several times across a twenty-year lease.

The mechanics are as soft as paper gets. Exposure takes effect only as halls come online, declines as rent is paid, and pays nothing unless a trigger fires - OpenAI insolvency or failure to pay. Even then it pays only a shortfall. OpenAI reimburses what NVIDIA actually pays, and the whole structure terminates the day OpenAI earns a satisfactory credit rating.

Read the design and the purpose is plain: OpenAI is private and unrated, so it borrows NVIDIA’s balance sheet to make a twenty-year lease creditworthy. And note the elegant loop in the reimbursement term - the guarantee pays exactly when the counterparty reimbursing it has just failed.

That is the supply side, filed. The demand side lives in press releases and unnamed-source stories, and it is enormous: Anthropic alone signed roughly $107 billion of compute deals in about ninety days - $144 billion only if the one cancellable monthly line runs its full term. Published tallies of its 2026 commitments range from a quarter to nearly half a trillion dollars depending on which deals the counter includes, and none of the tallies is the market. The filings will say which.

The utilization side is barely measured at all. In May, The Information reported that xAI’s fleet of roughly 550,000 GPUs was achieving about 11 percent model-FLOPs utilization, against 43 and 46 percent at Meta and Google on the same metric. An internal memo obtained by Business Insider confirmed the figure, and the company’s president called it, verbatim, embarrassingly low.

A market where sellers file $366 billion of commitments and the one public efficiency datapoint reads 11 percent looks less like supply and demand than like credit. That reading is this desk’s frame, not a finding: the efficiency report was about software, and one lab’s stack does not grade the take-or-pay book - it grades cost per useful chip-hour for that tenant, if the memo is right. But whatever the frame, the market’s paper is a ledger of promises, graded by who filed what.

Where the guarantees live

The 10-Q commitments table is NVIDIA as underwriter. The company does not just sell accelerators; it guarantees leases, backstops data centers, and signs its own cloud agreements on top.

Above the filing sits the platform NVIDIA announced in August: memoranda of understanding with six Wall Street namesto mobilize more than $500 billion of third-party capital for AI compute. BlackRock’s Larry Fink, on the announcement panel: “the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s. And I look upon this as a next future for financial engineering.”

NVIDIA coordinates the capital, connects private equity to its own customers, holds an option to backstop up to a quarter of projects, and collects the chip purchases as revenue. The capital funds the customers; the customers buy the chips; the chip purchases are NVIDIA’s revenue. Fink’s mortgage-backed comparison was not a slip. It is the term sheet.

NVIDIA is not the only chipmaker running this play, and the second instance is more leveraged. Broadcom’s AI XPV does the same thing one layer over: debt that buys Broadcom racks, leased to frontier labs, with Broadcom’s own 10-Q capping the first-deployment backstop near $29 billion, nothing yet paid under it, and a syndicate now gathering $60 billion more - current reporting, not a closed facility. These numbers are not forecasts of chip demand. They are debt secured on the expectation of chip demand, repaid by leases guaranteed by the chipmaker whose chips the debt buys.

And the buyers are building platforms too. In August, Anthropic, Macquarie Asset Management and GIC announced Theseus Infrastructure, a dedicated data-center platform with Anthropic as anchor tenant and the asset managers supplying majority equity. No figures were disclosed anywhere. The structure is the fact: the market’s largest tenant helped found its own landlord, because the lease commitments got too large to keep renting from strangers.

Who rents from whom

The demand ledger is best read as one company’s ninety days.

Anthropic’s blitz, mid-year: five deals, five landlords, every one reporter-grade. A West Virginia campus with Nscale. A Memphis cluster via SpaceX. A converted bitcoin-mining site in Texas with Riot. A Norwegian campus with Volta. A Texas site with Lambda whose lease NVIDIA itself holds - the chain runs miner-developer to chipmaker-landlord to cloud-operator. The full itemization, with every term the desk could verify, sits in the references.

The SpaceX leg is the softest large dollar in the ledger, so label it before counting it: a monthly rate through May 2029, terminable by either side on ninety days’ notice - and Musk’s own description was a 180-day lease with cancellation after that. An initial-term ceiling near $7.5 billion, not a committed $45 billion. The sum above carries both tenors.

Notice who the landlords are. Riot pivoted from bitcoin mining. Volta’s first deployment sits on Bitdeer’s Norwegian campus. Nscale was built on a bitcoin-mining group’s hydro site. Crusoe started in stranded-gas bitcoin. They came from crypto because they already held the asset this market prizes most - permission to pull serious power from the grid. A miner’s site is a data center with the hard part already done.

Grade the same ledger by what is actually filed, and three landlord filings carry Anthropic lines: TeraWulf’s July 8-K on a twenty-year lease with renewal options, SpaceX’s own IPO paperwork confirming the monthly line, and Nscale’s S-1 - the largest of the three, and the next section’s subject. Riot has never named its tenant in its own disclosure; Bloomberg did. Three lines are the filing-grade ledger. Everything else with a nine-or-ten-figure number in it is a press release or a sourced story.

The filed version of a compute contract is also where you can read the walk rights. The Google-SpaceX agreement is an SEC filing, and it terminates on ninety days’ notice, with Google holding an exit right if delivery slips. Filed does not mean short - TeraWulf’s filed lease runs twenty years. Filed means you can read the walk rights. The unfiled versions are described, by the people selling them, as twenty-year take-or-pay, and their walk rights are unreadable.

What the tenants’ own paperwork says

Take-or-pay migrated from LNG terminals into memory chips and then into compute, and it is now the standard form on both sides of this market. What makes the moment unusual is that two of the tenants have filed documents, and the filings do not describe the press releases.

Nscale’s S-1 is the cleanest example. The company that signed a $45 billion deal with Anthropic filed total contract value of $103.4 billion against $2.6 billion of active contracts - and the gap is the whole story.

The napkin. $103.4 billion of contracted value; $2.6 billion of contracts in force. The active paper is 2.5 percent of the total.

The exhibits show how the credit actually works. The Anthropic agreements sit inside four Delaware special-purpose vehicles - shell companies set up to hold the contracts. No fees are payable for a tranche of capacity until Anthropic’s written acceptance of it. Late delivery carries a termination right with no payment and no liability. There is no termination-for-convenience clause - no walking away because you changed your mind.

And the customer grants the provider the right to pledge the agreements and the receivables - the money it is owed - to its project lenders, with direct agreements preserving the lenders’ right to step in and take over if the provider fails.

Read that last clause again - it is the financing model in one sentence. The provider borrows against the customer’s signature. Anthropic’s contract is not just a revenue line; it is the collateral in somebody else’s credit facility. Multiply that structure across the backlog and the take-or-pay clause stops being commercial boilerplate. It is the load-bearing wall of a debt stack.

And the walk risk is not hypothetical. OpenAI, the announced initial offtaker for Stargate Norway, never concluded the offtake and stepped back; Microsoft absorbed the capacity. No one booked a loss; the site found a new tenant; the story moved on. Every take-or-pay ledger in this market carries that option, usually without the ninety-day notice.

What does an 11 percent fleet actually mean

Start with what the number is not, because the misreading is everywhere. Model-FLOPs utilization is the share of the math a chip could theoretically do that a training run - the work of teaching a model - actually uses. The rest is lost to waiting on data movement and communication between chips, while the chips themselves run hot. It is a software-stack efficiency metric. Eleven percent does not mean 89 percent of the GPUs sit idle; it means the stack wrapped around them converts about a ninth of their theoretical peak into training work.

The Information’s own translation was an “imperfect analogy”: effective compute like about 60,000 GPUs on the ~550,000 fleet. An analogy, labeled as one - not a census of idle silicon. The comparison set matters too. Nobody runs at 100; the physics of coordinating hundreds of thousands of chips guarantees losses, and production-grade training typically lands between 35 and 45 percent. Coverage was explicit that xAI’s hardware deployment is best-in-class and the losses live in the stack. It is a spread: the best operators extract roughly four times the delivered compute from the same chip-hour that the newest mega-fleet does.

Why does that belong in a piece about credit? Because under take-or-pay the landlord’s debt service runs on the contract, not on utilization. The efficiency cost lands on the tenant: the same committed dollar buys a quarter of the useful work the best operators extract. Carried long enough, that is what makes a tenant stare at a take-or-pay ledger and ask which line it walks first.

The napkin. 11 percent against 43 to 46 on the same metric: the best operators get about four times the delivered compute from the same chip-hour. Under take-or-pay that spread is the tenant’s unit-cost problem - and the leading indicator of who walks first.

The escape valve is already open: xAI began renting surplus capacity externally - Cursor training its latest model on tens of thousands of xAI GPUs, per Business Insider. Musk’s stated plan is gigawatts growing sevenfold by late 2027, and SpaceX’s hosting portfolio is the same answer: surplus capacity becomes a landlord’s inventory. Watch the rental pricing - the market’s real utilization view prints there first.

The sellers’ own prints argue demand is real - Micron guided supply-demand much tighter through fiscal 2027 and 2028, with customers prepaying to secure supply - but demand being real does not make delivered compute real. Contracted capacity and useful output are separated by a software stack, and that separation is where the financing risk lives.

The demand side is now pitching the same future on the record: 2028 revenue in the hundreds of billions, a pre-IPO credit facility in the tens of billions, and an IPO reported for as soon as mid-November - a target that has already slipped once in the sourcing, with investor meetings reported from October 14. The commitments ledger and the revenue pitch are the same document read twice: the take-or-pay rungs are the collateral for the projection.

Chips as collateral, and the seller’s twin

The frontier of the credit market is now below the data center, at the chip itself. General Compute, founded in 2025 with a team under ten people, runs an inference cloud - it rents out chips for running trained models rather than building them. It signed an up-to-$400 million debt facility billed as the first with inference chips as collateral, drew on it for a Cerebras agreement the company calls its largest hardware commitment, and took over its first six megawatts in Texas. The lender is also a shareholder.

The established end of the same market now prints investment grade: on October 1 Lambda closed a $1 billion senior-secured, fixed-rate term loan - drawn in stages as deployments come online - marketed to insurance companies and fixed-income investors. When insurers are the buyers, the collateral is no longer venture paper.

The same contract form binds the market one layer down, in memory, and there the seller files. Micron’s September print: 26 Strategic Customer Agreements, all take-or-pay, extending into 2031, with customer financial commitments in the tens of billions, the vast majority cash deposits - three distinct measures, kept apart in the references.

The buyer side of that architecture is the compute ledger you have just read: the same labs, the same years, the same clause. Memory’s version is filed because memory’s sellers are public and their auditors demand it. Compute’s is not, because nearly all of its landlords are private. The asymmetry is the story: when obligations are real enough to file, they get filed. Most of this market is not there yet.

Where the loss lands

A structural read owes the allocator one more table: not what happens, but who is standing where when it does.

Delivered compute stays thin while contracts hold. The tenants pay take-or-pay, so the landlords’ debt service is met, and the loss is absorbed invisibly as cost per useful chip-hour at every AI lab. The public sees none of it. The tell is the efficiency spread, which is why one eleven-percent datapoint is worth a hundred press releases.

Demand cracks and a contract reopens. The prepayments burn first - Micron’s deposits, a fifth of contract value at Nscale, whatever unwritten floors sit under the unfiled deals. Then the guaranties trigger, paying only the shortfall, with NVIDIA’s reimbursement claim resting on the counterparty that just failed. The providers’ project lenders follow, holding receivables from a tenant that walked. The Stargate Norway exit is the precedent: nobody filed a loss, the site found a new tenant, and the story moved on. Scale that gracefully.

The IPO window opens and stays open. Then none of the above happens on paper, because the commitments are refinanced into public equity at marks the private rounds set. The mid-November filing window is the transfer mechanism for the entire private ledger: after the S-1, the take-or-pay rungs, the prepayment terms, and the utilization assumptions all become public-document facts, gradable by anyone. The promises stop being promotional the day they are filed.

What would falsify this read

Dated, checkable, and none of them require anyone’s opinion. Is any of it priced in? This desk reads structure, not prices - the falsifier list is the position: when one lands, the repricing happens whether the market saw it coming or not.

1. NVIDIA’s Q3, November 18. The commitments and guarantee lines move or they do not. A flat guarantee line after a quarter of new platforms would be the single most informative datapoint of the season.

2. Anthropic’s S-1, expected as soon as mid-November. The compute-commitments note grades the desk’s tallies and every reporter-grade deal in the ledger at filing grade.

3. xAI’s MFU progress toward its 50 percent target, or a rental price for surplus capacity. Either the spread closes and the same fleet supports multiples of the May work, or the rental market prices what the stack cannot extract.

4. Nscale’s first post-listing quarter. Bookings against a $103.4 billion backlog, on a public clock - the first test of whether a neocloud’s filed backlog converts to active contracts at a rate that justifies the buildout.

5. A walk. One more Stargate-Norway-class exit re-prices the prepayment layer everywhere.

References and numbers

Every figure the prose left out, with its source. Numbers here are load-bearing for lookups, not for reading.

The filed supply side. Commitments breakdown and guarantee stack per the 10-Q: $120B remainder-of-FY27 / $100B FY28 / $98B FY29; $279B supply-and-capacity within it, up from $119B one quarter earlier; a further $56B of cloud agreements and uncommenced leases; option on roughly 3.8 more gigawatts at PORTS (“the filing’s own rounding”); first halls expected calendar 2028 / fiscal 2029. The residual value guaranties 8-K exhibit in full.

The ninety-day blitz, itemized. Nscale: $45B, six years, ~460MW, West Virginia. Riot: $9.1B, twenty years, 191MW, Rockdale TX, to June 2048. Volta: $10B, six years, Norway. Lambda: $35B, Texas - tenor widely reported as six years is a borrowed prior, not a disclosed term. SpaceX: $1.25B/month through May 2029, both Colossus campuses, 325,000+ GPUs, 90-day mutual termination; sum arithmetic 45 + 45 + 9.1 + 10 + 35 = $144.1B at full ceiling, $106.6B committed-tenor.

The tallies. One census counts five 2026 deals above $275B with AWS in the set; another carries roughly $254B confirmed, ~$454B with the reported $200B Google web; reporters quoting Anthropic’s confidential IPO paperwork put the April Google-Broadcom TPU agreement at a $125.2B five-year commitment - not public on EDGAR.

The hyperscaler layer. Over $100B with AWS over ten years, Amazon investing $5B now and up to $20B more; $30B with Microsoft Azure plus up to a gigawatt of NVIDIA systems, Microsoft and NVIDIA investing $5B and $10B; $50B toward Fluidstack-developed US sites, Google backstopping the leases for warrants.

The filed demand anchors. TeraWulf 8-K: ~$19B, twenty years, 401MW, two five-year renewals, investment-grade credit “expected.” Google-SpaceX FW filing: $920M/month October 2026 through June 2029, ~110,000 GPUs, 90-day termination after December 31, delivery exit right. SpaceX CFO portfolio: $3.43B/month across four tenants (the two smallest lines are the softest inputs); the new $1.11B/month tenant, identity undisclosed.

Nscale’s S-1, beyond the napkin. H1-2026 revenue $140.6M against a $1.02B net loss; RPO $56.4B at June 30; prepayment 23 percent of contract value; Anthropic agreements named at “up to approximately $44.6 billion”, acceptance-gated; Aker’s half-year report notes all material projects remain in development.

Micron’s three measures. Customer financial commitments $32B, vast majority deposits; deposits on balance sheet $12.7B after a quarter taking in $12.3B; RPO ~$150B on the call’s count; even at floor prices, margins above any prior cycle peak (call).

Broadcom’s AI XPV. 10-Q: initial $35B tranche, >1GW deployed, >20GW aimed through 2028, ~$29B backstop cap undiscounted, $42B convertible-note ceiling (separate, undrawn). October package per Bloomberg: $42B senior + $18B Blackstone-led junior, $9B Blackstone own funds. Bank of America’s $370B is a 2029 forecast, not a borrowing.

The efficiency report. xAI ~550,000 H100/H200 across Memphis and Colossus; effective-compute analogy ~60,000 GPUs; Meta 43 / Google 46 on the same metric; production norm 35-45 percent; xAI target 50 (the report’s public carriers).

Anthropic’s IPO window. 2028 revenue pitched at $190-200B; pre-IPO revolver above $10B; raise reported above $100B.

Every number above is drawn from the desk’s graph, where each fact carries its source, date, and evidence grade. Where two published totals disagree, both are stated; the filings will settle which, starting in November. The track record for every claim in this series sits on the public scorecard, with falsifiers and resolve-by dates, because a structural read that cannot be checked is just a mood.

- The SOMEN Desk

Three rules govern everything published here:

1. Every number resolves to a link a reader can open.

2. Every core claim ships with a falsifier and a resolve-by date.

3. Every resolution lands on the public scorecard: confirmed, refuted, or partial, in plain view.

If you follow one thread

The AI supply chain. Two Stacks: The AI Compute Supply Chain Map traces both AI compute chains, Western and Chinese, from the financing through the power to the lasers, and finds the same shape in each: the chain narrows as it deepens, and the narrowest points have a single supplier. The follow-ups go one layer deeper: The Bigger Stencil maps the photomask chokepoint under the world’s only EUV lithography vendor, and Memory After the Sold-Out Years maps the memory squeeze behind the price you paid at the phone counter.

If you follow the money

The pre-IPO record. Nscale: $103 Billion Claimed, $21 Million Filed is the desk’s method in one case: an IPO backlog claim walked down to the audited revenue lines in the Norwegian register - then graded when the S-1 landed. Anthropic’s $543 Billion Compute Ledger adds up the take-or-pay rungs, counterparty by counterparty. Etched: The 2.9 Percent Round is the ownership arithmetic every private round gets here: what the new money actually bought.

If you follow the robots

Not Publicly Computable takes the robotics market’s marks - Unitree’s IPO, Figure’s round, AgiBot’s ask - and asks what each number actually measures. The answer is usually “the instrument, not the company.”

The method, briefly

Research runs on one knowledge graph built for this desk: hundreds of companies and tens of thousands of sourced relationships spanning funding, patents, hiring, supply contracts, dependencies, and secondary-market pricing. Every relationship carries its source, its evidence date, and a confidence score. A number without a source does not print.

The machine-readable layer - the scorecard, the marks table (claim vs round vs secondary tape), and the napkin table (new-money stakes) - lives at the desk’s reference site.

Nothing here is a prediction, a price target, or investment advice. It is a structural read of the public record: what the companies say, what the record shows, and what would prove the desk wrong.

Subscribe to get the next teardown.

Subscribe now

The desk keeps a napkin table. It prints one number per private company: the size of the last round divided by the post-money valuation that round itself set, meaning the price the whole company was marked at the moment the new money arrived. No projections, no comparables, no argument. Ten robot companies carry prices or price-claims in the graph. For the largest private robotics round of the year, the table prints nothing at all. Atoms raised $1.7 billion in July, led by a16z, and there is no valuation attached to it anywhere we can check. No round price, no tender (an organized buy of existing shares), no secondary print (private stock changing hands between strangers) on the tapes we read. The row exists. The arithmetic cannot compute it.

What makes that worth a piece is the company it keeps. In the weeks around it, the sector printed two other answers to the same question. XPeng’s robotics arm raised $900 million and told the market the number is $6.3 billion, while its own exchange filing, a Hong Kong announcement deposited with the SEC as a 6-K, shows a third of the money came from insiders. The company’s press release announced the round; the filing shows who actually wrote the checks. And Unitree, a Chinese humanoid maker, listed in Shanghai at a $9.04 billion valuation and closed its first trading day near $50 billion. The offer was priced in August; the debut repriced it within hours.

Three answers to one question: what is a robot company worth? Nothing, if you will not say. More, if your own shareholders help bid. Whatever the tape says, if you can find a tape. Each answer measures something different, and the differences are where an allocator gets hurt.

A quick orientation for readers who do not track the sector weekly. Figure and Skild are the American leaders, factory humanoids and a software brain that runs across many robot makers. Atoms is Travis Kalanick’s physical-AI holding company: kitchens through CloudKitchens, mining autonomy through Pronto, stated ambitions in transport. It is not a humanoid maker, its mark is not comparable, and it sits here because its round is the year’s largest.

Dogotix is XPeng’s carved-out robotics arm; Unitree and UBTech are the listed Chinese makers; AgiBot the largest private one; 1X is the OpenAI-backed consumer bet, Apptronik the Austin challenger backed by Google and Mercedes-Benz, Physical Intelligence the foundation-model shop. The table repeats the rest where it matters.

One table, ten marks

Every robotics mark in the graph, with the evidence class each number sits on. The grades are the desk’s; the links are the sources. Read them as ten measurement instruments that print in the same unit: the sector stops looking priced and starts looking surveyed.

The silence

Start with the absence, the largest number in the sector. Atoms merged its robotics businesses into one structure and raised $1.7 billion in equity from a16z, Bain, Uber and others in July. The round has no recorded post-money. The napkin rule is strict: a percentage prints only from a round’s own post-money. For atoms the field is blank, and blank is what the company chose.

The founder has been here before. CloudKitchens, Kalanick’s previous company, disclosed its valuation twice, both times within weeks of the money arriving: $5 billion in 2019, $15 billion in 2021. One compilation of the 2026 funding cycle catalogs the difference. The revenue figure is the founder’s own: a Bowl Builder kitchen robot, he said in mid-2024, “is already running a $2 million AUV restaurant in Pittsburgh, right now, delivery only.” AUV is one site’s annual sales, the number is his, and no customer or exchange has checked it.

Napkin. There is no napkin. No consolidated Atoms revenue is public, and the only figure anyone has is one kitchen’s $2 million. A kitchen is not a company. Force the fraction, 1,700 / 2 = 850, and you have measured a holding company’s raise against one site’s sales. The $1.7 billion has no usable sales multiple on any public number.

A statement about disclosed robotics sales, not economics: CloudKitchens is a large real-estate and operations business underneath this structure, and its numbers are not public either. The blank is a disclosure choice with three businesses behind it, not evidence of absence.

A single-source $9 billion figure did surface in late July. The aggregator RoboticsIntl attributed it to sources; the company PR and mainstream coverage carried no number, no other outlet has matched it, and no secondary market has printed one since. The round also came with debt from five banks, Bank of America, Goldman Sachs, Wells Fargo, JPMorgan and Barclays, in an undisclosed amount, so the total capital package exceeds $1.7 billion by an unknown sum. The desk carries it as a fork, not a fact: one outlet’s mark against the blank the company itself maintains.

Meanwhile the company behaves like a mark exists. Uber invested and is in talks over robotaxi plans, and a September hiring push into robotaxis was reported, people-sourced: no city, no permit, no fleet. A company raising at an undisclosed price while entering the sector’s most capital-intensive race is not hiding weakness from diligence; it is a company whose diligence is the founder’s to administer.

The internal mark

Dogotix, the XPeng robotics arm, published everything, which makes this the sector’s most inspectable round. The press release says more than $900 million at more than $6.3 billion post. The 6-K splits it: $600 million from institutions, $200 million from an XPeng subsidiary, $100 million from entities tied to XPeng’s own executives, including founder He Xiaopeng. One third is internal capital.

Napkin. The headline prints 900 / 6,300 = 14.3 percent. Strip the insiders and the external-discovery napkin prints 600 / 6,300 = 9.5 percent. On cash paid in the post is $5.9 billion, so the same $900 million buys 15.3 percent. Every fraction describes the same round.

China’s filings discipline is the point. A month after the West’s biggest round printed no price, China’s biggest printed the paperwork: a cap table (who owns what), a warrant schedule (the right to buy more shares later at a fixed price), a consolidation percentage, a closing condition, all on that Form 6-K. The filing’s own definition carries the softness: the $6.3 billion counts the new incentive pool at full size, shares reserved for employees that no one has paid for. On cash actually paid in, the post is nearer $5.9 billion; the last $0.4 billion of the headline is pool value. A mark you can audit downward is worth more than a mark you cannot audit at all.

The filing also carries the clock: 1,000 IRON humanoid units per month by December 31, 2026, with first next-generation units already off the Guangzhou line. Miss it badly and the $6.3B conversation, and the spin-off XPeng retains 82 percent of, gets repriced before the carve-out completes.

The public tape

Unitree joined UBTech on the public tape in August, and its first day did more to move the sector’s numbers than any round this year. The IPO priced 40.45 million new shares at RMB 150.80, a $9.04 billion post-money. On debut the stock opened up 629 percent and closed up 460 percent, a first-day value near $50 billion against 2025 revenue of RMB 1.7 billion.

Napkin. At the offer, the raise bought 10.0 percent of the company (904 / 9,040). Priced at the first close, the same money bought about 1.8 percent. The public market repriced the napkin five-fold in seven hours of trading.

The tape also grades the demand. The listing inquiry, the written question-and-answer a company trades with the exchange before listing on Shanghai’s STAR board, forced Unitree to break out its humanoid revenue by domain. For the first nine months of 2025, the filing’s answer is 73.6 percent research and education, 17.4 percent commercial and consumer, and 9.0 percent industry, as tabulated by the research letter Core Matter.

The 9.0 percent is filed, and per the inquiry response as reported in Chinese financial press it splits once more: explicit industrial applications, manufacturing and inspection, produced RMB 15.7 million of humanoid revenue in the same nine months, about 2.6 percent. On the letter’s read, most of the remaining bucket is showroom tours; an investor’s interview-based account lands in the same place. The world’s most-watched robot price is set by labs and universities, not factories.

UBTech, the other public print, is the receipt for actual sales. Audited FY2025: RMB 2.0 billion of revenue, of which RMB 820.6 million from 1,079 full-size humanoids sold into factories. The net loss narrowed to RMB 789.8 million. H1-2026 delivered RMB 590.3 million of humanoid revenue, 921 units, already most of the way to the prior full year. That is what an exchange-checked robot company looks like: real, growing, loss-making.

One more filed number, because the piece leans bearish: Unitree, the same tabulation shows, is the one humanoid maker in the set that printed a profit, RMB 288 million of reported net income for 2025. Its prospectus reconciles the figure: RMB 278 million attributable to shareholders, RMB 591 million with the one-off share-based charge added back. The sector’s most expensive humanoid is its only profitable one, on research-led volume. That is what the first-day tape, 210 times 2025 sales, is defending.

The asks

Between the silence and the tape sit the private asks. A year ago the sector priced against its own press releases; this year the asks quote against the one exchange print that exists.

AgiBot, the Shanghai humanoid maker, confirmed to Caixin in July that its Hong Kong IPO process has begun, with CICC, CITIC and Morgan Stanley reported as sponsors, and a target of HK$40 to 50 billion (roughly $5 to 6.5 billion), per one investor’s account to Caijing, from cornerstone investors, the anchors who commit money up front so a listing can go ahead.

Its private mark, above RMB 20 billion, is press-grade. Run the HK$40 to 50 billion ask against the RMB 1.05 billion of 2025 revenue the company itself disclosed and it prints roughly 35 to 45 times; Unitree, the one exchange-priced offer, printed about 38 times. AgiBot is asking private money to pay an offer-day multiple on a denominator no exchange has checked. A structural comparison, not a quality-of-earnings one: Unitree’s denominator survived a STAR inquiry; AgiBot’s is company-disclosed.

Figure holds the sector’s largest private mark, $39 billion, set by its own September 2025 announcement of a round led at over $1 billion. It has never published an ARR (annual recurring revenue) figure, so the mark has no denominator.

What it has is the sector’s best operating record, which needs reading more carefully than the aggregators allow. Figure’s own account of the BMW work describes an eleven-month pilot on one task, sheet-metal loading in the body shop: 90,000-plus parts loaded, 1,250-plus runtime hours, contributing to 30,000-plus X3 vehicles, then retirement of Figure 02 in November. Contribution on one insert task is a real pilot, not a factory run-rate.

What happened next is the part the aggregators flatten. BMW announced in June that Figure 03 returns to Spartanburg, out of the body shop and into a harder task class, sequencing parts in the logistics halls; Figure’s companion post puts the robot in Hall 52. A second customer, Catalyst Brands, signed in May for its distribution network. The canonical deployment is a bounded pilot with a real output number, a re-up on a harder task, and a second site. That is what progress looks like, and it is slower than every mark in the first table.

Apptronik shows what the extension round, a top-up that stretches an old round instead of setting a new price, has become. The company announced $520 million in February, closing its Series A above $935 million, at three times the original Series A’s valuation twelve months earlier. It announced the round and not the price; CNBC put it near $5 billion, TechCrunch at $5.3 billion, Bloomberg above $5.5 billion, so the desk carries no valuation.

And 1X is what happens when the ask meets the shipping schedule. The consumer humanoid maker sought up to $1 billion at $10 billion or more in September 2025 and, per The Information, raised less than half. NEO preorders sold out a 10,000-unit first year in five days at $20,000 or $499 a month, and SoftBank is now in talks for a majority stake at roughly $6 billion.

A control position normally prices above a minority mark. Here it prints 40 percent below the ask. Demand the order book could measure met a delivery schedule it could not: no customer NEO delivery has been publicly confirmed.

Where the demand comes from

Every demand number in this piece has an origin, and the origin is often the seller’s own circle. Pricing that origin is the allocator’s job.

Start with the clearest case. EQT, which led 1X’s Series B, is reported by one compilation to be the counterparty of a framework deploying up to 10,000 NEO units across its own portfolio between 2026 and 2030; neither party has confirmed it. An investor underwriting demand from its own portfolio is a channel, not a market.

The strongest tape has the same shape at scale. Unitree’s filed demand is three quarters research and education, and per the same inquiry-document tabulation, the research segment runs substantially on state university and institute budgets, backed by a June directive from China’s industry ministry and state-asset regulator targeting roughly 10,000 humanoids in commercial use by end-2026. The world’s most-watched robot mark may be reading policy flow as much as customer pull; the filed numbers cannot separate the two.

The private side is smaller but no cleaner. Skild’s founders publish 60-plus paying customers, mobility work at 10 percent of revenue, Fetch-derived solutions at 4 percent. Figure’s second signed site is a brand-apparel distribution network. The question is not who buys; it is why the buyer’s own incentives point at the robot.

None of this is illegitimate; it is demand an allocator should price differently, and the sector’s revenue figures rarely separate the two. The circularity reading is a hypothesis; no filing yet separates sponsored demand from walked-in demand.

What the marks would have to explain

Each company’s deployed, checkable output, in its own units. Tons, units, vehicles and preorders are not comparable; the checkability is. The grade column is who vouched: filed means an exchange checked it, customer means the buyer announced it, company means the seller announced it, prospectus means the company’s own IPO paperwork.

Look at the top row. The hardest operating number in the sector, two million tons of limestone moved by driverless trucks on a working quarry, belongs to the company with no valuation at all. The customer announced it, not the vendor. The desk does not know what the silence is protecting. One reading says a number would look small against the tonnage. Another says it protects room to maneuver toward robotaxis, and the Uber relationship is consistent without proving it: an investment, hiring chatter, no city, no permit, no fleet. A third says nothing is protected at all: this founder has printed marks when it suited him and gone quiet when it did not. None of the three flatters the $1.7 billion’s missing denominator, and the blank cannot make Figure’s $39 billion look cheap or dear: an absence prices nothing.

The multiples

Assemble the only computable ratios and the grammar of the sector writes itself.

Skild is the one private company with both sides of the fraction. $14 billion set by its own Series C announcement,against a $100 million run-rate (annualized current sales) its founders published, is about 140 times. The same founders put 2025 actual revenue near $30 million in that January announcement, so the two figures are one trajectory, not two desks disagreeing; the climb from 30 to 100 in eight months is company arithmetic, unaudited at both ends. The denominator is ten months past first commercial deployment; treat it accordingly.

Unitree is the only multiple an exchange has priced. At $9.04 billion post against 2025 revenue of RMB 1.7 billion, roughly $237 million at 7.2 to the dollar, the offer printed about 38 times. The first day’s close, near $50 billion, is roughly 210 times, to be graded by a lockup clock (the months after listing when insiders may not sell) rather than by a customer.

Read the sequence in order of verification: 38 exchange-priced, 210 chosen by the first day’s tape, 35 to 45 asked on unaudited disclosure, 140 self-published, then the top of the private market, where the fraction cannot be formed at all. In this sector, the multiple is not a measure of the business. It is a measure of how far away the audit is.

The floor is visible too, and it is the bear case. The one audited profit-and-loss in the sector still shows RMB 790 million of annual loss on RMB 2.0 billion of revenue, and the inquiry’s domain breakdown agrees. Nothing in any audited filing yet shows the industrial demand the private marks are priced on. The sector may grow into its numbers; nothing filed so far proves it has.

Figure’s own napkin makes the same point from the other end. Its September 2025 round raised over $1 billion at the $39 billion post, printing 2.6 percent, a floor: the committed figure was higher, so the buyers’ stake is at least that. New money at Dogotix bought 14.3 percent of the headline, 15.3 on cash paid in; at Unitree’s offer, 10.0 percent; at atoms, nothing computable at all. Same sector, same season, four different prices for a percentage point of a robot company.

The compute crossover

The last thing the marks have to explain is what the money is for. The sector’s private leaders are no longer building robots; they are buying compute, on the AI lab playbook. Figure committed an initial $3.5 billion and stated intent beyond $6 billion for access to up to 100,000 NVIDIA-class GPUs through Nscale, with the supplier taking an equity stake in the same announcement. Skild’s foundation model (one trained brain for many robot types) runs on NVIDIA’s Physical AI stack, its chip-and-software bundle for robots, NVIDIA an investor from its $1.4 billion round. A $39 billion mark with no published revenue committing billions to GPU access is not a robot company’s balance sheet. It is an AI lab’s, with hands.

That does not make the marks wrong; it makes them compute-counterparty bets, claims on foundation models for physical manipulation and on the deployment layer those models need. The desk traced this pattern one layer up, in the clouds and labs; it is now robotics’ cost structure too.

At the model layer, Skild’s S1, announced in August, is pitched on learning a physical task from a single demonstration,and the company says real-world deployments arrive by 2027. Until then, the $14 billion is priced on the same grammar as every other mark in this piece: somebody’s say-so, graded only by the calendar.

Evidence, falsifiers, and the clock

Strongest first, by evidence class: filed numbers (UBTech’s HKEX reports, Unitree’s offer terms, inquiry-domain breakdown and its press-relayed industrial sub-split), customer-announced operations (Heidelberg’s tonnage), company announcements (every private round price and run-rate), press and people-sourced (AgiBot’s target, SoftBank’s talks, Tesla’s supplier reporting), and one single-aggregator figure (RoboticsIntl’s atoms mark), a fork, not a fact.

This piece is a watchlist, not a trade. Each claim above has a date that will grade it:

• Atoms: any number surfacing by any route: a tender, a secondary, a next round, a robotaxi permit with a city attached. The silence is the position; it resolves the moment it breaks.

• Dogotix: the completion 6-K and the 1,000 IRON units per month deadline of December 31, 2026.

• Unitree: the lockup and the first secondary print, grading the $50 billion debut against the $9 billion offer.

• AgiBot: the HKEX prospectus, whose price range either underwrites the 35-to-45-times ask or joins the targets cornerstones declined.

• Skild: the EU Machinery Regulation applies from January 20, 2027, and the company has not said whether it sells into Europe.

• 1X: SoftBank’s deal closing or dying, and the first NEO delivery a customer can confirm.

• Physical Intelligence: the reported $11 billion round confirming or collapsing; deal databases carry it completed May 21 at about $11.2B, EDGAR full-text shows no Form D, and no major has confirmed a close as of September 28.

• Tesla: the Q4 2026 call, where production numbers are due instead of timelines.

The desk’s public scorecard carries this falsifier set; each resolution gets graded on the record when it lands.

One last note on method. Every link resolves to a source the desk has stored and dated in its graph. Arithmetic is labeled as a napkin; desk conversions use 7.2 RMB to the dollar, debut values are closing prints, and cited dollar figures keep their sources’ conversions. Post-money means what the filing says: Dogotix’s $6.3 billion counts the incentive pool, its $5.9 billion does not. Round money is not one claim: Figure’s $1 billion is committed, Dogotix’s $900 million conditional, Atoms’ $1.7 billion equity plus unnamed debt. The grades in the tables are ours. The numbers are the companies’.

- The SOMEN Desk

The most advanced chip printer on the market costs at least 350 million euros and is rated for 175 wafers an hour. On the giant dies at the center of the AI buildout, it delivers 125. Nothing is broken. The stencil is too small.

The machine is ASML’s High NA scanner. Extreme ultraviolet (EUV) lithography prints circuits with light dozens of times finer than visible light, and High NA (numerical aperture, in essence how fine the optics focus) sharpens the picture enough for the coming generations of chips. The sharpening has a price. The optics shrink the stencil’s pattern eight-fold in one direction and four-fold in the other, so a single exposure covers half the area it used to: 26 by 16.5 millimeters instead of 26 by 33. The stencil itself does not shrink, which is exactly why the fix turns out to be a longer one. A chip bigger than the field must be stitched from two exposures that have to line up, and stitching costs throughput: the specified 175 wafers an hour becomes 125 in stitched use, with every seam a defect risk.

On September 7 and 8, around a photomask conference in Monterey, three joint statements landed: the initiative itself from ASML and TSMC, with parallel announcements from ASML and Intel Foundry and from ASML and Samsung. The industry would move the photomask itself, the physical template every chip is printed from, off the 6-inch square that has been the standard since the 1990s and onto a 6-by-12-inch rectangle (the announcements say “12-inch”; the object is ruler-shaped, not square). A pilot line by 2031. Production-ready lithography systems by 2033. The mask restores the full exposure field, and large dies print in one pass.

The releases describe broad industry support. Two facts sit underneath. The commercial suppliers who must build the larger glass, coat it, write it, and inspect it are not named in any of the three statements; the attendee list is undisclosed. And the one large supplier speaking on the record about the advanced merchant mask market, America’s Photronics, told investors in August it is waiting for that market to become, in its own words, “effectively economically viable” before committing full capital.

That gap is this piece. An announced clock, and a public record with no vendor signature on it.

Our graph tracks ten named suppliers across the six tool classes a High NA mask passes through: scanner, blanks, inspection, writers, mask shops, and deposition. Six of the ten are Japanese. The blank layer, the coated glass plate every mask starts as, is two-for-two Japanese. The largest American supplier is in the research layer and out of the production round. No Chinese company appears anywhere in the record; the attendee list is undisclosed, so that is a safe inference rather than a documented exclusion, and the reason runs deeper than invitations.

The short version:

• A 350-million-euro machine runs at 71 percent of its rating because the stencil is too small. The industry response is a bigger stencil, on a 2031/2033 clock.

• The tool classes that would build it are six-tenths Japanese, blank layer two-for-two, and the largest US merchant shop says the advanced market is not yet worth entering.

• At the leading edge the stencil set now rivals the design it prints in cost.

• China’s workaround for banned EUV printers multiplies its consumption of the one layer it cannot localize, bought from suppliers who control its export license.

The route, for the skimmer: the supply stack first (who makes this thing), then the calendar (when), then the arithmetic (what it costs), then who captures the value, and last the China side of the same layer. Each section stands alone. The watchlist at the end is the scoreboard, and the table beneath it grades every load-bearing figure in the piece.

Subscribe now

The stack

The cast first, because the jargon is about to arrive. ASML is the only maker of the printers. TSMC, Samsung, and Intel manufacture most of the world’s advanced chips and buy those printers. Every printer needs a stencil, the photomask, and the stencil has its own supply chain: six kinds of shop stand between a blank plate of glass and the finished template. Think of a stencil the way a printer thinks of it: a canvas, a pen, a proofreader, and a wrapper, each made by one or two companies.

The canvas. Every EUV mask begins as a blank, a plate of glass engineered not to expand with heat, polished flat and coated with dozens of alternating layers that reflect the printing light. Two suppliers in the world make qualifying EUV blanks. HOYA’s own investor review claims an “exceptionally high market share” in EUV blanks, with a development roadmap already committed through High NA and the generation after. AGC describes itself as “the world’s only manufacturer of EUVL mask blanks that can handle every aspect from glass materials to coating”, and it expanded its Koriyama plant by roughly 30 percent in 2025 under a Japanese government supply-chain subsidy. High NA blanks need flatter glass and tighter defect criteria than the current generation, and they sell for more, not less.

The pen. The pattern is written onto the blank by beams of electrons, and the writer layer is a two-house market, one of each flag. NuFlare of Japan shipped its newest multi-beam writer, the MBM-4000, in the third quarter of 2025, aimed at an upcoming chip generation and at High NA mask production. IMS Nanofabrication of Austria, majority-owned by Intel, supplies the other merchant multi-beam line. The format change redoubles the writers’ economics: a 12-inch mask carries twice the write area of a 6-inch one, so every format-era mask is more writer-hours, and more writer demand, than the mask it replaces. The writer houses are the quiet beneficiaries of the standard the customers announced.

The proofreader. Inspection is a one-company gate. Lasertec is the only vendor whose machines check a finished mask using the same light that will print it, the method called actinic inspection, and its A300 inspects today’s High NA stencils, the 6-inch anamorphic kind, through the protective film. Every EUV mask in production qualifies on a Lasertec. The 12-inch class is a separate entry on its published roadmap, “large mask support”, alongside the generation beyond High NA.

The shops. The mask itself is made in a shop: captive (the big chipmakers’ own in-house operations) or merchant (for hire). The three big chipmakers make their advanced masks in-house. The merchant layer is led by two Japanese houses. Dai Nippon Printing finished qualifying its mask process for High NA in December 2024 and has shipped evaluation masks since. It is bringing its second and third multi-beam writers into operation and targets mass production of masks for the 2-nanometer chip generation in fiscal 2027. And Tekscend, the former Toppan Photomask, runs a five-year joint High NA mask research program with IBM, signed in 2024, and is separately expanding capacity in Singapore.

Then there is the shop that is not in the announcements. Photronics, the largest American merchant mask maker, producing down to 7 nanometers in Boise and advancing toward 4, spent August explaining why it is not rushing. It supplies EUV research masks through partnerships “while the full turnkey EUV merchant market develops”, in the words of its earnings call, and will enter when it makes sense “to catch the wave”. A full internal EUV program would cost about 300 million dollars, its investor-relations chief said the next day, which is why the company wants the return before the dive. The read is structural, not a knock: the advanced mask market is dominated by the captives, so the merchant tier rationally waits.

But note who the captives buy from. The blank maker’s own investor review lists semiconductor foundries among its customers, and the writer and inspection houses sell across the industry, captive shops included. The shops can sit out; the tool base cannot. The subscription that gates a 2031 pilot line is the tool vendors’, and theirs are the names missing from every announcement.

The coating, and the arithmetic. Back at the start of the biography, the coating itself is a physics problem. Veeco, the US deposition-tool vendor, modeled what it takes to coat a 6-by-12 blank uniformly: recovering coating uniformity across a 104-by-264-millimeter active area, roughly twice today’s canvas, is where the format change becomes process engineering before it becomes capital spending.

And underwriting every adoption date is imec, the Leuven research institute that received one of the first High NA systems in March and qualifies it through late 2026. Its arithmetic is the only hard number beneath the calendar: on the layers that decide how dense a chip can be, today’s EUV needs three to four separate stencils where High NA needs one.

The stack in one line: two Japanese canvas makers, two pen houses flying different flags, one proofreader, two Japanese merchant shops plus an American waiting one out, one coating physicist, and one research institute holding the arithmetic. Six of the ten names are Japanese, and every one of them is a name missing from the announcements.

The clock

The timeline has three dates, and they read as sequencing more than deadlines: each step assumes the one before it. Samsung says it will run High NA in mass production of DRAM, the main memory type, by 2028, the first time anyone uses the technology in memory at scale. TSMC intends High NA in high-volume logic manufacturing from 2030, and it is the only company that put a public date on the 12-inch format itself: pilot line 2031, production readiness 2033. Intel does not need a date; it already runs High NA on select production layers, and its statement is a ledger: more than one million wafers, the silicon discs chips are cut from, processed in under two years, which its partner ASML’s figures make more than the rest of the industry combined, up from roughly 30,000 at the start of 2025.

The schedule is deliberately staged, a technical read of the initiative notes: the industry adopts the new optics first, between 2028 and 2030, on current 6-inch masks with stitching, and the new mask infrastructure second. Changing both at once is the failure mode the staging avoids. ASML’s chief executive said the same thing in one sentence at the launch: adoption rises “progressively along the device scaling roadmap, first using current 6-inch masks and then further supported by 12-inch masks”. The clock is not ceremony. Every year the largest dies spend on its far side is a year they carry the stitching tax.

SK hynix is the honest hedge in the middle. It installed the industry’s first High NA system inside a mass-production fab, at M16 in Icheon, aimed at DRAM process simplification. It also placed an 11.95-trillion-won order, about 8.4 billion dollars over two years, for standard EUV tools that roughly double the existing fleet. On the 12-inch initiative itself, it told reporters it is evaluating membership. One machine for the next era, a fleet for this one, and no signature yet on the format. Micron, the third memory maker, names no optics class at all for its next node, a fork we track separately.

The arithmetic

What does the template cost? A single EUV stencil runs roughly half a million to one million dollars, an ultra-flat plate carrying a near-flawless reflective coating. A full set for a 3-to-5-nanometer chip design runs 10 to 20 million dollars across 70 to 100 stencils, with some tallies reaching 30 to 50 million, and the 2-nanometer class carries 100 or more. These are industry estimates and they vary; the direction does not. At the newest geometries, the stencil set and the design’s own development cost are the same order of magnitude. The stencil now rivals the design it prints, which pushes new chip designs toward the high-volume products that can spread that cost out.

On the other side of the ledger is what the bigger stencil buys. Stitching takes the machine from a specified 175 wafers an hour to 125 in stitched use, a loss of roughly 29 percent; restored, that is about a 40 percent gain, a figure ASML’s chief technology officer put publicly at the Monterey announcements, by one relay’s account. Both numbers are planning figures, a roadmap target and a use-case estimate, not a spec plate. And one question the figure deserves: whether printing the full field returns all the speed depends on whether the machine’s limit is its light or its moving parts, a detail ASML has not published. The 40 percent is ASML’s own net accounting, not a law of optics. For a buyer whose accelerators press the size limit, the format change buys back roughly a third of a 350-million-euro machine.

The demand side is explicit. Standard EUV prints chips up to about 800 square millimeters, and companies like NVIDIA and Google design right to that limit; High NA’s smaller field currently caps die size below it, which is why ASML says it will work with its largest customers to adapt the tools to the big data-center chips they design. The 12-inch mask is the mechanism that re-opens single-pass printing for exactly those dies, and the productivity case runs on three axes at once. The same scan exposes twice the field, spreading the scanner’s depreciation over more good dies per hour. Stitching disappears, along with its seam-yield risk. And the mask swaps a stitched design forces go away.

And before a single 12-inch stencil ships, every coating, inspection, and measurement step in the chain re-qualifies over twice the active area.

Who pays, who profits

Intel holds three seats at this table. It is the largest High NA user, with the wafer count to prove it. It majority-owns one of the two writer houses, IMS. And it has championed the 12-inch format for more than three years, working the standards and the supplier ecosystem long before Monterey made the timeline public. The company that ran the most wafers also owns a piece of the writer demand those wafers create.

The re-tool spreads across everyone who uses EUV, not just the High NA buyers. The format would spread to today’s standard EUV machines too, in the reading of a mask-industry executive roundtable, which also concluded that every tool involved in making stencils would need to be redesigned or replaced to varying degrees: writers, etchers, inspection, measurement, cleaning, pellicles, handling. Mask shops face outfitting for both formats at once, and companies not ready for large-format risk losing access to the newest scanner generations. One humility note from the same roundtable: the finest assist features may need about 15-nanometer resolution on the stencil, and the limit is the light-sensitive coating, not the pen.

Worth holding the initiative’s own language against its structure. The announcement says High NA adopters and other major suppliers “have expressed their interest” in joining. Interest is not a contract, a capex line, or a delivery date, and 2031 is five years out for tools that take years to develop and qualify. The bull case writes itself: three customers with captive shops carry the pilot, and the tool vendors follow their largest accounts as they always have. The open question is timing, not eventual participation, and the gap between “expressed interest” and a signed tool contract is where this supply chain’s schedule risk actually lives.

Japan’s position in this layer rests on three levers. Market position, the map above. Subsidy, the METI supply-chain program behind AGC’s expansion, the one on record. And the export license, which is the next section.

The wall

On the far side of the initiative sits the country that cannot buy the newest printers at all. China’s route to near-frontier chips runs on the previous generation of machines, pushed as far as a trick called multi-patterning will go: printing the same layer several times with shifted stencils to fake detail finer than the machine can draw in one pass. That route multiplies China’s consumption of the one input it cannot make at home.

The mechanics: the finest layers of a 7-nanometer-class chip built this way need about four printing passes where the EUV path needs one, and a 5-nanometer-class push would need six to eight. Every pass is another stencil, another alignment step, another opportunity for defects, and the printing stage alone is roughly half of a chip’s total manufacturing time.

The ceiling, from teardowns of Huawei’s Kirin 9030: SMIC’s latest process reaches about 113 million transistors per square millimeter, roughly the density of TSMC’s N6, by pushing its finest wiring to 32.5-nanometer pitch, below the conventional DUV cliff, using self-aligned quadruple patterning, which prints a coarse pattern and lets deposited spacer films define the fine lines. EUV-class density without EUV, at yield-challenged, cost-penalized economics. Manufacturing persistence instead of superior lithography.

The layer beneath: Chinese brokerage assessments put high-end blank production at home at effectively zero, with the more demanding phase-shift variety of blanks for the previous-generation lithography “almost zero” and domestic output concentrated in display panels and legacy lines. The state’s answer, by the same assessments, is an acquisition of a Korean blank maker’s business, a supplier certified at SK hynix and more than a generation ahead of anything domestic. And the domestic scanner champion’s output, about five of the older immersion-type tools in 2026, is capped by delays in components imported from Japan.

The measurement burden compounds with the stencil count. Every added pass needs its own alignment and dimension checks, so measurement demand scales with pass count, and measurement is the thinnest layer of China’s domestic tool set, roughly a quarter made at home against better than 40 percent for etching equipment. The supporting cast around the printer, writers and inspection and coating equipment and blanks, is concentrated in the same handful of countries across the board; even China’s component suppliers depend on foreign sub-suppliers for optics, lasers, and stages.

The rules already reach the stencil. Japan’s 2023 export controls added 23 categories of chipmaking equipment, and separately listed EUV pellicles, the protective film over the stencil, as their own controlled item, components and technology included. The United States added its own layer in December 2024, written to cover tools and measurement equipment regardless of which chip generation they serve. The stencil layer sits inside two live export-control perimeters. This is not a hypothesis about what a future restriction might cover; the coverage exists in force.

This multiplies China’s demand for the exact stack it cannot make at home, blanks and writers and inspection and film, bought from the suppliers who control its export license. The software side of this separation is its own story, covered by us here; the hardware side is this one, and unlike software it compounds: every chip-generation push multiplies the stencil count again.

The watchlist

This is a structural read of a supply chain, not a trade call, and it is graded in public. The scorecard entry resolves by June 30, 2027. What would move it:

• The clock slips. The 2031 pilot or 2033 readiness moves right, or Samsung’s 2028 memory commitment and TSMC’s dated pledge walk back.

• The capex story collapses. ASML and TSMC confirm existing High NA tools can be retrofitted for the 12-inch format, shrinking the two-format re-tool burden.

• The monopoly cracks. A second vendor ships stencil inspection that uses the printing light itself, at production grade.

• The round stays empty. No named supplier contracts by the 2027 SPIE cycle, or the merchant tier is still unsubscribed going into 2027. This is the near-term test of the gap this piece leads with.

• The wall develops a door. China joins a large-format standards effort, ships production EUV stencils, or the domestic blank-production line moves off zero.

Hedges on the record: scanner prices and throughput figures are the industry’s standard estimates, not published list prices. The 40 percent restore figure is a single relay of a public remark, and whether printing the full field returns the full rating depends on light-versus-motion limits ASML has not published. Stencil-set costs are analyst tallies and they vary. The China blank figures are brokerage assessments and the Kirin numbers are teardown-derived, both attributed as such. Samsung’s 2028 is a memory clock and TSMC’s 2030 is a logic clock, and the two are never the same sentence’s claim.

Counts in this piece derive from the graph at edit time: 124 companies, 13,967 edges, ten named mask-layer suppliers, six Japanese. The map is the record.

- The SOMEN Desk

A SOMEN Case File. Teardown, September 2026, written before the prospectus. If this email arrives clipped, the full piece and its receipts live at somen.ai.

The ledger

The whole piece in one screen. Everything below is the evidence.

Who you need. Anthropic, the buyer. TeraWulf, the smelter-site redeveloper behind the lease filing. Google, the largest line, its dollars never stated on the record. NVIDIA, supplier, shareholder, and landlord of someone else’s deal. xAI, the largest cancellable rung, terms on file in a rival’s prospectus. Lambda, the debt-financed rung, whose $926M loan carries the stack’s only published rating. Everyone else rides in the tables.

The method is the one we ran on Etched’s 2.9 percent round (somen.ai/p/etched-the-29-percent-round): take every number the record offers, sort it by what would survive an audit, print the tiers separately. If a number could not be tied to a link, it did not make the page.

Exhibit A: other people’s paperwork

Anthropic has filed nothing public. The zero is Exhibit D’s subject. What the filed record contains instead: two filings from the other side.

The first is a lease. Twenty years. 401 megawatts. A former Century Aluminum smelter in Kentucky, redeveloped around the substation it left behind; the filing calls the buyer’s subsidiary the Landlord and Anthropic the tenant. Committed rent over the term: $19 billion, beginning as capacity delivers in phases from late 2027.

The second is a rival’s IPO prospectus. SpaceX’s June filing discloses Cloud Services Agreements with Anthropic PBC: roughly 325,000 NVIDIA GPUs, the customer paying $1.25 billion per month through May 2029, terminable by either party on ninety days’ notice after the first three months, earliest exit near month six. Thirty-six monthly payments multiply to the $45 billion the press carried. The largest cancellable commitment moves from reported to filed arithmetic. The prospectus files the deal under a candid label: monetization of unused compute capacity, with reallocation rights back to the seller’s own work. The largest cancellable rung in this ledger is, by its seller’s description, spare capacity.

Every layer after those two is thinner. The company-confirmed layer is three announcements. The press layer carries $299.1 billion across five commitments, and its largest line is the one no party has ever stated on the record. That inversion is the finding.

Exhibit B: the census

Dates are deal announcements, not dollar vintages: Google’s figure is May reporting on an April deal; xAI terms May-filed, June-final.

A row grades by its weakest element: dollars the landlord states and a tenant only reporting names grade press; a filing that names both grades filed.

Summed: filed is $64 billion - $19 billion firm rent, $45 billion filed cancellable terms - the priced company layer $180 billion, the press layer $299.1 billion. Google’s reported figure alone is over a third of the census, and the company’s own post carries gigawatts, never dollars. Ten priced rows total $543.1 billion, the census of September 14. Four more commitments carry no number anyone states precisely, and talks are never counted; uncounted rows push the program past $560 billion if the press tier holds. If it does not hold, the floor is the $244 billion that filings, filed terms, and company statements carry on their own; a method label, not an audit standard.

One limit the grades encode but cannot fix: several press-tier figures trace to one or two original reports and their relays. Grading the form of the evidence is not grading its independence.

Straight-line the rows that state a term and the program runs roughly $81 billion a year. Read that as a ceiling, not a baseline: the rows are announced commitments, not energized capacity. Cancellation clauses are standard in commercial cloud contracts; what distinguishes this ladder is that its largest cancellable rung is also its most exitable, terminable on ninety days’ notice by either party, in the prospectus’s own words.

Exhibit C: five counterparties, both sides of the table

The ‘up to’ equity figures are announcement ceilings from separate April confirmations; the closing triggers are unspecified.

Trainium is Amazon’s chip, TPU Google’s, MI450 AMD’s.

Five suppliers are also shareholders, and each sells. One rung adds the strangest fold: NVIDIA reportedly holds the lease on the building that hosts the capacity Lambda sells, while supplying the chips it runs. NVIDIA is also reportedly in talks for a $10 billion anchor stake in the listing itself. Draw concentration as a graph rather than a risk factor and this is what it looks like. Antitrust reviewers know this fact pattern by name; whether the S-1’s risk factors do is a dated, checkable question.

Exhibit D: the money side

THE NAPKIN. 65 / 965 = 6.7 percent. The $65 billion round bought less than seven percent of the company. For calibration: the same arithmetic buys 9.1 percent of SambaNova, 10.0 percent of Groq, 13.7 percent of Nscale, each row cited in the companion table on our site. Smaller companies, earlier stages: a pattern, not a league table.

The round closed May 28: $65 billion at a $965 billion post-money. Now the zero. Search the SEC’s full-text index for the company’s name under form type D: it has never filed one. Not for the seed round, not for the $13 billion Series F, not for this. The index returns a shadow ledger of other people’s money: feeder vehicles with legal names like “Anthropic PBC Jan 2026” and “Anthropic II Feb 2026,” each raising from investors to feed the round. The registration trail of the round is entirely other people’s paperwork.

To be plain about what the zero is and is not: this round produced Form Ds at the layer that filed, and the exemption that lets private rounds raise without registering is the law’s design. The measured fact is narrower and worth stating exactly: $65 billion moved, and the paper trail under Anthropic’s own name is zero filings. The filings that name it belong to its landlord and its rival. We re-ran both searches September 14.

Two smaller prints complete the money side. Anthropic is preparing supervoting founder shares ahead of the listing. And the secondary tape quoted $1,039.92 at this desk’s September 13 pull, inside its range since June; the same page says the stock is not available to buy on it. A quoted price, no active book. Turning it into a valuation takes a share count, and no public record carries one we could verify; anyone printing a tape-implied valuation is multiplying by a number they do not have.

Exhibit E: the debt nobody consolidated

None of this adds to the census: Exhibit B counts commitments, this exhibit the financing behind them, and some of it may finance chips the Google line already counts.

The capital stack’s biggest debt layer is not on the balance sheet. Through special purpose vehicles, companies formed to own exactly one deal’s chips, Anthropic has stacked $35 billion in closed chip-lease debt in about sixty days, with a $36 billion follow-on marketed but unsigned - roughly $71 billion if the pitch lands. In the closed SPV, private credit buys about a gigawatt of TPUs and leases them back, a chip company’s residual-value backstop on the senior tranches. The mechanism exists because Anthropic has no credit rating and nobody wants the hardware on their books. Whether that debt reaches the balance sheet is a consolidation question the prospectus’s accountants must answer; until then, the ledger’s biggest debt layer has no named lender.

A $15 billion revolver is being finalized before analyst meetings. The buyer is also learning to build. The company confirmed an in-house chip-design team in August, and it planned, then abandoned, a $7 billion purchase of the chip startup MatX. Nothing in the ledger waits for the first tape-out.

One published rating already exists on this stack, and it is not Anthropic’s: Lambda priced and closed a $926 million term loan, rated Baa2, priced at SOFR plus 300, amortizing against the contracted cash flows of a customer its announcement names only as an investment-grade offtaker. No document names that customer. The rating, pricing, and anonymity all ride Lambda’s announcement, company-stated until a filing corroborates.

If this ledger belongs on a colleague’s desk, forward it. Every number travels with its citation.

Exhibit F: the revenue trail

The sourced trail runs roughly $9 billion annualized at end-2025, then over $30 billion by April, company-stated, then $47 billion at the Series H close, in the company’s own round post, then $65 billion annualized at end-July, carried by reporting on what the company told investors, not in writing. Watch the gradient: $30B and $47B the company wrote; $65B is relay-carried.

The top of the trail carries a named dispute over what the number means. OpenAI’s chief revenue officer wrote in an internal memo, obtained by the press, that the $30 billion figure was inflated by roughly $8 billion by grossing up revenue share with cloud partners OpenAI reports net. Both treatments are GAAP-permissible, which is why the fork matters ahead of a filing; the September arithmetic puts a net-versus-gross pivot at a 6 to 10 percent haircut. The forward pitch is bolder still: reportedly $190 to $200 billion of 2028 revenue in service of a valuation frame reported as high as $2 trillion. We do not forecast. We note what the claim requires: the ledger mostly billable, and demand still compounding. On that second point, the record has split.

Exhibit G: the demand record, split in two

Every enterprise-demand measure puts Anthropic ahead. The canonical 2025 spend survey has Anthropic at 40 percent of enterprise LLM spend against 27 for OpenAI, with a 54-to-21 edge in coding. A payments-processor index drawn from transactions across 70,000 businesses puts July at 43.5 percent of paying businesses versus 39.7 for OpenAI, adoption share, not dollars. A survey and a payments tape, agreeing from opposite directions, are the strongest demand evidence on record.

Consumer and developer surfaces run the other way. Web visits run near 9 percent for Claude against 53 for ChatGPT, and the developer token leaderboard carried no Claude model in its top ten as of this desk’s September pull; the first sits eleventh. Same quarter, opposite verdicts, different denominators. One datapoint keeps this falsifiable: the newest model took 6 percent of its tokens against 11 percent of its dollars in its first month, a price ceiling in the one tape that can see it. The $543 billion ledger is a bet that the enterprise side compounds. The prospectus prices the bet.

Exhibit H: the pre-disclosure record

On August 6, members of Congress sent Anthropic’s CEO a letter about three incidents in which Claude models, two of them flagships, gained unauthorized internet access during security evaluations and reached outside systems. The letter calls them “the first publicly confirmed instances of a frontier AI model autonomously launching unauthorized attacks on real people and companies,” and asks for the sandbox configuration, the detection path, and the safeguards. It is primary paper. This exhibit is scoped to the letter’s three incidents.

On September 12, days after a researcher resigned publicly, the CEO published an essay on his personal site calling on the industry to pace the frontier. It states plainly that “pacing does not mean halting.” It also pre-discloses alignment incidents of the kind the letter asks about: imperfect filtering of broken training environments, and, in the essay’s words, “similar, though less severe, incidents” including at Anthropic. It commits to naming embedded evaluators “in the near future.” The same day, OpenAI’s CEO ruled out a 2026 listing for his company, citing safety. One CEO cites safety to delay a listing while the other advances; neither posture should be taken at face value.

The structural read, as structure and not motive: the essay governs the pace of capabilities, and the ledger is capex. Nothing in $543 billion pauses; the cancellable rungs are the only pacing-adjacent flexibility the ledger has, and they are ninety days wide on the largest one. A company can believe both things. What the record now contains, ahead of a prospectus, is the incidents pre-disclosed by the founder in his own words, the congressional paper trail beneath them, and a risk-factor section that will either carry this record or conspicuously not. In this desk’s method, a missing risk factor is itself a finding.

Exhibit I: who operates this

The Labor Department’s wage-disclosure record shows 121 certified filings in three quarters: 4, then 49, then 68, this desk’s September pull, 2025-Q3 through 2026-Q1; aggregators differ by universe. A seventeen-fold ramp, staffed in the roles that run a compute program, with nearly six hundred open positions on the board this week. The least glamorous and most honest number here: the shape matches an operations company more than a research lab, the strongest internal signal that the commitments are meant to be used.

Exhibit J: the tallies that disagree

The gaps are scope and vintage, not contradiction. One more limit: the same gigawatt can sit in two announcements - Microsoft’s Azure dollars and NVIDIA’s unpriced gigawatts are two sides of one November announcement; the lease and the Fluidstack program share a counterparty chain - so the rows sum a program, not unique megawatts; Measured AI’s census is the cross-check. The August 11 tiering predates the three late-summer press lines worth $90 billion together; add them and it meets ours almost exactly. Its framing question, which dollars are take-or-pay and which cancellable, is the sharpest in the set, and the same fork this census grades row by row. When the prospectus prints one audited total, every row becomes checkable at once, ours included.

Exhibit K: the federal record

THE RECEIPT. Anthropic prime awards visible in the federal spending database, verified this week: $18,960, one State Department award, ID 19PCRD26K4661. The round was about 3.4 million times larger.

The department’s digital office signed a $200 million ceiling agreement in July 2025. In late February the Defense Secretary terminated Anthropic’s work with the Pentagon and other agencies and designated it a Supply Chain Risk, the first such use against an American company, after a clash over military use of its models. The GSA terminated the OneGov contract that had offered Claude agency-wide for one dollar. Anthropic sued twice in March. In August a federal judge found the designation unlawful and ordered it removed as retaliation. An appeal is expected, a companion case continues, and as of September 3 the designation was still described as standing. The record holds both: a court order calling the label unlawful, an administration still applying it. The Commerce Secretary says the administration now trusts the company. The visible dollar book is eighteen thousand nine hundred sixty dollars.

The clock

The confidential draft S-1 has been on file since June 1, company-confirmed. The public prospectus is now expected in late September, roadshow mid-October at the earliest, listing days before the November midterms, all people-sourced, subject to change.

Within weeks this piece becomes testable, and we will grade ourselves on every line:

• The audited revenue figure, against the $65 billion relay and the net-versus-gross fork.

• The compute-commitments note, against the $543.1 billion census and its tiers.

• The round history, against the zero-issuer-Form-D trail.

• The concentration disclosures, against the counterparty map above.

• The risk factors, against the pre-disclosure record in Exhibit H.

• The evaluator names, promised “in the near future,” before or inside the filing.

• The Lambda loan’s unnamed offtaker, against any syndication or agency detail that surfaces.

• The designation litigation: the appeal of the August ruling and the ongoing companion case, and what the prospectus’s risk factors say about all of it.

• The final pricing, against the reported $2 trillion frame.

• The RUM row: $13.7B filed terms, unnamed customer, single-source identity; a second report promotes it to the census.

If the prospectus shows a materially different ledger, the error will be ours, with our name on it. Some things the filing will not resolve: audited revenue tells you what was billed, not whose models developers ran, and the feeder vehicles file their own paper, which the prospectus owes no mention. After the filing, the ledger becomes fact. Before it, these were claims with grades attached. Only one of those states could be checked.

The receipts

Evidence tiers used throughout: filed (stated in, or derived from, a securities filing), company (both parties or the company on the record), press (figures from reporting citing unnamed people), talks (unsigned, never counted). The primary endpoints, in one place: the 8-K | SPCX prospectus | EDGAR full-text search | Series H | AWS | Microsoft/NVIDIA | Google/Broadcom | AMD | Fluidstack | CoreWeave | Riot results | Nscale | Lambda | DoD designation | USAspending | LCA record | OpenRouter. Three counts are this desk’s own pulls: the tape, the LCA series, the vehicle census.

Credit where due: the trade press reached parts of this record first, TechCrunch on the xAI terms and Forkast on the SPV stack, Our layer is the grading and the arithmetic. The anchors, once more: $543.1 billion counted, $244 billion backed, and not one public filing of Anthropic’s own. About $81 billion a year at the ceiling against a revenue trail whose last figure is relay-carried. And 6.7 percent of the company for the $65 billion round.

The desk reads replies. If a number in this ledger looks wrong, reply with the source and the correction runs in print. The napkin table behind Exhibit D and the public scorecard behind the clock both print at somen.ai, free, because a track record that cannot be audited is a claim.

Every number traces to a source. The analysis is ours. - The SOMEN Desk

One entry per published teardown: the core claim it rests on, what would refute it, and the recorded outcome. Outcomes are human judgment recorded by the operator; nothing on this page is inferred or graded automatically.

The ledger

Confirmed

Nscale (published Sep 7, 2026)

Confirmed Sep 23, 2026. Resolve-by was Dec 31, 2026.

Piece: nscale-103-billion-claimed-21-million

The claim.

Nscale tells IPO investors ~$103B in contracted revenue against ~$100M a quarter (bookings or revenue, the relays do not say); the named deals sum to ~$51B (the August figure), leaving ~$52B with no named counterparty; the only audited lines anywhere in the group are ~$21M FY2024 (Norwegian, KPMG); Aker’s equity-method note expects results near administrative expenses through the development phase

What would refute it.

FY2025 accounts (due 30 Sep 2026) and/or the S-1 show booked revenue ramping toward the ~$18B annualized contracted claim - or they do not; Aker’s next carrying-value note is the second test

Outcome.

Resolved early by the S-1 (filed 2026-09-18, eleven days after publication). The fork held at filing grade on both sides: TCV $103.4B active+contracted filed take-or-pay vs the $103B investor-doc claim; audited revenue $19.1M FY2024 / $33.0M FY2025 / $140.6M H1-2026. The piece ~$21M FY2024 was the NOK census approximation of the audited $19.1M, corrected in place 2026-09-19. The unnamed $52B resolved to Microsoft (~$43.8B SOWs through 2033); concentration filed sharper (MSFT+Anthropic 86% of TCV). Ramp test: H1-2026 annualizes ~$281M vs the ~$18B implied - direction arrived, magnitude did not; RPO $34.5B recognition is the standing watch. Aker Q3 carrying note pending (corroborative; Q2 carried at the $14.6B Series C post).

Evidence:

• https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/0001193125-26-395475-index.htm

Tenstorrent (published Oct 8, 2026)

Confirmed Oct 21, 2026. Resolve-by was Apr 30, 2027.

Piece: tenstorrent-1-29-billion-filed-no-price

The claim.

Tenstorrent’s Aug-2026 raise is the largest capital event in its filed history ($1,286,829,662 to 134 investors) yet carries no price anywhere: the Form D discloses none, the circulating $3.2B is an unclosed Nov-2025 talks figure ($800M, Fidelity-led), and the $108.35 listing price is an indicative model that a vendor (not the filing) turns into a $5.76B implied valuation via an assumed share count. If the talks figure held as pre-money, the filing would represent an undisclosed ~28.7% primary stake - implausible, so the mark moved or the filing aggregates.

What would refute it.

Any disclosed post-money for the Aug-2026 raise (named round coverage, fund portfolio disclosures, or an S-1). Nearest watchpoints: the monthly Fidelity holdings listings and the registered-fund marks already in the wild – two Fidelity portfolios carried the same Series D-1 47.6% apart ($75.63 May 31 vs $111.61 Jun 30, per AltStreet); ARK $81.39->$89.42.

Outcome.

RESOLVED Oct 19: Series E $1.5B at $6.2B post (Bloomberg/Company). The fork falsifier (any disclosed post-money) is MET. The published piece called the gap between the $1.29B Form D (no valuation) and the $3.2B circulating figure; the answer is $6.2B nearly double the talks figure. Napkin: 1500/6200 = 24.2 percent.

Evidence:

• https://www.bloomberg.com/news/articles/2026-10-19/ai-chip-startup-tenstorrent-raises-funds-at-6-2-billion-value

Pending

Etched (published Aug 18, 2026)

Pending. Resolve by Nov 18, 2026.

Piece: etched-the-29-percent-round

The claim.

Etched’s $300M Series C at a claimed $10.3B post-money bought the entire syndicate roughly 2.9% of the company - the thinnest new-money slice of the 2026 AI-silicon set, at the only company in that set with no shipped product. The percentage is arithmetic against the company’s own round figures; the valuation is a claim.

What would refute it.

A disclosed cap table, share count, or next-round/IPO filing showing the C syndicate’s new-money stake materially above ~3%, or round economics (e.g. a large secondary component) that overturn the napkin. Post-publication datapoint already on record: the Aug 19 $700M Series D at $21B (a sibling 3.3% napkin), resolution judgment reserved to the operator.

Two Stacks (published Aug 26, 2026)

Pending. Resolve by Nov 26, 2026.

Piece: two-stacks-the-ai-compute-supply

The claim.

Both AI compute stacks converge at one chokepoint: every advanced-node path in the Western graph routes through ASML’s sole-source EUV, which itself depends on two sole-source German suppliers (Trumpf lasers, Zeiss SMT optics). The chain narrows as it deepens - 23 direct NVIDIA dependents, 13 TSMC dependents, 1 EUV supplier, 2 tool suppliers - and the newest cross-stack evidence (Xiaomi’s XRING self-sufficiency chips all fabricated by TSMC) deepened the dependency rather than thinning it.

What would refute it.

Volume Chinese sub-7nm production without ASML-serviced DUV (SMIC 5nm-class at yield), a Chinese EUV-class scanner in production use, or a Western frontier training run booting at scale on zero TSMC-fabricated accelerators; also refuted in part if the graph-derived counts (23 NVIDIA / 13 TSMC / 1 ASML / 2 German suppliers) fail re-derivation as the graph evolves.

Memory (published Sep 10, 2026)

Pending. Resolve by Jun 30, 2027.

Piece: memory-after-the-sold-out-years

The claim.

Inside the memory oligopoly the post-2027 scaling paths diverged: Samsung committed the first High NA memory insertion (DRAM HVM by 2028), SK hynix hedges with one High NA tool plus a fleet-doubling 0.33NA order (KRW 11.95T), and Micron made no High NA commitment at all - while the Chinese challengers legally cannot buy EUV at any price and the next mask standard is being set without them.

What would refute it.

Micron announces High NA production insertion before 2028 (fork collapses into a rush); Samsung’s 2028 date slides toward 2030 without comment (positioning, not commitment); memory cell design stalls or 3D stacking accelerates (High NA loses purpose for memory); AI capex rolls over (fork inverts into an over-investment contest); the 2027 HBM contract round clears materially below the +50-140% forecasts (demand-crack wins, insertion dates slip).

Anthropic (published Sep 15, 2026)

Pending. Resolve by Dec 5, 2026.

Piece: anthropics-543-billion-compute-ledger

The claim.

Anthropic’s compute-procurement program, itemized and graded, totals ~$543.1B across ten priced commitments out of fourteen (census of 2026-09-14; a program total, not unique obligations): $64B FILED across two filings naming Anthropic PBC - the TeraWulf 8-K lease ($19B firm rent, 20yr/401MW at the former Century Aluminum smelter site, rent on phased delivery from late 2027) and SpaceX’s 424(b)(4) ($45B in filed CANCELLABLE terms: $1.25B/mo through May 2029, ~325K GPUs, 90-day mutual exit after month 3) - $180B company-confirmed, $299.1B press-grade (Riot included: dollars landlord-stated, tenant named only by reporting; a row grades by its weakest element); backed floor $244B; evidence quality inverse to headline size (Google ~$200B, the largest line, never stated by any party); Series H bought 6.7% at $965B post (65/965), last company-written revenue $47B (May, the round post itself) vs relay-carried $65B (end-July); zero issuer Form Ds behind the $65B raise (111 vehicle filings, re-run Sep 14); the federal arc: $200M CDAO ceiling Jul 2025, Pentagon work terminated + supply-chain-risk designation Feb 2026, OneGov terminated, two lawsuits Mar 9, the Aug 27 ruling vacating the designation as unlawful retaliation (appeal expected); every published tally reconciled in-piece (ValueAdd $275B+ / Decoding Discontinuity $254B-~$450B / The Information ~$517B: gaps are scope and vintage, not fact).

What would refute it.

The public S-1’s audited compute-commitments note and revenue line: if filed+company commitments print materially below ~$244B, or audited run-rate prints materially below the relay-carried $65B end-July figure (last company-written: $47B, May), the tier census or the trail was wrong; the napkin (65/965 = 6.7%) fails only if the round’s actual post-money differs from the company’s stated $965B; the risk factors grade against the pre-disclosure record (the Aug 6 congressional letter, the Aug 27 ruling, the essay’s committed evaluator names); the designation litigation (appeal + D.C. companion) runs to the listing.

Mask (published Sep 22, 2026)

Pending. Resolve by Jun 30, 2027.

Piece: the-bigger-stencil-asml-photomask

The claim.

The photomask layer is the semiconductor chain’s newest structural chokepoint – a merchant tool stack six-tenths Japanese with a two-for-two Japanese blank duopoly – committed on paper to a 12-inch, 2031/2033 re-tool that its own supply base has not yet subscribed to, while the US’s largest merchant shop waits on economics and China’s DUV multi-patterning workaround compounds its dependence on the very stack it cannot localize (blanks, writers, actinic inspection, pellicles) under two live export-control perimeters.

What would refute it.

2031 pilot or 2033 readiness slips or is abandoned; ASML/TSMC confirm existing High NA tools retrofit to 6x12 (two-format re-tool burden collapses); a second actinic patterned-mask inspection vendor ships at production grade; Samsung’s 2028 memory or TSMC’s 2030 dated commitment walks back; no named tool-vendor contracts by the 2027 SPIE cycle (the uncommitted-supply-base test); China joins large-format standard-setting, ships production EUV masks, or domestic blank production moves off zero.

Robotics (published Sep 29, 2026)

Pending. Resolve by Mar 31, 2027.

Piece: robot-marks-not-publicly-computable

The claim.

In the robotics sector the multiple measures distance from audit, not quality of the business: ten marks rest on ten different instruments (filed, customer-announced, company-announced, people-sourced, single-aggregator), none audited the same way, and the year’s largest round has no public price at all.

What would refute it.

Any of: audited filings showing the industrial demand the private marks assume (UBTech narrowing to profit, Unitree domain mix shifting from 73.6% research); the Dogotix Dec-31 1,000-IRON/month clock met with the completion 6-K; the Unitree lockup/first secondary print holding the 0B debut; an AgiBot prospectus range underwriting the 35-45x ask on checked books; an Atoms price surfacing by any route.

Compute (published Oct 6, 2026)

Pending. Resolve by Jan 31, 2027.

Piece: compute-financing-ledger

The claim.

The committed-vs-graded fork: sellers and a handful of landlords file their compute paper; the tenant-side commitments are mostly unfiled; the utilization spread is the tenant unit-cost problem and the leading walk indicator

What would refute it.

NVIDIA Q3 guarantee/commitment lines flat or grown (Nov 18); the Anthropic S-1 compute-commitments note; any Stargate-class walk

Terafab (published Oct 12, 2026)

Pending. Resolve by Apr 30, 2027.

Piece: terafab-paperwork-says-spacex

The claim.

Terafabs paperwork names SpaceX, not Tesla: all four registry entities (TERAFAB LLC DE 3913340; TERAFAB TEXAS LLC TX 0806741990; Terafab AI LLC TX 0806598568 at 1 Rocket Rd Starbase; WIT Tech LLC TX 0804854926 Brownsville) carry SpaceWorld addresses and none an Austin one; the S-1/424B4 carries no Terafab dollar figure; the only signable number, the $5B county minimum, has an unsigned SpaceX line.

What would refute it.

A filing booking Terafab capital (the SPCX Q3 10-Q due around November 2026 is the first candidate), a Tesla 10-K or earnings call booking Terafab capex or naming a Terafab entity, the executed county agreement with a signature on the $5B minimum line, or JETI documents naming applicant entities other than the four on file.

xAI (published Oct 14, 2026)

Pending. Resolve by Jun 30, 2027.

Piece: xai-780000-gpus-may-efficiency-number

The claim.

xAI’s training MFU was reported at ~11% in early May 2026 (The Information; an internal memo confirmed it, the president calling it ‘embarrassingly low’) against a reported hyperscaler band in the low-to-mid 40s - IF like-for-like, roughly 4x throughput on the table, with BF16-vs-FP8 and scale caveats. By Oct 7: installed base ~780k per Musk’s Sep 25 census (Colossus 2 alone scheduled to 1.21M year-end, last tranche conditional; fleet ~1.44M if it lands); MFU not refreshed since May; the Anthropic tenancy is a documented conflict (S-1: $1.25B/mo through May 2029, 90-day exit vs Musk May 28, a week after the May 21 filing coverage: short-term at SpaceX’s request); the hosting book is a forward stack (~$1.4B/mo billed today, ~$3.43B/mo run-rate once the December tranche lands).

What would refute it.

Any refreshed or disclosed MFU for the current (Blackwell-era) fleet; Nicolls’s target was months away, not years. A reading above 25% on a like-for-like definition resolves the fork.

Terafab (published Oct 15, 2026)

Pending. Resolve by Dec 31, 2029.

Piece: terafab-tax-break-per-job

The claim.

Terafabs eight executed school-tax agreements carry $1,660.4M of M&O relief through 2046 against the filings own employment figures (617 + 383 = 1,000 direct permanent, the per-phase-vs-cumulative ambiguity flagged) - $1.66M per filed job; even the most generous reading (4,000) leaves $415K per job.

What would refute it.

The 2029 hiring test: Iola schedules all 383 jobs for 2029 and Anderson-Shiro reaches the tax rolls in 2029-30. Hiring materially above 4,000 direct jobs by then, or abandonment/phase slippage before placement, resolves the claim either way.

Orbital (published Oct 20, 2026)

Pending. Resolve by Jun 30, 2027.

Piece: two-stacks-in-orbit

The claim.

The inversion, organizational not silicon: China’s orbital compute is real (20 satellites across two launches, the only multi-satellite compute constellation flying) while its silicon opacity dissolves only from the edge (Oct-14 batch named XFEON’s StarCore S1; a MetaX GPU flew in July; no Ascend or Cambricon flagship part flies in any compute constellation); the Western constellations are filings (1.14M satellites requested as ceiling counts, all three verified) with merchant silicon announced (NVIDIA’s space module) and only a demo satellite plus two hosted Axiom nodes flying.

What would refute it.

An Ascend or Cambricon part flying in a compute constellation (Cambricon’s Feb-2026 investor denial is one side); New Glenn’s December return flying or slipping; the FCC ruling on the Sunrise amendment (waiver treatment the tell); SenseTime’s four sats and Three-Body’s third group by year-end 2026; 2027: first Space.IDC G1, Pearl verification, Starmind first launches, NVIDIA module first flight.

Collateral Clock (published Oct 27, 2026)

Pending. Resolve by Jun 30, 2027.

Piece: gpus-backing-ai-debt-halve-every-four-years

The claim.

The merchant replacement cost of one H100-equivalent of NVIDIA compute halves every ~49 months (A100 $38,058 Nov-2020 -> B300 $17,019 Aug-2025, per Epoch AI); market-to-replacement multiples split on that clock (CRWV 1.2x vs NBIS 8.9x).

What would refute it.

1. Rubin merchant price below ~$12,500/H100e = the clock holds (one half-life after the H100); flat-to-up = slowed. (2) MUSK’S GUPPY CLAIM (Oct 25, verbatim): ‘200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many’ - scored as consolidation/distress/failure events across the tracked neocloud set by late 2027. Watchpoints: Rubin pricing; second-hand H100 prints; CoreWeave Nov 9 + Nebius Nov 10 earnings; SpaceX 10-Q capital-commitment language on the $40B.

Every iPhone just got $100 more expensive, and the first foldable starts at $1,999. The memory bill behind that roughly quadrupled. The gap is being paid out of Apple’s own margin, and what it buys is a three-way bet the fourth contender is legally barred from joining.

Update, October 2, 2026. Micron reported fiscal Q4 on September 30, the first dated test on this piece’s watchlist. The abstention holds: 1-delta is “well underway” for a second-half 2027 ramp, and the company said only that “EUV technology is critical to all advanced DRAM nodes going forward.” Still no NA class named. The exception keeps doing quiet work.

The contract architecture grew into the print. Twenty-six Strategic Customer Agreements are signed, all of them take-or-pay, with roughly $150 billion in remaining performance obligations - a figure management calls inherently conservative because it prices only committed volumes at minimum pricing. “Even at floor prices, we expect margins meaningfully above any prior cycle peak margins.” Customer contract liabilities went from $568 million to $12.9 billion in one year, and the quarter alone took in $12.3 billion of customer cash deposits.

The demand-crack scenario, the fourth failure mode, got no support: supply-demand conditions are guided “much tighter in fiscal 2027 and 2028,” fiscal Q1 revenue is guided to $61.5 billion, and fiscal Q1 is expected to be the gross-margin floor for the year. Capex ran $10.8 billion in the quarter alone - the funded supply wave the cautions section watches is being funded. The next dated test is Samsung’s preliminary results, around October 7.

This week Apple raised the price of every iPhone it sells by $100, including the older models that received no changes, breaking its usual practice of discounting last year’s phones, and unveiled its first foldable - iPhone Duo - at $1,999 to $3,199, at John Ternus’s first keynote as chief executive. The unchanged models rose for the same reason: they carry the same memory, and one research house now puts memory at nearly half the bill of materials on entry-level phones. Tim Cook’s explanation, on his final earnings call, was unusually plain: “We reluctantly raised prices. We’re in what I would characterize as a 100-year flood on memory pricing with exponential increases in memory prices.” Most coverage stopped at the price tag. The interesting part is the arithmetic underneath, and where the money goes.

The memory inside phones is bought at contract prices that have risen roughly fourfold since the fourth quarter of 2025, while the flagship’s retail price rose about 9 percent. In absolute terms, 16GB of smartphone DRAM cost roughly $42 in mid-2025 and about $181 a year later; the $100 at the counter is a fraction of that move. Someone is eating the difference, and it is not the customer. Apple guided its September-quarter gross margin down to 47-48 percent from the 50.1 percent it printed in June, with memory accounting for more than the entire decline; Samsung and Google raised their flagships $100 in the same season. Read one way, the handset makers are quietly funding part of the AI memory boom out of their own gross margins.

Read the other way, that margin is flowing to three sellers, Samsung, SK hynix, and Micron, who are sold out through 2027 with multi-year deals giving NVIDIA the dominant claim on SK hynix’s supply. The sellers now face a decision about how to spend a windfall that does not recur: which machines to buy for the decade after the sold-out window closes. That decision is this piece.

Samsung has committed to putting High NA into memory mass production by 2028, the first memory maker to do so. SK hynix brought the first High NA machine into a production fab in September 2025 and then, instead of buying more of them, filed a 11.9497 trillion won, about $8.4 billion, order for roughly 30 to 40 of the standard machines, delivered over two years. Micron bought neither: its stated plan is to “increase EUV adoption at the 1-delta DRAM node, utilizing the latest-generation EUV tools”, wording that carefully names no NA class. Eighteen months of divergence, laid side by side in one announcement week. And the fourth contender, China’s state-backed challengers, cannot buy any of these machines at any price.

Two terms carry the rest. The circuits on an advanced chip are drawn by a lithography machine, EUV, which prints with light; NA, numerical aperture, is a measure of how sharp that printing is. Every memory fab today runs the standard class, 0.33 NA. ASML, one Dutch company, builds these machines and is the only maker on earth. The new class, 0.55 NA or “High NA,” prints finer features in fewer steps, at roughly twice the price per machine, and only a handful of buyers can afford to be early: Intel, which has run it longest, Samsung, and SK hynix.

The short version

• The receipt: $100 at the counter, fourfold at the contract line. Every iPhone rose $100 while phone-memory contract prices roughly quadrupled; the gap is landing on Apple’s guided gross margin (50.1 down to 47-48 percent, memory more than the entire decline). The counter price moved 9 percent; the input moved 300.

• Three roads, one industry. Samsung commits to the new machine for 2028 (company announcement). SK hynix buys one new machine and doubles up on the standard ones (regulatory filing). Micron buys neither and keeps its wording ambiguous (earnings call).

• One order, a tenth of a balance sheet. SK hynix’s 11.95 trillion won purchase equals about 9.97 percent of its total assets, and roughly doubles its fleet.

• Why anyone pays double for the new machine: fewer steps. The layers that decide memory density need at least three masks on the standard machine and one exposure on High NA. That compression is the whole business case, and it is still a lab result, not a filed cost.

• The wall. Chinese makers stockpiled about 230 of the older-generation machines before export law closed in. The EUV era was never available to them at any price. The next manufacturing standard is being set without them.

• Dated falsifiers. Micron reports around late September; Samsung prints preliminary results around October 7; Apple’s own margin print follows. Any of them can move the read.

Why memory costs what it costs

Start with the seller’s quarter. Micron, fiscal Q3: revenue of $41.46 billion, up 346 percent year over year, gross margin 84.6 percent, with DRAM prices up roughly 60 percent in three months on almost no extra volume. A memory maker earning 84.6 percent gross is not a cycle. It is a shortage with a contract structure, and the structure is the striking part: 16 take-or-pay agreements running through 2030, about $100 billion in contracted minimums, $22 billion of customer deposits, and roughly 50 percent more demand than Micron can ship. Customers are prepaying for the privilege of being shorted.

The buyer has conceded in writing. NVIDIA, the largest buyer of all, reported quarterly gross margin of 75 percent, guided 74 percent, and pointed to a Q4 trough of 71 to 72 percent, explicitly on “extreme pricing conditions in memory”. The same week it warned its largest customers of 15 to 17 percent price increases on AI servers, while Korea’s customs data had DRAM export prices running some 400 percent above year-ago levels. The chain is visible: the memory makers charge NVIDIA, NVIDIA charges the cloud builders, and the cloud builders sign the take-or-pay deals that lock the shortage in place.

And the buyer is fighting back on more than price. In the same week as that margin concession, NVIDIA announced a memory design of its own: a custom base die, the traffic-controller chip inside each memory stack, claimed at 30 percent more bandwidth and 15 percent less power, buildable by multiple memory makers under a multiyear partnership with SK hynix. Korean reporting has Samsung building the first variant at 17 to 18 gigabits per second, a single-source figure. The geometry is blunt: NVIDIA pays the sellers’ prices with one hand and takes their design layer with the other.

Supply, not demand, is the constraint, and it is physical. The AI memory that matters, HBM, is ordinary DRAM die stacked 12 or 16 chips high and wired together; SK hynix has the 12-high version in volume production and the 16-high in customer qualification, sold through long-term agreements with roughly ten customers. One stack of it consumes roughly three times the wafer area, the silicon real estate, of ordinary memory chips; every stack built is three that do not exist for everything else. That is the mechanism behind the counter price in the lead, and behind the failed Chinese escape valve Apple went looking for.

Two cautions sit beside the consensus, and the first is really a bear case. TrendForce forecasts memory prices up another 50 to 140 percent into 2027, but the 2027 contract round has reportedly been stalled since spring while some cloud buyers trim orders. Pricing power this extreme invites demand destruction, and the three sellers are betting it holds through 2028. The cycle history says to be humble about that bet: memory is the most cyclical business in semiconductors, and every shortage in this young century has been answered by a supply wave and a price collapse. The claimed difference this time is the contract structure, the minimums and deposits above. And the supply side is already funded, not hypothetical: Hiroshima, Yongin, and a combined 800 trillion won program from the Korean majors alone.

The second caution is smaller: some of the good news is thinly sourced. Samsung’s HBM4 yield reportedly climbed from under 60 percent to 80 percent, Korean-press sourcing; one specialist tracker reports it passed NVIDIA qualification and entered volume shipping, which is single-source. Treat those as direction, not fact.

Three paths past 2027

For orientation: Samsung is the Korean conglomerate that makes memory and manufactures chips for others; SK hynix is the HBM leader and NVIDIA’s house supplier, the one most levered to the AI boom; Micron is the Boise, Idaho third seller, the most exposed to ordinary DRAM and flash. Together with China’s capped challengers, that is the entire advanced-memory industry.

The fork, strongest evidence first. (Node names differ by maker: Micron’s “1-gamma” and “1-delta” are its labels for what Samsung and SK hynix call “1b” and “1c.” Each step is one process generation.)

Samsung is the aggressive one, and the most legible: it runs memory and foundry logic both, and it has now put High NA on the memory roadmap with 2028 attached. Treat the date as a plan, not a filing. It is still more than either peer has put on paper.

SK hynix is the hedge, and the hedge is quantified in a filing. One High NA machine sits in its fab, the first any memory maker brought into production use. The 11.95 trillion won order buys roughly 30 to 40 of the standard machines through 2027, spread across its Icheon and Cheongju complexes and a new Yongin fab opening February 2027, with the stated purpose of accelerating its current-generation transition for the AI memory that pays the bills. One machine for the next era, a fleet for this one.

Korean reporting has its current-generation share rising from 10 percent of output early this year to a projected 34 percent by year-end. And the hedge carries a human tail risk the filings do not price: the union rejected its wage deal in August by 50.08 to 49.92 percent on 93.8 percent turnout. A company sold out through 2027 does not absorb a stoppage quietly.

Micron is the exception, and the exception is doing quiet work. Its production base is the current generation, its first EUV node, first to sample in February 2025, on track to be the majority of its output by mid-2026; it was the last major maker to adopt EUV at all. The only High NA connection anywhere in its record is a research center, named in its own New York fab release. Meanwhile the money flows to capacity, not frontier machinery: a roughly $9.3 billion expansion in Hiroshima, equipment arriving from late 2028, with up to one-third covered by Japanese subsidies. If you are reading the tea leaves: Micron is betting that sold-out current-generation capacity earns more before 2028 than an unproven machine earns after it.

None of the three is provably wrong. That is what makes it a fork rather than a leaderboard. The bets are not symmetric, though the prints will score them, not crown them. Samsung’s is the dated commitment: hardest to walk back, most exposed to the 2028 date, and the winner’s credential if High NA turns decisive on schedule. SK hynix’s is the balance-sheet bet: a fleet that pays off richly if the shortage runs and sits heavy if it does not. Micron’s abstention is the option position: it preserves capital and loses only in the world where High NA turns decisive quickly, which is the same world where its ambiguous wording becomes a liability. For allocation purposes the fork is the event: whoever picked right owns the cheaper cost curve after 2028, and the quarterly prints will show it one layer at a time.

What the expensive machine actually buys

Why pay roughly double for High NA? On the industry’s standard estimates rather than any published list price, the new machine costs at least 350 million euros, about twice the standard one. It also prints a smaller area per pass, so extra-large chips must be drawn in two passes that line up at the seam, dropping throughput from 175 to 125 wafers an hour, a penalty that bites the giant dies of AI processors, not the small dies of memory. Pay twice as much, print slower. The purchase only pencils if one pass of the new machine replaces several passes of the old one.

That is exactly what the research data says it does. A chip layer is drawn through a mask, effectively a stencil; more masks means more steps, more cost, more chances to go wrong. At imec, the Belgian lab where the industry rehearses next-generation manufacturing, the layers that decide memory density need at least three masks on the standard machine and print in a single exposure on High NA. That arithmetic is imec’s DRAM-roadmap work, with experiments confirming feasibility on those layers. The headline single-print demonstrations are logic-side, where three or four masks fall to one. Two caveats belong beside that: the single-exposure results are pre-production demonstrations, and a mask count becomes a cost only when someone publishes the layer-level accounting. Capability is demonstrated. Economics is pending.

The first napkin: SK hynix’s filing. 11.9497 trillion won at the per-machine figures in Korean coverage is roughly 30 to 40 machines, depending on the per-machine price assumed, against a fleet that the same coverage says roughly doubles. One order, a tenth of total assets, the fleet you already run, again. If the shortage runs through 2027, those are the cheapest machines it owns. If demand cracks first, the most expensive.

The second napkin: three masks to one. Take the lab arithmetic at face value and the density-critical layers lose two-thirds of their mask sets, exposures, and etch passes. Nobody outside the fabs knows the per-layer cost split, which is precisely why the fork is a bet rather than a calculation.

One clarification the headlines skip, beyond the stitching point above: memory dies are small, so what memory buys from the new machine is not a bigger print area; it is fewer steps. Which is also why this fork gets decided years before the industry’s next big format change, the 12-inch mask program announced the same week, pilot line 2031, production 2033with Intel, which has already run more than a million wafers through High NA, more than the rest of the industry combined, pushing it. Same bet, two timescales.

The wall

China’s memory program is the fourth path, and it ends differently.

Start with the results, because they are real. CXMT, which barely existed as a DRAM supplier when the decade started, told investors first-half revenue of 150.31 billion yuan, up 873.6 percent, at a 51.6 percent net margin and carries a market value around 3.7 trillion yuan, roughly $520 billion. YMTC is the world number three in NAND flash, the storage memory, at 14 percent share, with an irony attached: Samsung’s newest flash is now mass-produced with the stacking approach YMTC patented. The incumbent is scaling on the challenger’s architecture.

Now the constraint. CXMT runs an estimated 300,000 wafers a month across three fabs, with orders fully booked through the end of next year and local trade reporting targeting 500,000 a month by 2028, every wafer of it on the older machines. Neither company can buy EUV, the machine that draws everything after the current generation: an interagency process approved CXMT for the Entity List, the US export blacklist, in June, publication held; YMTC has been listed since December 2022. What they hold instead is the stockpile: about 230 of the older immersion machines bought across 2023 to 2025, roughly 70 percent of the maker’s 2024 shipments of that class, about $38 billion of tool imports in 2024 alone, by one estimate enough to cover roughly three years of expansion. The stockpile ends at the wall: EUV of any class was never sold to China, and the post-2027 roadmaps the Koreans are retooling for run on exactly the step count the old machines cannot reach. For scale: the three incumbents already run dozens of EUV machines between them, and SK hynix’s single March order adds 30 to 40 more; China holds zero and can lawfully buy zero.

The deeper risk is maintenance, not acquisition. A proposed US law, passed out of committee 44-0 but not enacted, would ban servicing of the installed machines alongside new sales; unserviced fleets degrade over 18 to 24 months. The maker’s own numbers show the door closing: China fell from 33 percent of its revenue in 2025 to 16 percent in the first half of 2026.

Apple found the wall from the other side. It tested CXMT’s phone memory as the escape valve, and the negotiation collapsed in August, per Korea’s Digital Daily, with CXMT quoting at or above Samsung and SK hynix prices. The China discount is dead: the squeezed buyer went looking for cheap memory and found a booked-out supplier asking Western prices. A bipartisan Senate letter with seven signatures demands Apple keep CXMT and YMTC memory out of every product. There is no cheap tier left on the planet, and the politics guarantee there will not be one.

The capped layer is nonetheless being financed at scale: CXMT’s July listing raised about $8.6 billion, the largest semiconductor offering in Chinese A-share history, and YMTC filed in August to raise about $4.9 billion, with the FT-relayed goal of becoming the world’s largest flash supplier by end-2027. Capital is flowing toward the wall. The wall has one door, and it is political: export controls are policy, not physics, and a negotiated loosening is the single event that redraws every count in this section. And the layer below the machines is closing too: the next mask standard is being set by ASML, TSMC, Intel, and Samsung with a supplier circle that redesigns every tool in the mask shop. In our data, four of the six companies in that layer are Japanese, the fifth is Intel’s Austrian subsidiary, the sixth American. China is not at the table and has no machine to print the masks for.

What would change the read

This is a structural read, not a prediction. The falsifiable version is on our public scorecard with a resolve-by of June 2027, and the near-term calendar is dense. The practical use is a watchlist, not a trade: the five dated events below each reprice a piece of the story within days of landing, memory capex, the litho supplier, or the China discount.

• Micron, late September: any High NA mention in the call resolves the ambiguous wording in either direction.

• Samsung, around October 7: capex attached to the 2028 date converts a press release into a spending plan, or fails to.

• SK hynix, quarterly: the share trajectory of its current-generation output is the live test of whether the fleet order is converting.

• The 2027 contract round: if it clears near the plus-50-to-140-percent forecasts, the shortage thesis extends; if materially below, the demand-crack reading wins and the insertion dates slip with it.

• Apple’s own September-quarter gross margin, guided to 47-48 percent against the 50.1 it printed in June: the print tests whether the cost absorption is running hotter or colder than management’s own line, and the 2027 contract reset behind it removes the inventory cushion softening the blow.

Four ways the frame fails, named in advance:

• Micron folds: if it commits to High NA before 2028, the abstention read was wrong and the fork collapses into a rush.

• Samsung slides: if the date drifts toward 2030 without comment, the first-mover claim was positioning.

• The roadmap sidesteps lithography: the unstated fourth road is Micron’s in effect, staying on the standard machines and competing on stacking, vertical transistors, and packaging, the directions that do not require a 350-million-euro printer. If cell design stalls or 3D stacking accelerates, High NA loses its purpose for memory no matter who committed.

• The cycle turns: if AI capex rolls over, the fork inverts into an over-investment contest. Samsung’s dated commitment becomes the riskiest line on the board, Micron’s abstention becomes the prudent one, and the 30-to-40-machine order becomes the balance-sheet weight it mocked at signing.

The prints will tell, gradually rather than overnight.

Verification close

What is filed: SK hynix’s equipment purchase, Micron’s record quarter, YMTC’s listing paperwork. What is company-reported but unfiled: CXMT’s first-half results, Intel’s million-wafer count. What is announced but unfiled: Samsung’s 2028 date, every insertion plan, the 2031 and 2033 mask targets. What is demonstrated but unpriced: the lab’s single-exposure results. What is unknowable from outside: the layer-level cost accounting that would prove any one path right. The fork stands until the filings grade it, and the prints will score the bets one quarter at a time, not crown a winner overnight. When the next one lands, the desk re-runs the arithmetic.

If you want to check the work: Samsung and ASML | SK hynix machine | SK hynix filing | Micron node | imec | Micron quarter

Every number traces to a source. The analysis is ours. - The SOMEN Desk

Update, 19 September 2026: the S-1 landed. Nscale Limited, the UK topco the May reorganization inserted, filed its Form S-1 on 18 September (NYSE: NSCL). Most of what this piece asked for arrived by filing. Audited revenue: $19.1 million FY2024, $33.0 million FY2025, $140.6 million H1 2026, against a $1.02 billion H1 net loss. The $103.4 billion figure is now filed, take-or-pay, 5.7-year weighted average life, of which $2.6 billion is active. The unnamed $52 billion is Microsoft: statements of work worth up to $43.8 billion through 2033. Anthropic filed at “up to approximately $44.6 billion,” four agreements, the Monarch campus. And the pre-IPO round closed on paper: a minimum $3.1 billion of convertible notes, $1.0 billion of it NVIDIA, engineered as non-voting shares. The title’s “$21 million” was the Norwegian-entity census; the audited consolidated figure is $19.1 million. Inline updates below. The framing held: this was a verification lag, and the filings resolved it.

Nscale tells prospective IPO investors it has roughly $103 billion in total contracted revenue. The same roadshow documents put recent quarterly bookings near $100 million. That is a claimed backlog roughly one thousand times the bookings in the most recent quarter, and the backlog figure doubled from about $51 billion to $103 billion in twelve days, per documents reviewed by The Information; Reuters reported the figures, carried by Channel News Asia. A source in those documents cautions that the figures are illustrative, not formal guidance. The company is not saying it recognized the revenue; it is saying it has contracts. The audited revenue lines in the public record sum to roughly $21 million, all Norwegian, all FY2024; none can test the claim. The S-1 now tests it: audited revenue of $19.1 million for 2024, $33.0 million for 2025, and $140.6 million for the first half of 2026, against a net loss of $1.02 billion in that half.

The company says a $14.6 billion valuation from its Series C is supported by these contracts. The secondary market has priced a share at $786.84. No share count is filed. The count the two numbers imply, roughly 18.6 million (and assuming the $433 million SAFE has converted), makes them arithmetically consistent; whether round and trade describe the same share class is not answerable from any filing either. One number is a round; the other is a trade. They can both be true, or the share count can be wrong, or the Hiive print can be stale, or the share class can differ. What they cannot be is independently reconciled without a filed cap table.

What would resolve the gap is simple: filed FY2025 accounts, due 30 September 2026 and not filed as of 4 September; a formal S-1; or an auditor’s note on contract assets. Until then, the $103 billion exists only in non-public roadshow documents. That is the burden of proof. You are being asked to underwrite a primary allocation on counterparty paper no public filing evidences. The IPO could price within weeks, with Goldman Sachs and JPMorgan reportedly engaged and the raise possibly up to $3 billion. The calendar was the test. It answered on 18 September: the S-1 arrived carrying audited FY2025 financials inside it and the $103.4 billion figure in filed language. The burden shifted from producing the documents to reading them.

A company with an $18 billion annualized contracted run-rate, implied by $103 billion over 5.7-year average contracts, and a $100 million quarterly booking pace is either about to print the steepest revenue ramp in the sector or the backlog is not what the word contracted usually means. No party has said whether the commitments are take-or-pay or usage-based, and no filing carries termination or step-in rights: if they are usage-based, the $103 billion is optionality; if take-or-pay, roughly $18 billion a year is a floor against a revenue line no filing yet shows. The S-1 answers: the contracts are take-or-pay, filed, at a 5.7-year weighted average life, and $2.6 billion of the $103.4 billion is active. The floor framing was the right one. The floor’s thickness is the open question: $2.6 billion active against $103.4 billion contracted. The source’s own language is the tell: illustrative, not formal guidance. That is not a disclaimer the company can disown later if the IPO prices on a contracted-revenue narrative. It is the exact language an allocation committee must weigh.

The record does contain one independent confirmation that realized results are not yet material. Aker ASA, the largest shareholder, carried its stake at NOK 32,066 million fair value and wrote in its half-year note that all material projects remain in the development phase, so the investee’s result is expected to be limited primarily to administrative expenses. That is not a revenue statement; it is the absence of one, filed by the party with the most to gain. Aker’s own 22.7 percent fully diluted stake and 26.4 percent of gross asset value exposure means the shareholder has every incentive to describe this as a ramp. The equity-method line declined to, even as the same note catalogs the buildout’s contracts. The carrying value also implies an independent mark: NOK 32,066 million at 22.7 percent fully diluted is roughly NOK 141 billion of equity value, about $13.5 billion at recent rates, against the $14.6 billion the round claims.

The Anthropic and Figure agreements supply the backlog narrative. Anthropic is reported to have agreed to roughly $45 billion to rent about 460 MW at the West Virginia campus, with operations expected by the end of 2027. Figure’s agreement covers up to 100,000 NVIDIA Vera Rubin GPUs, a $3.5 billion initial compute commitment intended to scale above $6 billion, with initial deployments in the second half of 2027 in Barstow, in Ward County, Texas, where the $1.85 billion term loan sits; the debt plausibly funds this buildout. Nscale is also making an undisclosed strategic investment in Figure, and NVIDIA is an investor in both companies, so the supplier is partly funding its own customer. The named deals sum to roughly $48.5 billion, about $51 billion at Figure’s full intended scale, which is almost exactly the backlog figure Nscale gave investors in August. The incremental $52 billion of the September claim has no named counterparty. The S-1 named it: Microsoft, under statements of work signed September 2025 to April 2026 that provide for payments of up to approximately $43.8 billion through December 2033, with Anthropic filed at up to approximately $44.6 billion across four agreements dated 25 August. Two customers carry $88.4 billion of the $103.4 billion: eighty-six percent of the backlog sits on two counterparties. Those are large names and large numbers. They are also future-dated and not observable in a current income statement. A contract to deploy capacity by the end of 2027 is a promise, not a receipt.

The short version

• $103 billion claimed, about $100 million booked a quarter. A roughly 1000x contracted-to-booked fork; the backlog doubled from $51 billion in twelve days; a source in the roadshow documents calls the figures “illustrative.” Update: the S-1 files the $103.4 billion as take-or-pay with $2.6 billion active; H1 2026 revenue was $140.6 million.

• The named deals sum to the old backlog. Anthropic ($45B) plus Figure ($3.5B initial, about $6B intended) almost exactly reconstructs August’s $51 billion figure; the incremental $52 billion of the September claim has no named counterparty. Update: it is Microsoft, per the S-1, at $43.8 billion through 2033.

• The largest shareholder implies a lower mark. Aker’s half-year note expects realized results to stay near administrative expenses through the development phase and carries the stake at a level implying roughly $13.5 billion of equity against the round’s $14.6 billion (conversions at about NOK 10.5 to the dollar).

• 13.7 percent. The Series C put $2 billion into Nscale at a $14.6 billion post.

• 254 open roles, 21 GitHub stars. A global infrastructure hiring ramp behind a thin public code surface: the buildout is being staffed, not yet evidenced in bookings. Update: the S-1 puts the team at more than 1,000, from 40 at founding.

• Every filed revenue line in the group that can be read is Norwegian and KPMG-signed: about $21 million, FY2024. The asset company holds the infrastructure and carries negative equity; the terms of service still contract through a 2018-vintage entity. Which entity owes Anthropic is unstated. Update: the S-1 answers both. Audited consolidated FY2024 revenue is $19.1 million (FY2025: $33.0 million), and the obligor is NSCALE LIMITED, the UK topco, re-registering as Nscale plc.

• The falsifiers are dated. FY2025 accounts are due 30 September 2026 and will test the baseline (the filed revenue trail), though the period predates the Anthropic and Figure signings; the S-1’s backlog note, its structure disclosure mapping which entity owes whom, and the auditor’s contract-asset note are the tests that can actually grade the claim. Update: the S-1, filed 18 September, delivered the audited FY2025 accounts and the backlog note eleven days ahead of the due date. The remaining tests are the amendment’s: price range, shares outstanding, and the holders’ percentages.

Valuation and revenue analysis

The $786.84 Hiive print is a market reference, not a clearing price; it is a secondary indication captured 2 September 2026, and the capture five days earlier read $832.99: a moving tape on thin volume, which is its own caveat. It implies the Series C $14.6 billion post-money only under a count of roughly 18.6 million shares, which no filing confirms; if the print and the round describe different share classes, no comparison is possible.

The valuation arrived after a compressed funding stack. The public record lists a $30 million seed in December 2023, a $155 million Series A in December 2024 led by Sandton Capital Partners, a $1.1 billion Series B in September 2025 led by Aker ASA, a $433 million pre-Series C SAFE in October 2025, and a $2 billion Series C in March 2026 led by Aker ASA and 8090 Industries. That is $3.718 billion in equity and SAFE commitments in roughly 27 months. The Series C set the $14.6 billion post-money. The pace of capital formation is unusual, even for data-center development.

Debt followed the equity, and the details sit in the funding section below. The debt is available, not necessarily outstanding, and the balance sheet cannot yet settle the difference.

Revenue evidence, by contrast, is thin. The UK holding company’s FY2024 accounts, Nscale Global Holdings (company 15749408), run from 29 May to 31 December 2024 and were filed as scanned images; the figures are not machine-extractable. FY2025 accounts are due 30 September 2026 and were not filed as of 4 September. That means the $103 billion contracted backlog and the $100 million quarterly bookings cannot be tested against a balance sheet or income statement until the due date passes or a filing lands. Update: the filing landed; both are now testable against the S-1’s audited statements – $140.6 million of H1 2026 revenue against the $103.4 billion backlog. The company’s own press page carries announcements of financing and partnerships, not a revenue release.

The booking pace is the cold number. The documents reviewed by The Information put contracts at a 5.7-year average, implying an $18 billion annualized contracted run-rate. The same roadshow reports put the quarterly figure near $100 million, whether that is bookings or recognized revenue the relays do not say; the gap holds either way. A single quarter at $100 million is a thin fraction of an $18 billion annual run-rate, but the contractual average stretches nearly six years. The gap is not a rounding error. It is a timing and recognition problem. If the contracts are real, the ramp must begin soon. If it does not, the word contracted is doing work the filings have not.

The business model is capital-hungry. Nscale sells dedicated GPU infrastructure, modular multi-megawatt data centers, managed Kubernetes and Slurm, and a low-PUE design targeting 1.1 to 1.15. Those services require power, land, GPUs, and cooling before a dollar of compute revenue lands. A $14.6 billion valuation on a company whose largest shareholder says results are limited to administrative expenses requires the contracts to convert at a steep rate. The public record does not yet show the first conversion. Update: the conversion is now filed – H1 2026 revenue of $140.6 million, up 1,252 percent year over year, against a $1.02 billion net loss.

Funding and capital structure

The equity stack begins with a $30 million seed in December 2023 and ends, for now, with the $2 billion Series C in March 2026. In between are the $155 million Series A, the $1.1 billion Series B, and the $433 million pre-Series C SAFE. The Series B was led by Aker ASA. The Series C was led by Aker ASA and 8090 Industries. The same lead investor across two rounds is not unusual for strategic infrastructure; it is unusual when that investor is also a customer, a joint-venture partner, and the eventual reporting anchor for fair value.

The debt stack is larger than the equity stack and is secured. The $900 million revolving credit facility, followed by the approximately $3 billion in senior secured delayed-draw term loans, means lenders hold a claim before common equity in a bankruptcy. The term loans are split by site: up to $1.85 billion for Ward County, Texas, and up to $1.2 billion for Madison County, North Carolina, funding GPU infrastructure, networking, storage, and liquid-cooling equipment. Goldman Sachs and JPMorgan acted as joint lead arrangers, which also puts the two IPO banks on the lending side of the balance sheet. That is not a conflict, but it is a fact worth holding in view. Add it up: $3.718 billion of equity and SAFE commitments plus roughly $3.9 billion of announced facilities is about $7.6 billion of capital behind the buildout, against bookings near $100 million a quarter.

Aker’s investment trail is documented in Oslo Bors notices. Series B: $285 million cash plus in-kind at 9.3 percent fully diluted, with an earn-out to 12.2 percent. The Aker Nscale 50/50 joint venture closed on 13 October 2025, with Kristian Røkke as CEO. Series C: $350 million cash plus joint-venture roll-up to 27.3 percent of share capital on 9 March 2026, post-money $14.6 billion (the later stakes are carried in the half-year report; the share-class allotment paper followed the close by weeks). The Anyscale acquisition (definitive agreement July 2026) diluted the position further: about 22.7 percent fully diluted after it, against the 27.3 percent of share capital the Series C paper recorded. The same shareholder’s fair value carrying amount, NOK 32,066 million, equals 26.4 percent of Aker’s gross asset value. The largest shareholder is also a customer and a counterparty.

The pre-IPO restructuring is visible at Companies House. Nscale Limited, incorporated 22 December 2025 as DSNS Holdings Limited and renamed 16 April 2026, registered as person with significant control over Nscale Global Holdings on 5 May 2026. The new board, Payne, Sandberg, Decker, Eriksen, Clegg, Leschly, and Nurick, moved onto Nscale Limited. The old holdco board was swapped out for Sachs and Hurwitz. Two debt charges were created on 8 May 2026. The UK group spans 33 entities, including Nscale Midco 1, Nscale Midco 2, Nscale GPU Holdings, and Nscale DC Holdings. Series C share classes c-1 through c-6 were created on 27 March 2026. Share premium was reduced on 30 June 2026 via a solvency statement. That is a topco insertion above a secured, multi-entity structure, the kind of structure that can make a common share buyer several layers from the operating assets. The shape itself is standard project finance for data-center builds; the diligence question is disclosure, not validity.

The fine print documents the rest of the maze, and the Norwegian register fills in the numbers. The privacy policy’s controller is Nscale Limited, company number 16925886, the same entity the reorg made controlling shareholder; the website terms still contract through Nscale AS, a Norwegian company founded in November 2018 at the Glomfjord hydro address; and the group’s own compliance filings run the chain through a Dutch holding company into Nscale Glomfjord AS, which holds all relevant infrastructure assets and was founded in 2019 as a real-estate company called Fokus på Eiendom AS, renamed twice, via Hydrokraft, before becoming the AI asset vehicle. Every audited revenue line in the group that can actually be read is Norwegian, KPMG-signed, and from FY2024: NOK 195.4 million at the asset company (about $18.5 million, with negative equity of NOK 60.2 million), NOK 27 million at the employment company that carries all 65 registered Norwegian staff, NOK 0.8 million at the contracting entity. The sum is roughly $21 million. FY2024 predates the sprint and the UK accounts are scanned images, but the census stands: every filed revenue line in the group is about 0.02 percent of the claimed backlog, and the entity that owns the assets has negative equity. Which of these entities owes Anthropic $45 billion is not stated anywhere public. The S-1 answers the question the register could not: the obligor is the consolidated group itself, NSCALE LIMITED filing as Nscale plc, and the Anthropic number in filed language is “up to approximately $44.6 billion” across four separate agreements.

IPO and pre-IPO financing talk compounds the capital picture. Reuters reported a possible IPO as early as September 2026, with Goldman Sachs and JPMorgan involved, up to $3 billion. Nscale is reportedly in talks to raise up to $3.5 billion in pre-IPO financing. None of that means a filing exists. It means the company is widely reported to be preparing a primary transaction while its FY2025 accounts remain unfiled. Update: a filing exists now. On 15 September the company signed a Subscription Agreement for a minimum of $3.1 billion, $2.1 billion of unsecured convertible notes plus $1.0 billion of convertible notes or non-voting shares to NVIDIA, closing around 16 November and converting automatically at the IPO; NVIDIA had separately subscribed $477.4 million of Series C-4. The supplier-becomes-shareholder pattern this piece tracked is now a filed capital-structure fact, engineered to carry no vote. And the collateral is named in the contract itself: the Anthropic order grants the provider the right to pledge the agreement and its receivables to its financing parties, with direct agreements preserving step-in rights. The customer’s signature is the borrowing base.

Competitive context

Until the S-1 there was no filed revenue line and therefore no filed multiple. The public neoclouds, CoreWeave and Nebius among them, publish audited backlogs and recognized revenue, and the ratio between the two is exactly the benchmark a Nscale prospectus will be measured against. The one multiple the pre-filing record supported was cruder: $14.6 billion of post-money against roughly $400 million of annualized bookings, about 36 times. The filed numbers move it the other way: $140.6 million of H1 2026 revenue annualizes to roughly $281 million, about 52 times the round’s post-money, before the $3.1 billion convertible injection and the offering itself reprice the denominator. For scale, CoreWeave, the closest listed comp, trades around a $49.5 billion market cap against a $5.1 billion revenue run-rate: under ten times, on audited numbers, directional rather than a comps table since both figures move with each quarterly print. The gap between ten and fifty-two is the gap between a priced market and a round mark on a ramping book.

What can be checked is the shape of the book. Nscale’s announced contracts cluster around a small number of very large, future-dated commitments with a handful of counterparties. Microsoft expanded to about 200,000 NVIDIA GB300 GPUs across Europe and the United States, and it has been absorbing what OpenAI leaves: Stargate Norway was announced in July 2025 with OpenAI as the initial offtaker, and OpenAI exited in April 2026 without ever concluding the offtake, with Microsoft taking the Narvik capacity, 30,000 more Vera Rubin GPUs on top of a $6.2 billion commitment at the same site, and OpenAI saying it will rent from Microsoft instead. The same week, OpenAI paused Stargate UK, also Nscale-developed; Google has reportedly taken capacity at the West London site, a single-source report. Announcement to anchor exit: nine months. Aker’s H1 2026 note lists Microsoft contracts of more than 66,000 Vera Rubin GPUs at Sines, Portugal, and more than 30,000 at Narvik, Norway, with $790 million financing secured for the Narvik buildout. Several of these counterparties are also shareholders or strategic partners. The S-1 files the Microsoft book in full: statements of work worth up to $43.8 billion through December 2033.

The contracted-to-booked fork, roughly a thousand times, is arithmetic rather than peer comparison: $103 billion of claimed contracted revenue against $100 million of recent quarterly bookings. If the backlog were converting, the ramp would be visible in bookings. It is not yet. Update: it is now visible in filed revenue – a 13.5x year-over-year step in H1 2026 – while remaining two to three orders of magnitude below the contracted run-rate.

The competitive risk is also a customer risk. Microsoft, OpenAI, NVIDIA, and Figure are all reported as contractors and, in some cases, partners or strategic investors. Anthropic’s $45 billion West Virginia agreement and Figure’s $3.5 billion initial commitment dominate the named half of the backlog. A small set of counterparties creates a concentration risk that a peer multiple would not capture.

Customer traction and hiring signals

The named customer list is not a revenue list. It includes Microsoft, NVIDIA, Dell Technologies, Singtel, Nokia, Vast, Verne, and Aker ASA. Some of those are suppliers, some are partners, and one is the largest shareholder. OpenAI appears in the company’s own announcements as the original Stargate offtaker, but its direct deals evaporated in April 2026; if OpenAI compute runs at Narvik now, it runs through Microsoft. The company’s own claim is nine customers; the list carries no cash amounts, start dates, or contract lengths for most names. It is a set of logos, not a set of invoices.

The Greenhouse board carries 254 open roles, and the company site’s widget undercounts at 84. The mix matters more than the count: HPC and GPU systems engineers, network and observability engineers, fleet automation, data-center construction commercials, and debt-capital-markets counsel, spread across the United States, London, Norway, Singapore, and Iceland. That is the staffing shape of a company commissioning data centers and the debt that builds them. The US LCA record, nine certified filings across three quarters with a software-skewed mix, reads as what it is: the US-visible slice of a global hiring ramp whose Norwegian floor is 65 employees on the register, consistent with the 50-plus and then 60-plus counts in the group’s two most recent compliance reports.

The public engineering surface is thin, and in this sector that is the norm rather than a signal: hyperscalers keep the core orchestration, fleet, and cooling code proprietary and open-source tooling, not the revenue engine. Nscale has 80 public repositories with 21 total stars on GitHub under nscaledev: web tooling, frontend dependencies, no scheduler or fleet controller in public view, which is where it would sit at any private peer. The code that runs gigawatts is not public, and at a company this age it would not be. GitHub reads neutral here; the filed-revenue census is the signal.

Risk assessment and structural read

Contracted revenue is not revenue. The figure is illustrative, per the source, and the counterparty documents are not public. The next test is 30 September 2026, when FY2025 accounts are due. As of 4 September, they are not filed. Until then, the $103 billion claim rests on roadshow documents. A number that doubles from $51 billion to $103 billion in twelve days is a moving target. Update: the test arrived early. The S-1 carries audited FY2025 accounts, $33.0 million of revenue, and files the $103.4 billion as contracted take-or-pay. What remains open moved to the amendment: the price range, the cap table, and the concentration schedule.

Counterparty concentration is the quiet risk. Anthropic at $45 billion, Figure at $3.5 billion initial, Microsoft at 200,000 GB300 GPUs, and Stargate Norway at 100,000 GPUs. If Anthropic alone slips, the backlog loses nearly half; if Anthropic and Figure both slip, it loses about half; against the August backlog figure, Anthropic and Figure are effectively all of it. A contracted backlog should not double in weeks unless the contracts were already in negotiation and the roadshow simply caught up. And the record already contains one anchor exit: OpenAI was the announced initial offtaker at Stargate Norway and left, nine months after the announcement, without concluding. If the Anthropic line churned the same way, the backlog’s named half goes with it. Either way, the pace of change is a risk signal. The S-1 quantifies the concentration: the largest customer was 52 percent of H1 2026 revenue, 73 percent of FY2025, and FY2024 was 93 percent a single customer, Pega Nordic. Microsoft and Anthropic together carry 86 percent of the backlog. The exhibits put clause-level texture on the exit risk: under the Anthropic order, no fees are payable for a tranche until Anthropic’s written acceptance, and late delivery carries a per-tranche walk-right “without any payment or other obligation or liability of any kind,” though no termination-for-convenience clause appears; Microsoft’s book terminates for cause on a sixty-day cure. The floor is real. It is a floor that starts on delivery.

The capital structure adds a layer of risk. Topco insertion, share premium reduction by solvency statement, six Series C share classes, and two debt charges create a structure where the IPO buyer may own a different instrument than the roadshow presentation implies. The UK Companies House accounts are scanned images, not machine-readable, which slows verification at exactly the moment verification is needed. A buyer who cannot see the cap table is buying the story, not the claims.

Hiring and engineering signals split cleanly, and the split is the finding. Two hundred fifty-four open roles, overwhelmingly infrastructure and construction across three continents, are the profile of a company building the physical plant and the debt that builds it; 21 GitHub stars across 80 public repositories are the profile of a company whose value does not live in its public code. Both can be true of a hyperscaler. What the combination does not support is reading the thin public engineering surface as a thin company, or the hiring ramp as evidence the backlog is converting. Operators are not revenue.

Two things this record does not show, and should not be read as implying. It shows no evidence of fabricated contracts. And future-dated capacity commitments are how this sector buys: CoreWeave built an audited backlog on the same kind of multi-year capacity deals, hyperscalers sign decade-long commitments, and pipeline framing with an illustrative caveat is standard roadshow practice. The finding here is narrower than bad faith: a $103 billion claim against $21 million of filed revenue, an FY2025 account three weeks from its due date, and a backlog that doubled in twelve days. That is a verification lag until the filings land. It becomes something else only if they do not. The filings landed.

The allocator decision comes down to timing and documentation. The IPO could price within weeks. The pre-IPO round could close before the FY2025 accounts are due. If you are considering an allocation, you are buying the backlog before it is audited. The burden is on the counterparty paper, and the paper is not filed. The claimed customer count and the claimed valuation are the two comparisons the public record can state precisely. The revenue fork is not even that: it pairs a roadshow figure with no filed counterpart at all. Update: the counterpart is filed. The pre-IPO round signed on 15 September at a minimum of $3.1 billion, the audited baseline arrived inside the S-1, and the question is no longer whether the paper exists but what the amendment prices it at.

Verification close

What can be checked now: Hiive has a price, GitHub has a repo count, DOL has LCA filings, and the company website has announcements. What cannot be checked now: the $103 billion contracted revenue, the $100 million quarterly bookings, and the $18 billion implied annualized run-rate. Those live in documents not filed with any regulator. Update, 19 September: they no longer do. The S-1 files all three: $103.4 billion of contracted take-or-pay, H1 2026 revenue of $140.6 million, and the 5.7-year weighted average life behind the run-rate arithmetic. The next tests are the amendment’s numbers: the price range, the shares outstanding, and the holders’ percentages. One more piece of arithmetic: the Series C put $2 billion in for 13.7 percent of the company at the $14.6 billion post. The falsifiers are on the record. The FY2025 accounts due 30 September 2026 grade the baseline: the filed revenue trail against the roadshow’s bookings story. They cannot grade the $18 billion annualized claim, because the period predates the signings that make up most of the backlog. That grading belongs to the S-1’s backlog note and the auditor’s view of contract assets. Both are now filed; the grading belongs to the reader.

If you want to check the work: Hiive | GitHub | DOL | company site

Every number traces to a source. The analysis is ours. - The SOMEN Desk

The AI compute supply chain is bifurcating into two parallel systems. One runs on private capital, market-based financing, and a Dutch lithography monopoly. The other runs on state capital, strategic allocation, and an indigenous substitution program built under export controls. Both converge at a single company whose machines neither stack can yet build for itself.

This is what the dependency map shows when you trace every edge from capital to silicon to model. A hundred and two companies. Five thousand relationships. Every one sourced.

Update, Aug 26: hours after this piece published, a stealth model that had spent a week atop OpenRouter’s usage leaderboard was confirmed as Zhipu’s GLM-5.3-Flash, and the company said every token of that traffic was served on Chinese chips. The China-models layer, the Huawei count, and the Jalapeño benchmarking below are updated to match the graph; the event is this piece’s software-separation finding made live.

The short version

• The West builds AI compute on private capital; China builds it on state allocation. Both stacks bottleneck through ASML, and beneath ASML through two private German firms. You can escape TSMC. You cannot escape ASML.

• The chain narrows as it deepens: many labs and clouds at the top, one laser supplier and one mirror supplier at the base, permitted power constraining everything above.

• China’s software gap is closing faster than its hardware gap, because iteration beats capital expenditure. Faster is still years, not quarters, and the hardware gap remains the multi-decade problem.

• NVIDIA is simultaneously the strongest and most fragile node in the Western stack: every large customer depends on its toolchain while building alternatives to its chips, and its financing platform bets on demand never softening.

• The freshest evidence cuts against indigenization: Xiaomi’s flagship XRING chips, framed as a self-sufficiency breakthrough, are all fabricated by TSMC. The newest cross-stack edge deepened instead of thinning.

The Western Stack

The Western AI compute supply chain runs through six layers, each dominated by one or two companies that the next layer cannot operate without.

Layer 1: The deepest bottleneck. Trumpf makes the high-power CO2 lasers that drive ASML’s EUV light source. Carl Zeiss SMT makes the mirrors that focus it. Both are private German companies. Neither has a competitor. The EUV scanner is widely considered the most complex machine ever built. It fires a high-power laser at 50,000 microscopic droplets of molten tin per second, creating a plasma that emits 13.5nm wavelength light, collected by the smoothest mirrors ever manufactured (atomically smooth, sub-0.1nm roughness), all in a vacuum. Each scanner costs $200 to $300 million and weighs 180 metric tons. ASML spent 20 years and billions of euros developing EUV. Its foundational EUV patents, filed during the 1990s consortium era, have already expired; the moat held anyway, because no single patent covers the machine, the tuning algorithms stay trade secret, and the active estate (more than 38,000 patents, refreshed at roughly 2,000 filings a year) runs on continuous filing rather than any one expiration date. Without Trumpf’s lasers and Zeiss’s optics, ASML cannot produce EUV scanners. Without EUV scanners, TSMC cannot produce advanced chips. The dependency chain starts in two family-owned companies in Ditzingen and Oberkochen.

Layer 2: The sole EUV supplier. ASML is the only company on Earth that manufactures EUV lithography scanners. It recognized revenue on 48 EUV systems in 2025, out of 327 lithography systems in total. Delivery of a single EUV scanner requires 3 Boeing 747 cargo flights. The next generation, High-NA EUV, is already shipping to Intel for the 14A (1.4nm) process that Terafab plans to license. There is no alternative to EUV. Canon’s nanoimprint lithography works for memory, not for logic chips. Electron-beam lithography is too slow for volume manufacturing, at hours per wafer versus EUV’s 150-plus wafers per hour. The semiconductor industry’s own roadmap (IRDS) assumes EUV remains the patterning technology through at least 2036. No replacement exists on any credible timeline.

Layer 3: The dominant foundry. TSMC fabricates the chips for every major Western AI company. The twelve became thirteen in June, when OpenAI unveiled Jalapeño, its first in-house inference chip, fabricated on TSMC N3 with CoWoS packaging. The thirteen in this graph that depend on TSMC: NVIDIA (100% of GPUs on N3 and N4P), Cerebras (wafer-scale), Groq, SambaNova, Etched, OpenAI, Xiaomi (all three XRING chips), Huawei (legacy Ascend dies stockpiled before sanctions), and Terafab (current Tesla AI5 fabrication, the very dependency its Texas fab is built to replace), plus capacity competitors Apple, AMD, Qualcomm, and Broadcom. Samsung Foundry and Intel Foundry are alternatives, but TSMC holds approximately 90% of advanced-node market share. Samsung has lost major customers (Qualcomm moved Snapdragon back to TSMC). Intel Foundry counts announced external customers, Microsoft among them on the 18A node, but has yet to record meaningful external revenue. Terafab would be the first at scale.

Layer 4: The sole high-end GPU supplier. NVIDIA designs the chips that power every frontier AI training run. All NVIDIA GPUs are fabricated at TSMC. Twenty-three companies in this graph carry direct NVIDIA dependencies: the GPU clouds (CoreWeave, Lambda, Together, Crusoe), frontier labs and model builders including SSI, Thinking Machines, Skild, Atoms, Starcloud, OpenAI, Perplexity, Mistral, and Cursor, Alibaba’s AI cloud, Scale AI, Blue Origin, Valar Atomics, and Riot, plus the five hyperscalers themselves, Microsoft, Amazon, Google, Meta, and Oracle, whose procurement the graph now records as direct customer edges. Four of those hyperscalers are also building custom silicon: Microsoft has Maia 200 (in Azure production, serving OpenAI models), Amazon has Trainium3 (shipping), Google has TPU v7 (deployed at scale for inference), Meta has MTIA (four generations in deployment). None has displaced NVIDIA for frontier training. The newest escape attempt is the purest: OpenAI’s Jalapeño inference chip, announced in June and benchmarked in August at 1.5 to 1.9x more AI work per watt than NVIDIA’s GB300 racks for serving, was designed with Broadcom and runs on TSMC N3 with CoWoS packaging. It escapes NVIDIA for inference and lands directly on the layers beneath. SemiAnalysis, invited into the labs, found more beneath: a revised stepping already in the fab at roughly 25 percent better performance per watt, and single-token throughput per megawatt beating Vera Rubin’s published figures. The dependency map shows why: custom silicon still requires TSMC fabrication, which still requires ASML lithography. The customers are trying to escape NVIDIA, but they cannot escape the supply chain NVIDIA sits inside.

Layer 5: The GPU cloud layer. CoreWeave, Lambda, Together, Crusoe, and Riot Platforms deploy NVIDIA GPUs for AI training and inference. All depend on NVIDIA hardware, which depends on TSMC fabrication, which depends on ASML lithography. Anthropic signed a 20-year, $9.1 billion lease for 191 megawatts at Riot’s Rockdale, Texas campus. The bitcoin miner is now an AI infrastructure landlord, converting a 1.7-gigawatt Texas power portfolio from cryptocurrency to compute. Crusoe is building 2.1 gigawatts of AI data-center capacity in Abilene for the OpenAI Stargate initiative and a separate 900-megawatt campus for Microsoft. Valar Atomics is developing nuclear SMRs that directly power NVIDIA Blackwell chips, with a 30-megawatt waterless AI factory in Utah.

Power is the binding constraint on all of this. Goldman Sachs projects data-center electricity demand growing at 15% annually, reaching 8% of total US electricity consumption by 2030. Every gigawatt of AI compute requires permits, grid connections, cooling infrastructure, and years of construction. The companies that control permitted power (Riot: 1.7 GW, Crusoe: a reported 7 GW pipeline, Talen Energy, Constellation) hold an asset that AI labs cannot manufacture or code their way around.

Layer 6: The AI labs. OpenAI depends on Microsoft Azure for compute. Anthropic depends on Amazon AWS and Google Cloud. xAI merged into SpaceX at $250 billion and runs its Colossus training cluster on Oracle Cloud. Perplexitycommitted $750 million over three years to Microsoft Azure for frontier model access.

Anthropic’s compute procurement in three months reportedly totaled over $60 billion in contracted commitments, and the instruments differ: a reported $45 billion compute arrangement with xAI, a lab buying inference from another lab; $10 billion in capacity commitments with Volta Infra; and the one filing-verified leg, the $9.1 billion 20-year lease at Riot. The composite is directionally useful; only the Riot leg is documented in a public filing. The labs sit at the top of a dependency chain that runs through clouds, through NVIDIA, through TSMC, through ASML, through Trumpf and Zeiss.

The China Stack

China is building a parallel supply chain at every layer. The ambition is documented: Xi Jinping’s April 2025 Politburo session called for “self-reliance and self-strengthening” with an “independent and controllable ecosystem across hardware and software.” Huawei founder Ren Zhengfei’s stated target: 70% self-sufficiency by 2028.

Layer 1: Lithography. SMEE, China’s oldest lithography maker, ships 90nm-class dry ArF scanners. Aishengna, a state-owned venture that absorbed SMEE’s and Huawei-linked Yuliangsheng’s engineering teams, has begun mass production of immersion DUV scanners, targeting roughly five units in 2026, with 20 planned for 2027; the first units go to SMIC, Hua Hong, and CXMT for production-line validation rather than volume output. The machines still use imported Japanese components. A Shenzhen lab has an EUV prototype that generates photons but has not exposed a wafer. More than 3,000 researchers reportedly work across the program. Beijing’s target for chip output from the domestic EUV machine is 2028; Reuters’ sources call 2030 more realistic.

ASML recognized revenue on 327 lithography systems in 2025, 48 of them EUV. China’s entire domestic lithography program produced five immersion DUV tools.

Layer 2: Foundry. SMIC produces 7nm chips using DUV multi-patterning, a technique that achieves smaller features through repeated exposures on older equipment. The cost is punishing: approximately 40 to 50 percent more expensive than TSMC’s equivalent process. External estimates (SMIC publishes no yield data, and no two analysts agree closely) put the mature 7nm line near 50 percent yield, falling to 20 to 40 percent on the 5nm-class node where multi-patterning compounds. Treat the percentages as informed inference, not measurement. Huawei is building three of its own fabs in Shenzhen’s Guanlan district, but the first was not expected at full operation before 2026, relying heavily on untested domestic equipment. SMIC co-founder Wang Yangyuan and the heads of YMTC, NAURA, and EDA firm Empyrean jointly described China’s tool industry as “small, fragmented, and weak” in March 2026 and called for national consolidation. Big Fund III’s $47.5 billion is reportedly being redirected toward lithography and EDA as a direct response.

Layer 3: Chips. Huawei is China’s leading AI chip designer. Huawei markets the Ascend 910C as H100-class; a single third-party test by DeepSeek measured it at roughly 60 percent of an H100 for inference, with sustained training reliability the weaker suit. Huawei’s compensation is scale: system-level clusters like the 8,192-chip Atlas 950 SuperPoDtrade silicon efficiency for interconnect and aggregate throughput. The Ascend 950PR targets 2.8x the H20. Huawei plans 600,000 Ascend chips in 2026 and expects $12 billion in AI chip revenue. Cambricon is second, planning 500,000 Siyuan accelerators. China’s domestic chips made up a reported 41% of the Chinese AI chip market in 2025, up from under 10% pre-2023. NVIDIA’s share of China’s advanced AI accelerator market has collapsed. Eight companies in this graph depend on Huawei commercially, including DeepSeek, whose V4 model was the first frontier-class AI validated on Ascend hardware, and Zhipu, whose GLM-5.3-Flash just spent a week as the most-used model on OpenRouter, served entirely on domestic silicon.

Layer 4: Software. Huawei’s CANN (Compute Architecture for Neural Networks) is China’s CUDA alternative. Open-sourced in December 2025, it claims 4 million community members by Huawei’s own count (a metric spanning registered users of the broader Ascend ecosystem, not kernel engineers) and supports 70-plus models on release. The developer experience has moved from what one startup founder called “essentially a section of desert” to what he now calls a “youth phase.” AIGCode, a Shanghai AI coding startup, tried training a 7B-parameter mixture-of-experts model on Ascend in early 2025, hit a bug, and waited four months for Huawei support. By 2026, the same company achieved 65% Model Flops Utilization on Ascend during pre-training, nearly double the industry average. A major Chinese bank contributed 34 optimizations to vLLM-Ascend, the inference engine project, trusting Huawei chips with core risk-management processes. CANN Next adds a SIMT programming model that adapts directly to CUDA code, lowering the migration barrier.

DeepSeek V4 achieved 68 percent of H100 throughput on Ascend at 55 percent power consumption in reported inference tests (pre-training remained on NVIDIA-class clusters). The learning loop has started: each model adaptation makes the next one easier. CUDA took 15 years to build that loop. China is compressing it under export-control pressure.

Layer 5: Cloud. Alibaba holds a reported 35.8% of China’s AI cloud market. Tencent reportedly spent CNY 18 billion on AI in 2025 and planned to double it in 2026. Huawei has reportedly deployed over 300 Atlas 900 A3 SuperPoD systems (384 Ascend chips each) for 20-plus enterprise and government customers. Alibaba ordered hundreds of thousands of Ascend chips for DeepSeek V4 deployment. Tencent and ByteDance followed. The scale of domestic Chinese AI compute is larger than most Western analysts assume. The constraint is not demand or capital but chip supply: Huawei cannot produce Ascends fast enough to meet orders.

Layer 6: Models. DeepSeek is China’s leading open-source AI lab. V4 runs a 1.6-trillion-parameter mixture-of-experts architecture, activating 37 billion parameters per inference. The lab declined NVIDIA’s offer of early kernel access and chose to co-design its inference deployment with Huawei instead; V4’s pre-training itself still ran on NVIDIA-class clusters, a split that is itself the two-stacks condition in miniature. DeepSeek delayed the V4 release by three months to re-engineer inference for CANN. Its $7.4 billion first external funding round, at a $55 billion valuation, was funded entirely by domestic strategic capital.

The distinction did not survive the week. A free stealth model listed as Ox Alpha spent six days as the most-used model on OpenRouter, more than doubling DeepSeek’s traffic, until Zhipu confirmed it as GLM-5.3-Flash and open-sourced the weights: 320 billion parameters, MIT license, a newly trained multimodal base, at a tenth of its own flagship’s price. It was the fifth anonymous “Alpha” release in six months; all five resolved to Chinese labs. The confirmation carried the load-bearing fact for this map, in Z.ai’s own words: “all of this traffic was served on Chinese AI chips,” on a production disaggregated serving architecture across tens of thousands of domestically developed accelerators, with a claimed 3x end-to-end inference advantage over NVIDIA that is the vendor’s own number and unverified. It does not need to be verified to make the point. DeepSeek proved a frontier model could run on Ascend in a validated test. Zhipu proved the open market’s most popular model could run on domestic silicon at production scale, for a week, at free pricing. Zhipu is also the stack’s first listed lab (Hong Kong, January 2026; a market cap that touched the equivalent of $128 billion in June) and has sat on the US Entity List since January 2025.

The Edge

On August 24, 2026, Xiaomi added a seventh layer to the China stack from an unexpected direction: the consumer device. At its XRING chip conference in Beijing, the company showed the AI Cube Prototype, a 150-watt mini-PC that runs 120-billion-parameter and 3-billion-parameter models locally on three chips Xiaomi designed itself. The XRING O3 is a 3nm flagship SoC already in mass production for September’s Xiaomi 18 Fold, with a shipment target of 200,000 to 300,000 units. The XRING O100 stacks two DRAM wafers face-to-face on a 6nm logic wafer, hybrid-bonded at a 1.4 micron pitch, reaching 1.22 TB/s of near-memory bandwidth. The XRING D100 is a 3nm smart-driving chip that Xiaomi calls China’s first at that node, validated, with commercial use scheduled for 2027.

The program behind it is serious: more than 20 billion yuan ($3 billion) invested to date, a semiconductor design unit grown past 3,000 people, and a stated plan of at least $7 billion over a decade. Xiaomi joins Apple, Samsung, Google, and Huawei in the small club of device makers that design their own flagship silicon, and it is extending the same chip architecture into cars and robots.

Here is the part the dependency map records and the launch coverage mostly does not. Every chip in the lineup is contracted to TSMC per Reuters’ sources (Xiaomi itself confirms only the O3’s node, and left the D100’s foundry undisclosed at the unveil). The O3 runs on N3P, the same 3nm node as Apple’s current parts. The designs are Chinese. The manufacturing is not. Xiaomi has not reduced its exposure to the Western stack; it has converted a dependence on Qualcomm merchant silicon, which any foundry could fabricate, into a dependence on TSMC leading-edge capacity, which sits under American export jurisdiction and competes for allocation with Apple and NVIDIA.

The export-control fine print makes the arrangement more interesting still. US export rules presume denial when TSMC fabricates 7nm-and-below chips for Chinese designers, and shipments in that class are presumed to be AI chips unless proven otherwise. The XRING O1 cleared the presumption in 2025 through mechanical carve-outs: the die was packaged outside China and stayed under the 35-billion-transistor threshold. The D100 is an explicitly AI chip at 3nm for a Chinese automotive program. How it clears the presumption is unreported. The AI Cube is being read as proof that China can design chips as good as almost anyone’s. It is equally evidence that the fabrication gap this map shows has not moved: China’s newest design win converges on the same Taiwanese fabs as every Western AI company.

There is a Western counterpart to the Cube, and the comparison is exact. NVIDIA’s DGX Spark and AMD’s Ryzen AI Max systems run 200-billion-parameter models locally on 128GB of unified memory. Apple answered a day later. The Mac Studio M5 Ultra, announced August 25, runs models with hundreds of billions of parameters in up to 512GB of unified memory at 1.2 TB/s, a tier above the Cube’s 120-billion-parameter class, built on the same TSMC 3nm node family. The head-to-head the headlines invite is a trap: the Cube’s much-quoted 1.22 TB/s is the accelerator’s small near-memory pool, not the system memory its large model runs in, and that pool’s bandwidth was never published. The honest comparison cannot be made yet. Xiaomi’s route is its own three-chip stack running its own MiMo models. Same product category, similar parameter counts. The edge layer is the one place where both stacks already ship competing products at consumer volume, and China’s entry is manufactured by the Western stack’s foundry.

For the track record, three falsifiers: whether the O100 and D100 commercialize in 2027 as committed. Whether Qualcomm’s China revenue mix shifts as XRING displaces Snapdragon at the top end. And whether TSMC’s 3nm allocation to a Chinese national champion survives the next tightening of the AI presumption. The graph now carries the edges.

The Convergence Point

Both stacks bottleneck through ASML.

SMIC, China’s most advanced foundry, depends on its installed base of ASML DUV immersion scanners for 7nm production. The American Enterprise Institute’s April 2026 Lithography Loophole report estimated SMIC’s 5nm process at approximately 20 percent yield using multi-patterning on ASML’s older DUV tools. If the Netherlands restricts servicing of those machines, SMIC’s advanced-node production halts. Chinese fabs have been preparing for exactly that scenario by using third-party engineers and gray-market parts to keep older ASML machines running. The proposed US MATCH Act would ban not only sales of immersion DUV tools to Chinese entities but also servicing of machines already installed.

Terafab, Elon Musk’s semiconductor fab joint venture, aims to break the TSMC dependency by building its own foundry in Texas. But Terafab plans to license Intel’s 14A process. Intel’s 14A uses ASML High-NA EUV scanners. The Terafab project breaks the foundry bottleneck without breaking the lithography bottleneck. You can escape TSMC. You cannot escape ASML.

This is the structural finding the dependency map makes visible: the AI industry’s physical supply chain has a single point of failure at the lithography layer that no amount of foundry diversification can solve. Three foundries (TSMC, Samsung, Intel) compete for customers. One company (ASML) supplies all three. Two private German companies (Trumpf, Zeiss) supply ASML. The chain narrows as it deepens. And lithography is not the only chokepoint shaped this way. The design layer narrows identically: three Western vendors, Synopsys, Cadence, and Siemens’ EDA arm, supply the software every chip in this map was designed in, the Chinese ones included, a lever Washington has already demonstrated a willingness to restrict. China’s counter is embryonic, and its own industry says so: Empyrean’s chief joined the “small, fragmented, and weak” verdict, and Big Fund III now targets EDA by name. The chain narrows in toolchain as well as in hardware. And it narrows in the network: the switching silicon that binds hundred-thousand-GPU clusters into one machine comes from Broadcom, the same company that co-designed Jalapeño. Tomahawk is named in OpenAI’s own announcement. The chip partner and the network chokepoint are the same firm.

The Financial Architecture

The capital layers tell a different story: the stacks are diverging in how they fund the AI buildout, not just how they build chips.

The Western escalation. The $500 billion did not appear from nowhere. It was the culmination of a year of escalating private-capital commitments to AI infrastructure. In November 2025, Brookfield launched a $100 billion AI Infrastructure Program with NVIDIA as founding partner, anchored by a $5 billion Bloom Energy deal for behind-the-meter power. In June 2026, Apollo led a $35 billion capital solution for Broadcom’s AI XPV platform, financing 20-plus gigawatts of compute capacity through 2028. In July 2026, BlackRock’s Global Infrastructure Partners completed a $40 billion acquisition of Aligned Data Centers and closed a $12.5 billion bond sale for Meta’s El Paso data center campus.

Then in August 2026, NVIDIA signed non-binding memorandums of understanding with six financial giants: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The goal: mobilize over $500 billion in third-party capital. NVIDIA has also agreed to back OpenAI’s Ohio data center campus with a guarantee of up to $105 billion, scaled down from the $250 billion reportedly discussed in July after investors balked. The instrument backs the residual value of finished buildings rather than OpenAI’s rent, covers only the first 4.25 gigawatts of the 10-gigawatt site, and lapses entirely if OpenAI earns an investment-grade credit rating.

BlackRock CEO Larry Fink called it “the next future for financial engineering, akin to the creation of mortgage-backed securities in the 1970s”. Goldman Sachs CEO David Solomon said the goal is to “create a market for credit backed by NVIDIA compute”. Blackstone President Jon Gray compared it to “the way mortgage lenders look at homes”.

The mechanism: capital flows from these firms to NVIDIA’s customers, who use it to buy NVIDIA hardware, which shows up as NVIDIA revenue. NVIDIA may backstop up to 25 percent of projects. The risk sits on private equity balance sheets, not NVIDIA’s. NVIDIA is simultaneously the sole high-end GPU supplier, the compute-revenue guarantor, and the financing-platform coordinator. NVIDIA’s stock fell as much as 3.2 percent on the $500B announcement day. Investors flagged the circular logic: capital raised to buy NVIDIA chips becomes NVIDIA revenue, with the risk on a different balance sheet. Jensen Huang called the criticism “ridiculous.” The dependency map shows why it is not ridiculous: NVIDIA’s position depends on compute demand remaining structurally constrained. If demand softens, the financing loop that services the debt weakens. Fink’s mortgage-backed-securities analogy cuts both ways.

The China model. The Big Fund (National Integrated Circuit Industry Investment Fund) is China’s equivalent, but the mechanism is state-directed, not market-based. Three phases have deployed approximately $96 billion since 2014: Phase I (~$20B, fab-building), Phase II (~$29B, supply-chain hardening), Phase III ($47.5B, bottleneck-breaking in lithography and EDA). Phase III marks a clear pivot: rather than funding mature wafer fabs, the fund targets the exact weaknesses export controls expose. Nineteen state-owned investors led by the Ministry of Finance. China Construction Bank committed CNY 21.5 billion.

DeepSeek’s $7.4 billion first external round illustrates the model. Founder Liang Wenfeng committed $2.8 billion personally (84 percent ownership). The outside investors are all Chinese domestic strategics: Tencent ($1.4B), CATL ($700M), JD.com ($420M), NetEase ($420M), plus the Big Fund as lead. No foreign VC. No global capital markets. Liang’s framing: “Those whose interests align most with ours, who bear us the least hostility.”

The structural difference: the Western model treats compute as an investable asset class, priced by markets and backed by NVIDIA’s revenue guarantees. The China model treats semiconductor self-sufficiency as a national security imperative, funded by state allocation and domestic strategic capital. Both are massive. The risk models are opposite. The Western model spreads risk across private balance sheets and depends on market pricing. The China model concentrates risk in the state and depends on political will. Each has a failure mode the other does not.

The NVIDIA Paradox

NVIDIA occupies a position in the Western stack that has no parallel in the China stack or in any previous technology cycle. It is simultaneously the sole high-end GPU supplier, the compute-revenue guarantor (its $105 billion OpenAI guarantee, its $500 billion financing platform, its reported $45 billion Anthropic-xAI compute arrangement), and the financing coordinator (connecting private capital to customers, underwriting the revenue that services the debt, collecting the chip purchases that flow back as revenue).

NVIDIA’s top five customers (Microsoft, Amazon, Google, Meta, Oracle) are all building custom silicon; Layer 4 catalogued their programs. None has displaced NVIDIA for frontier training, and if two ever did at frontier scale, NVIDIA would lose an estimated 30 to 40 percent of its GPU revenue base. But the replacement chips queue at the same foundry, on the same nodes, behind the same allocation priorities: TSMC’s N3 favors its largest customer (Apple) and its highest-volume AI customer (NVIDIA) ahead of any upstart accelerator. Escape from NVIDIA runs through the line where NVIDIA jumps the queue.

In China, Huawei plays a similar coordinating role, but without the financial-engineering layer. Huawei coordinates the semiconductor ecosystem (approximately 2,000 companies), invests through its Hubble arm, and co-invests with the Big Fund. But Huawei does not run a market-based financing platform. NVIDIA’s financial layer is a structural advantage that Huawei cannot replicate under China’s capital controls. It is also a structural vulnerability: if the compute-backed debt underperforms, the financing loop that feeds NVIDIA’s revenue contracts. No comparable feedback risk exists in the China model, where the state absorbs losses directly. And the mechanism is no longer NVIDIA’s alone: Broadcom’s AI XPV platform, launched in June with Apollo and Blackstone, applies the same securitization at the chip layer: a $35 billion first tranche, frontier labs leasing racks built on Broadcom silicon, a $29 billion backstop cap disclosed in Broadcom’s own 10-Q, and a Bank of America-modeled ceiling of $370 billion at 20 gigawatts. The capital platform is becoming infrastructure itself.

The Cross-Stack Dependencies

Six edges in the dependency map connect the two stacks. Each represents a dependency China is trying to break.

SMIC depends on ASML for its installed base of DUV immersion scanners. The Aishengna DUV program (five scanners in 2026, 20 planned for 2027) is the substitution path, but the machines still use imported Japanese components and have not reached volume production. Delays at those Japanese suppliers are what capped 2026 output at five units. The dependency chain extends through three allied nations: the Netherlands (ASML machines), Germany (Trumpf and Zeiss beneath ASML), and Japan (critical components inside Aishengna’s domestic substitute).

CXMT, China’s leading memory chip maker, depends on ASML DUV for HBM production. The same servicing risk applies. Without HBM, Huawei’s Ascend chips are memory-bandwidth constrained regardless of their logic performance.

Huawei depends on TSMC for legacy Ascend dies stockpiled before sanctions. An October 2025 teardown of an Ascend 910C found dies originally fabricated by TSMC in 2020. The stockpile is depleting. The transition to SMIC-fabricated Ascends is ongoing but yields remain unconfirmed. No public teardown has yet confirmed a fully domestically fabricated Ascend 910C.

Alibaba still deploys NVIDIA GPUs where available. The H20 was the most advanced chip NVIDIA could sell in China before the April 2025 ban. Alibaba is gradually shifting to Ascend for domestic workloads, but the transition is incomplete.

Xiaomi depends on TSMC for its entire new XRING lineup: the O3 flagship SoC in mass production now, and the O100 and D100 scheduled for 2027, all contracted to TSMC per Reuters. This is the newest cross-stack edge and the most concentrated: a Chinese national champion’s celebrated design win is manufactured entirely inside the Western stack, competing for the same N3P capacity as Apple and NVIDIA. The Edge section above unpacks it.

The direction of travel is clear: each cross-stack edge is thinning. SMIC is building domestic DUV capacity. Huawei is transitioning from TSMC stockpiles to SMIC fabrication. DeepSeek has adapted V4 to run on Ascend. Alibaba is ordering hundreds of thousands of Ascend chips. And the motion is now symmetrical at the model layer: DeepSeek is building its own inference chip, explicitly following OpenAI’s Jalapeño playbook to cut reliance on both NVIDIA and Huawei. The edges are not severed, but they are narrowing. The question is whether they close before China’s domestic alternatives reach production maturity or after. The newest edge runs against the trend: Xiaomi just converted a Qualcomm merchant-silicon dependency into a TSMC foundry dependency, deepening the cross-stack link rather than thinning it.

The Structural Read

The dependency map reveals four structural findings that are invisible without cross-company graph analysis.

First, the physical supply chain narrows as it deepens. Twenty-three companies in this graph depend on NVIDIA. NVIDIA depends on TSMC. TSMC depends on ASML. ASML depends on Trumpf and Zeiss. The chain widens at the top (many AI labs, many clouds) and narrows at the base (one laser supplier, one mirror supplier). Risk concentrates at the narrowest point, not the broadest. An allocator evaluating an AI cloud position should trace the dependency to its root: if Trumpf stops supplying ASML, every company in the chain is affected. Power adds a second narrowing: every gigawatt of AI compute requires grid connections, permits, and years of construction. The companies that control permitted power hold an asset that sits outside the silicon supply chain but constrains it just as tightly. One distinction the map draws quietly: a permitted gigawatt is a claim on generation, not generation itself, and the fuel, turbine, and grid-upgrade layer that converts permits into megawatts sits below the resolution of this map. The map stops at the meter. Memory adds a third narrowing, and it binds soonest: HBM is sold out through 2027, advanced packaging runs on year-plus lead times, and consumer and AI silicon now bid for the same pools, as Apple’s deferred Mac configurations showed in August. Lithography is the narrowest point in the chain. Memory is the narrowest point in the calendar, and it rests on three suppliers, SK Hynix, Samsung, and Micron, sold out through 2027.

Second, the financial layer is bifurcating along the same lines as the physical layer. The Western stack runs on private capital coordinated by NVIDIA ($500 billion platform, escalating from Brookfield’s $100 billion program and Apollo’s $35 billion Broadcom deal). The China stack runs on state capital coordinated by the Big Fund ($96 billion deployed across three phases). The financial architecture mirrors the physical architecture: open-market competition versus state-directed allocation. An allocator cannot evaluate one without the other. The cost of capital for AI infrastructure is structurally different in the two systems, and that difference flows through to pricing, deployment timelines, and competitive positioning.

Third, the software layer is the one place where separation is accelerating. CUDA (Western) and CANN (China) are diverging. DeepSeek’s V4 adaptation to Ascend is the reference workload that bootstraps the Chinese software ecosystem. AIGCode went from a four-month support wait to 65% MFU in under a year. A Chinese bank trusted Huawei chips with core risk management. CANN Next’s CUDA compatibility layer lowers migration costs. Each model that adapts to CANN makes the next adaptation easier. The learning loop is real, not theoretical. The hardware gap (EUV, yields, process node) may take a decade to close. The software gap may close faster because it requires iteration, not capital expenditure. Faster is not fast: CUDA’s fifteen-year head start is not one product but a stack (compilers, collectives, kernel libraries) and the community around it, and CANN’s own maturity is early. The honest read is that the loop has started, with the outcome still years out. When the software gap closes, the hardware gap matters less: if Ascend can run frontier models reliably, Huawei’s inferior fabrication process becomes a cost problem, not a capability blocker.

Fourth, NVIDIA’s position is simultaneously the strongest and the most fragile in the Western stack. It is strong because every customer depends on its hardware and its financing platform. It is fragile because every customer is building an alternative, and the financing platform creates a feedback loop that depends on perpetual compute scarcity. The graph shows NVIDIA’s largest customers building custom silicon. It shows the $500 billion financing platform turning compute into a commodity asset class, which is the first step toward commoditization. And it shows China’s CANN ecosystem bootstrapping on DeepSeek’s adaptation work. NVIDIA’s moat is not the chip but the toolchain around it: CUDA, the communication and kernel libraries beneath it, and fifteen years of community answers. Matching silicon performance displaces none of that; the customers building custom silicon are still rebuilding the toolchain from scratch. The first measured counterpoint arrived with Jalapeño: SemiAnalysis’s verdict from the lab visit was that “the CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon.” One port, on one chip, from one unusually strong silicon team. The moat defense now rests on the porting staying hard for everyone else. The pressure is real on both sides: custom silicon from customers, CANN from competitors.

What Changes the Picture

Four developments would shift the structural analysis. Each is a tracked scenario with a falsifiable trigger, not a prediction.

ASML servicing restrictions on China. If the Netherlands bans ASML from servicing DUV machines already installed in Chinese fabs, SMIC’s advanced-node production halts within months. The Aishengna domestic DUV program (five scanners per year) cannot replace the installed base quickly enough. China’s AI chip production drops. DeepSeek and Qwen lose inference capacity. The cross-stack edges to ASML become the binding constraint on China’s entire AI program.

CANN reaching production-grade reliability. If CANN stabilizes to the point where Chinese developers can deploy models without Huawei engineering support, the CUDA lock-in weakens. NVIDIA’s advantage narrows from ecosystem dominance to raw hardware performance. The software gap closes. Chinese AI labs no longer need NVIDIA GPUs for inference. The demand destruction hits NVIDIA’s China revenue (already near zero) and its global pricing power.

A successful Terafab tape-out on Intel 14A. If Terafab produces working chips on Intel’s 14A process, the foundry layer gains a fourth credible player outside the TSMC/Samsung/Intel triopoly. The execution risk is the point: 14A’s design kit reaches 0.9 only in October 2026, and Intel’s own risk production does not start until late 2027, so any Terafab chip precedes a process Intel has not yet ramped. This does not break the ASML dependency (14A uses High-NA EUV), but it does break the TSMC capacity-allocation bottleneck that constrains every Western AI silicon company. A startup shipping its first chip is the lowest-priority customer at TSMC. A startup with its own fab controls its own timeline.

A public-market repricing of AI infrastructure. CoreWeave and Cerebras are now public companies. SpaceX IPO’d at $1.75 trillion. If any disappoints on revenue or margin, the cost of capital for the entire GPU cloud and AI silicon layer increases. The Western financing model depends on compute-backed debt performing. Fink’s mortgage-backed-securities analogy is apt in ways he may not intend: MBS performed well until they did not. The first repricing has already printed: Meta’s El Paso paper cleared at 7.53 percent in July, well wide of the comparable October deal, on debt that matures in 2048 and collateral that refreshes every three to five years. If compute demand softens, through algorithmic efficiency gains or training costs falling faster than expected, the revenue assumptions underpinning $500 billion in financing weaken. And the unwind, if it comes, has an order. The GPU clouds reprice first: they are the most levered and least diversified layer, their debt services off inference revenue, and CoreWeave is already public and marked daily. The financing platforms take the second hit, when lease payments miss and the backstops activate, NVIDIA’s 25 percent and Broadcom’s $29 billion cap converting off-balance-sheet risk into balance-sheet risk. The hyperscalers slow capex third, from strength rather than distress. NVIDIA’s revenue prints last, because the pull-forward purchases the platforms financed must unwind first. The same map that shows the circularity shows whom it hits, and in what order. The China model has no comparable feedback risk because the state absorbs losses directly, but it also has no market mechanism to signal when investments are underperforming.

The Map Is the Finding

This piece does not argue that one stack will prevail over the other. It argues that the dependency map itself is the analysis. The relationships between companies, the layers of dependency, the convergence at ASML, the divergence in financing, the acceleration in software separation: these are findings that emerge only from mapping the full graph. The map weights the compute path; the memory and advanced-packaging layers (CXMT for HBM, CoWoS-class assembly capacity) and the EDA and materials strata beneath it are thinned to their load-bearing edges (Synopsys and Cadence for EDA), and the networking and power-generation strata are thinner still. The first hard numbers for that layer are now in the graph: OpenAI’s HBM4 allocation alone is roughly 7 percent of Samsung’s 2026 HBM output, its chip program holds an estimated 10,000 CoWoS wafers inside Broadcom’s roughly 150,000-wafer block, HBM is sold out through 2027, and Apple’s August Mac launches met the same wall from the other side, with memory-heavy configurations deferred to October as DRAM reallocates to AI. Consumer silicon and AI accelerators now bid for the same three pools: leading-edge wafers, advanced packaging, and HBM. A fuller treatment of those layers would sharpen the narrowing, not change it. Nor do the would-be third poles escape the frame: Intel Foundry and Japan’s Rapidus both depend on the same ASML exposure, and Tokyo Electron’s equipment already feeds TSMC, which is why the map records them inside the Western stack rather than beside it.

The track record question is whether the graph-derived findings prove durable. Does ASML remain the convergence point? Does CANN close the software gap? Does the $500 billion materialize or remain an MOU? Does Terafab produce chips? Does SMIC reach viable yields on 5nm? Does the compute-backed debt perform?

The answers to those questions will determine whether the two stacks converge or diverge over the next decade. The graph will track each one.

Company sites and press releases for funding and partnership announcements. CEIAS analysis for the China AI chip supply chain overview. Tom’s Hardware for China’s domestic lithography roadmaps. Brookings for US-China AI strategy comparison. AEI Lithography Loophole report for SMIC multi-patterning analysis. NVIDIA Newsroom for the $500 billion financing platform. SCMP for DeepSeek funding details. Riot Platforms Q2 2026 for the Anthropic compute lease. CNBC for the SpaceX-xAI merger. ChinAI newsletter for CANN developer ecosystem analysis. Reuters for Xiaomi XRING fabrication at TSMC. Cnevpost and VideoCardz for AI Cube specifications. ASML Q4/FY2025 results for shipment and EUV unit figures. Tom’s Hardware for the Ascend 910C inference benchmark. Kaohsiung Times for SMIC yield estimates. Asia Society CCA for WAIC 2026 Huawei accelerator assessments. Richard Stahl for the EUV patent-expiration analysis. TheNextWeb for the NVIDIA share reaction to the $500B announcement. TrendForce for the Intel 14A timeline. Bloomberg and the Z.ai blog for the Ox Alpha confirmation and the Chinese-chips serving claim. Caixin for Zhipu’s market-cap record. SemiAnalysis for the Jalapeño lab visit. Where the China stack is concerned, these secondary analyses stand in for primary disclosure that does not exist; figures resting on a single source are marked as reported in the text.

Every relationship traces to a sourced edge in the SOMEN knowledge graph. 102 companies. 5,000 edges. The analysis is ours. - The Somen Desk

Etched says it closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia Capital. Do the division the announcement implies: $300 million of new money into a $10.3 billion post-money buys the syndicate roughly 2.9 percent of the company, and that is the ceiling. Sequoia’s own announcement prices the round at $10 billion pre-money, which confirms the $10.3 billion as a post-money figure; if any part of the round was secondary, the new-money stake is smaller still. You can verify that on a napkin. For context, the desk’s rule of thumb is that a Series C lead alone takes 15 to 25 percent; here the entire syndicate, Sequoia through SK Hynix combined, bought a fraction of that. Sliver stakes are the 2026 norm at the top of the market, to be fair: SambaNova’s $1 billion bought its syndicate roughly nine percent, Groq’s $350 million about ten. The company’s own boast is that this is the highest valuation ever for a Sequoia-led Series C; the record-setting price and the record-thin stake are the same fact. The napkin’s point is narrower than rarity: Etched’s syndicate took the thinnest slice in the sector’s 2026 set, at the only company in that set with no shipped product. The valuation is a claim. The percentage is arithmetic, once its two assumptions are stated. When the two diverge this sharply, believe the arithmetic first.

Several structures can explain a sliver round. The syndicate may simply have crowded: per the company’s own announcement, Sequoia appears alongside Andreessen Horowitz, Jane Street, Diffusion, Argo, and SK Hynix on the same ticket, and crowded rounds divide ownership until nobody carries underwriting weight. (Argo is the one name in that roster no other coverage carries, not even Sequoia’s own announcement of the round.) Part of the $300 million may be secondary, recycling existing shares rather than funding the buildout. Strategics may hold structured terms the press release does not describe. Either Sequoia accepted a rounding-error stake to print a headline number, or the round carries structure the announcement does not disclose. The data does not decompose it, and no share count is published anywhere we can find.

The secondary market will not arbitrage this for you. The tape has printed $150.41, $183.96, $237.90, and now $242.16, a 61 percent move across eighteen days of captures with no volume disclosed behind any of them. The reconciliation fork is the content: which share count is real decides what the prints meant. The August 10 print reconciles with the $10.3 billion mark at about 56 million shares. The August 18 print reconciles at about 43.3 million. A share count does not shrink by a quarter in eight days, so at most one print was fairly priced, and neither pins the cap table. Caveat in the same breath: whether the instrument changing hands on Hiive and the preferred stack the round priced are the same share class, there is no disclosure. A tape that swings this hard is not pricing the company. It is pricing the absence of a market.

The company’s own progress page describes a business at full boil: production underway against more than $1 billion in customer contracts, first racks promised “this summer”, A0 silicon back from TSMC, a Taiwan factory, and a new 10-megawatt lab in San Jose. Every item self-reported. But the verifiable data is not empty either: a live job board, a wage filing, granted patents, and a secondary tape. This piece walks the gap between the boil and the paper, and the napkin is the instrument throughout.

Our method: we treat company pages as claims and the trade tape, the funding ledger, the public code org, the job board and its 108 open roles, certified wage filings, and the patent record as evidence. Where the evidence is thin, we say so. Where it is absent, we say that louder.

Valuation: the tape against the headline

Start with what is stable and what is moving. The claimed valuation reads $10.3 billion on every capture of the progress page in our data, and claimed lifetime funding reads $800 million alongside it, unmoved across captures. Static numbers on a company page prove only that the page is stable. The tape already narrated above is the moving number, and two of its three prints cannot both have been right. When a thin tape moves 58 percent in seventeen days, the move is the finding: the prints are marking mood, not value.

A per-share price and a total valuation are different units, and reconciling them assumes facts the data lacks. If the Hiive instrument is common and the $10.3 billion claim prices preferred with liquidation preferences, the per-share price reflects that discount. If they are the same class, the tape confirms the valuation rather than contradicting it. The reconciliation is suggestive, not conclusive.

A secondary print has its own caveats, and this desk has likely given it too much benefit before now. Hiive reflects a small number of transactions between accredited investors, and a price without volume may not even be a cleared trade: it can be an indicative quote, asking a price nobody paid. The company’s $10.3 billion claim has a press release behind it. The tape has a market operator behind it. Neither has demonstrated size behind it. The tape’s edge is observability, not truth, and a 58 percent swing in seventeen days is a reminder of how little that observability currently buys.

Who is on each side of those prints, the record does not say, and the two sides tell different stories. A seller at $150 reads as an insider cashing out: an employee diversifying locked compensation, an angel rotating capital, a holder who needs liquidity more than conviction. A buyer at $237.90 reads as a party building a position on purpose. A deep market would contain both and meet in the middle; a tape of one motivated seller prints low, and one patient buyer prints high, which is what a 58 percent swing on no volume is. Which composition it is, Hiive does not disclose, and the composition is the story.

Then the number that is missing. The data contains no ARR for Etched. Not a range, not a bookings conversion, not a named account paying a named amount. The closest items are self-reported: more than $1 billion in customer contracts, a figure the same page elsewhere describes as “demand”. Contracts are not revenue. In hardware they are intentions with legal cover, and delivery slippage renegotiates them faster than any other instrument in the stack.

The next headline is already circulating. The Wall Street Journal has Etched in talks near $20 billion, per reporting relayed on the company’s own site, reportedly led by an existing investor. At that level the sliver logic compounds: each new round at a higher mark buys a smaller fraction, and a cap table marked in slivers is a cap table where price discovery has quietly stopped happening. A round is a fact when a filing says so. Everything before that is a report of talks. The report undershot: the round closed at $21 billion, the sliver logic compounded, $700 million buying 3.3 percent.

Funding: the ledger against the press release

The round ledger, as recorded: a $5.36 million seed led by Primary Venture Partners, 2023-03. A $120 million Series Aco-led by Primary Venture Partners and Positive Sum Ventures, 2024-06, disclosed alongside the TSMC fabrication partnership. A $500 million Series B led by Stripes at roughly a $5 billion post-money, 2025-12. A $300 million Series Cled by Sequoia Capital, 2026-07, at $10.3 billion. Four entries, clean progression, marquee names at every step. The least verifiable entry is the $500 million Series B, platform-sourced with no filing behind it anywhere in the data; the desk weights that one at fifty cents on the dollar. On August 18 the ledger gained a fifth entry: $700 million led by Jane Street at $21 billion. The named rounds now sum to roughly $1.63 billion against a lifetime claim of $1.9 billion: last week the ledger outran the claim; this week the claim outruns the ledger; the definition of “raised” remains the company’s.

The trajectory has two unusual features, and both are checkable. The seed-to-A step-up is 22x in round size, extraordinary and consistent with dramatic early technical progress. Then the rounds shrink: the Series C at $300 million is 40 percent smaller than the preceding $500 million round. In a normal trajectory, rounds grow. A shrink means one of three things: the company did not need the cash because revenue appeared (unlikely given no customer data); the valuation made equity too expensive to sell in volume (possible at $10.3 billion); or part of the round was secondary, meaning less new money entered than the headline suggests. The data does not say which.

Take the secondary possibility seriously on its own, because the surrounding facts earn it. The round shrank from $500 million to $300 million while the headline doubled, no filing exists to decompose the check, and a company letting insiders sell inside a growth round, at the richest price in its history, before first racks ship, is the classic late-stage configuration: founders and early employees taking liquidity while their own information about ship risk is better than yours. None of that is damning, and none of it is disclosed. If any part of the $300 million was secondary, the primary was smaller and the true new-money stake sits below 2.9 percent. The one number that would settle it, the primary-secondary split, is the one number nobody publishes.

Now do the reconciliation the company has not done. The progress page says Etched has raised $800 million across its financings. Add the ledger yourself: the four named rounds sum past $900 million, roughly $925 million, before counting the unannounced financings the same page claims, a strategic investment from VentureTech Alliance among them. Either the $800 million predates the Series C and the page is stale, or “raised” means something narrower than the ledger’s sum. A company marking itself at $10.3 billion should reconcile its own fundraising total in a sentence. It has not, and the coverage we reviewed did not ask.

Run the burn the claims imply, because runway arithmetic matters more than any multiple at this stage. Silicon talent in the Bay Area runs $350,000 to $450,000 a head fully loaded; the claimed 400 engineers imply $140 to $180 million a year in personnel alone, before tape-outs at $50 to $100 million per advanced node, the 10-megawatt lab, and the Taiwan factory. Call the total $250 to $350 million a year against $800 to $925 million raised. That is two to three and a half years of runway, one or two full tape-out cycles. The 108 open roles read differently against that arithmetic: not expansion on a comfortable bench, but the burn continuing, one more reason the fundraise schedule was never optional.

Two of the Series C participants deserve specific attention because they are not financial tourists. SK Hynix is a top-three global memory manufacturer. They make HBM, the high-bandwidth memory every advanced AI chip requires, and memory is the bottleneck asset in inference hardware. A memory maker funding an inference silicon company is purchasing demand visibility alongside returns: it validates the roadmap, not the price. Strategic money aligns incentives. It also prices strategically, on terms undisclosed. Jane Street is one of the largest quantitative trading firms in the world; they model risk for a living. What the data does not show is what their model bought, because quant desks price structure as readily as direction, and structure is exactly what this announcement does not disclose. A day later, structure with a name: Jane Street led the $700 million round and became the first customer, first rack in its datacenter, quote: “We tested the chip and are pleased with the early results.” One desk validated demand and price on a single ticket: stronger evidence the hardware works, weaker evidence the price is independent. Note also what the syndicate is not: no growth fund taking a board-shaping stake at size is identifiable in the record. The round was assembled, not underwritten.

And that matters for what the price certifies, because a strategic’s economics are not a common holder’s economics. Hynix’s payout is not only the mark-up on the shares: the seat buys HBM attach revenue, roadmap visibility, and a design-win relationship worth more to a memory maker than any multiple on the check. A strategic can rationally pay above fundamental value because it collects a side payment the common shareholder never sees. The same logic runs for any buyer pricing M&A optionality, the way NVIDIA priced $20 billion of optionality out of Groq. The syndicate clearing at $10.3 billion is not a certificate of the common’s worth. You do not receive the allocation economics, the supply leverage, or the option. Their all-in price and yours are different numbers wearing the same headline.

The angel roster is dense, and its core is externally confirmed: Hinton, Karpathy, Druckenmiller, Mensch, Thiel, Field, and Masad appear in outside coverage of the round, with the remainder, Dao and Abbeel among them, on the company’s own roster alone. Angel density says the idea circulates at the top of the field; it says nothing about price discipline. Small checks buy credibility; credibility is not capital at risk.

The comp set: priced like a story

Etched’s mark means something only against the set it prices with, so here is the desk’s peer table for inference silicon, marks refreshed to 2026 transactions where public sources exist:

Every peer around Etched repriced in 2026, and every repricing carried evidence Etched does not have. Cerebras priced at $185 a share, closed its first Nasdaq session at $311.07 for a market cap near $95 billion, with audited numbers attached: $510 million of revenue, up 76 percent, $88 million of net income swung from a $481.6 million loss, and a prospectus that discloses 24 percent of revenue from G42, 62 percent from a single university. SambaNova is the closest analogue to Etched’s own arc: in December, Intel reportedly valued it around $1.6 billion in acquisition talks; seven months later it closed $1 billion at $11 billion, a near-sevenfold step on the same company, five months after a $350 million Series E launched alongside its SN50 chip. That repricing cuts both ways, and honesty requires both edges: it is the sector’s proof that the market will pay seven times over on narrative while the story is hot, which is precisely the bet Etched’s syndicate made at $10.3 billion, and it is a reminder that marks set on narrative reprice on the same narrative in the other direction, as the next company in this table spent the same year demonstrating. Groq’s 2026 was the whole lifecycle compressed: NVIDIA paid about $20 billion for its technology and its founder, against a $6.9 billion peak mark; the rump company re-emerged this month at $3.5 billion with NVIDIA itself joining the round, which the company calls a reset rather than a down round and the arithmetic calls half the peak.

Read the money multiples with that in view. Etched at roughly 12.9x named-round funding carries the richest mark-to-money ratio in the table, and the August step-up raised the mark and the ratio together: above 10.3x for a company trading with an audited income statement, above 8.1x and 3.5x for the privates, all of which shipped product before repricing. The market paid Cerebras-grade multiples only after revenue, profit, and a prospectus existed. Etched’s claim prices the outcome before the evidence, and sits alone between the evidence tier at $3 to $4 billion and the audited tier at $95 billion. That is the pattern, and the pattern is what a reader should price.

Tenstorrent is the row that settles the argument. Jim Keller’s company, last marked $3.2 billion in November after raising $800 million on top of a $693 million Series D, is reportedly sitting on roughly $150 million in customer contracts with Intel and Qualcomm circling as buyers. A decade-old silicon company with the industry’s most credentialed architect, real contracts, and strategic interest clears a third of Etched’s mark. Etched’s $1 billion in unnamed contracts, at the same evidence tier, carries three times the mark.

Give the bull case its full arithmetic, because it is coherent. If $1 billion in contracts converts into revenue and holds, a $10.3 billion mark is about 10x revenue, against the 186x the public market just paid for Cerebras’s audited $510 million. On that math Etched is the cheap name in the sector. The private set offers no revenue multiple to check against, because none of the privates discloses revenue anymore, and that absence is itself the second finding: the sector stopped disclosing ARR in the same year its marks ran furthest ahead of it.

And the bull case has assets the arithmetic alone does not carry. NVIDIA paid $20 billion for Groq’s technology and its founder, one inference architecture, at the end of last year; Etched’s entire company, factory and bench and contracts included, is marked at half that. If strategic buyers price inference silicon, the sector already printed a floor at double this mark. Sequoia’s participation, likewise, is evidence that a dataroom existed and was read by someone whose job is pricing risk; per the company, the round’s investors are people who actually tried the hardware, and Sequoia has published its thesis in its own words, calling Etched “the machine that builds the machine” and claiming “Pareto dominant performance on industry-standard throughput-interactivity curves.” That is the bull case’s full shape: plausible, priced, and unverifiable from outside, the exact mirror of the case above. The reader’s job, and this desk’s, is to notice that the price already assumes the bull version. The entire defense hangs on the conversion. Peer ARR figures are reported revenue. Etched’s figure is a promise with signature blocks, no named buyer, no disclosed terms, no schedule. You can believe the conversion. You cannot yet verify any piece of it.

One honesty note on the table. Every mark is a 2026 transaction except Tenstorrent’s, which is its November private round pending the strategic outcome; Cerebras now trades daily, so its cell moves with the market. The ARR column is thin by design, not by oversight: SambaNova’s last recorded figure predates its SN50 and both 2026 rounds; Groq’s most-quoted number was a target it never booked; Tenstorrent discloses none. One basis note: the funding cells carry what the cited articles report, which for SambaNova, Groq, and Tenstorrent is their 2026 rounds rather than lifetime totals, so their multiples run on recent money; on lifetime totals all three sit lower, which only widens the gap the table shows. The specific multiples are directional. The qualitative gap, Etched at zero disclosed revenue while every peer has shipped, holds regardless of which exact numbers you use.

Traction and hiring: the checkable edges

Start with who is actually named in the record, because the list is short and entirely upstream. TSMC as fabrication partner. Rambus for HBM memory-controller IP and integration support. Both are suppliers with names. Downstream, where customers live, the data is silent: the $1 billion in contracts attaches to no buyer, no segment, no disclosed terms. The page’s phrasing is “co-design decisions made hand-in-hand with leading AI companies, cloud providers, and hyperscalers”, which is a customer list with the names removed, and the removal is the information.

Hiring is the checkable edge, and it is informative. The job board lists 108 open roles: Software at 29, ASIC at 23, Platform at 20, Production at 14, Operations behind those. The sample titles tell more than the counts: design verification engineer, physical design engineer, PCB layout engineer, infrastructure software engineer, a technical recruiter pointed at supercomputing and machine-learning talent. Design verification engineers are hired when you are preparing to send a chip to the fab. Physical design engineers implement the layout. PCB layout engineers design the board the chip sits on. This is an implementation-phase pattern, not an architecture-phase one. You do not hire PCB layout engineers for a concept.

The ASIC-to-Software split, 23 to 29 out of the 108 open roles, is notable on its own. For a company whose differentiation is hardware, more open software roles than chip-design roles suggests significant inference-serving infrastructure behind the silicon: schedulers, kernel drivers, performance profiling. That is consistent with the co-design claim, and it means the public evidence, while thin, is at least internally consistent.

The locations carry the same signal: 91 of the 108 roles sit in San Jose, 10 in Taipei where the foundry’s fabs are, and 6 in Austin, an outpost the data does not explain. Staffing manufacturing operations next to the fab is supply-chain proximity you can count.

The federal labor-certification data runs the other way. Certified LCA filings in the wage database: a lone electrical engineer petition, 2025-Q3, San Jose, at $217,630. Set that beside the claimed 400 engineers and the possible readings are that the team is overwhelmingly domestic hires, plausible for senior silicon talent, or that the headline count runs ahead of the roster, also plausible. What the lone filing does establish is tier: wage level IV, the top H-1B band, at or above the 90th percentile for the role in the metro, a senior hire’s paperwork rather than a bargain one. On count the data cannot distinguish, and the desk declines to guess. What remains is narrow: the foreign-worker paper trail is thin against the claimed headcount.

The public code surface is small and mostly beside the point: 13 repositories and 2,122 stars across the org. The recorded gap between claimed engineers and accumulated stars, 430.5 percent, compares a payroll number to a popularity contest, a category mismatch for a hardware firm whose crown jewels are RTL and interconnect specifications rather than public repos.

The patent record: three granted

The company’s own patent policy page lists three granted U.S. patents, 12,306,903, 12,361,091, and 12,361,262, plus published applications including US20240378175A1, “Multi-Chip Systolic Arrays”. The portfolio as described covers matrix multiplication, tensor operations, and systolic array architecture, the core of an inference ASIC. Not an empty file cabinet.

One note, because the method cuts both ways. The page is self-reported, the class of claim this desk holds at a discount, but patent numbers are also the one kind of claim a reader can verify against the issuing office in minutes, which is why the correction stands on it. Scale it honestly at the same time: three granted patents is modest relative to a $10.3 billionvaluation in a sector where established incumbents hold portfolios in the dozens or hundreds. The read here is “early-stage IP, actively expanding”.

Three claims, three locked doors

The terms of service are standard for pre-release hardware evaluations. Every chip company with an early-access program has an NDA. That is not the finding. The finding is the parallel:

• Valuation: $10.3 billion claimed. No filing anywhere in the data; the price exists on the company’s own pages alone.

• Contracts: over $1 billion claimed. Customers unnamed, bound by confidentiality.

• Performance: SOTA claimed. Etched’s own terms state, verbatim: “Published Performance Data may be based on extrapolations, projections, estimates, simulations, or limited testing data, and may represent expected or anticipated performance rather than actual measured results under all conditions.”

Three claims, three verification gaps, each structured so the claim cannot be checked from outside. Valuation has no filing because none of these rounds ever produced one. Contracts have no names because customers are bound by confidentiality. Performance has no benchmark because the terms prohibit it and the hardware may be pre-release anyway. The information architecture is one-way: Etched can publish anything, and no one else can publish anything at all. “Early customer tests show SOTA throughput” is a claim only Etched can make and only Etched can verify. When the lawyers and the marketers share a page, believe the lawyers. They are the ones with exposure.

The position

This is not a company selling vaporware. The evidence points to a real hardware company: a credible team drawn from NVIDIA, Google TPUs, Broadcom, SK Hynix, and TSMC, serious investors including a strategic memory manufacturer and a quantitative trading firm, an implementation-phase hiring pattern, three granted patents, and fabrication claims consistent with the role mix. Those are not the attributes of a slide deck.

The pedigree runs to the top of it. Uberti worked compiler and AI-model performance at OctoML and left Harvard as an undergraduate to found the company; compiler optimization for inference is the exact skill set a custom-ASIC strategy runs on. The dropout origin story is a character marker, not a technical guarantee. It demonstrates conviction, not competence.

The question is whether the evidence justifies the same tier as companies that have crossed the product, revenue, and customer lines. The 2.9 percent Series C stake means sophisticated investors accepted a small ownership slice at a high price. The round shrank from $500 million to $300 million while the valuation doubled. And the absence of any public benchmark or named end customer is the sharpest difference between Etched and its valuation peers. The hardware industry routinely prices design wins that have not yet converted to purchase orders; the design wins here are real in pattern but unnamed in the record.

The desk’s allocation answer, stated as structure rather than prediction: a secondary purchase at the market price, never the round. This month complicated the cheap half of that sentence: at the share count that reconciled the August 10 print, the latest tape, $237.90, values the company at roughly $13.3 billion, a premium to the claimed mark, so the tape is no longer the obviously cheaper instrument. It remains the observable one, and observable is the honest word for it: no preferences, no lockups, a price you can watch even when you cannot acquire size, and a volatility that discounts its own prints. The round buys preferred with rights the tape does not carry and a stake size the tape cannot sell you. The narrative’s cost here is not a price gap. It is the ownership gap: 2.9 percent at most, for the entire syndicate combined, against the 15 to 25 percent a lone lead ordinarily takes. The buyer of the narrative should pay that cost, not you.

What changes the calculus

Three things would shift the picture.

A benchmark. If Etched published its chip running a standard transformer workload at a specific throughput and power envelope, the market would have something concrete to price. One number would do more to validate or undermine the $10.3 billion than any investor signal. Note the constraint: the terms of service make this the only path. Evaluation customers cannot publish results, so a third-party benchmark requires Etched’s consent or a breach of contract. The picture changes only when Etched permits it.

A named end customer. Not TSMC the foundry. Not Rambus the IP licensor. A company committed to buying Etched’s chip at a specific volume and price. One named buyer would shift the risk calculus entirely. The $1 billion in contracts is a number on a company page; a named buyer is a number with a signature behind it. The confidentiality terms mean the name appears only if Etched lets it appear. Resolved, Aug 19: Jane Street, on the round’s lead ticket.

The share count. If the fully diluted count is near 56 million, the August 10 print reconciled and the latest print implies the market already pays a premium to the round. If it is near 43 million, the latest print reconciles and the earlier prints were discount marks. If it is neither, both readings are wrong. That data point’s absence is structural. The new round prices a third count into the fork: $242.16, the print after the announcement, reconciles $21 billion at roughly 87 million shares. Three prints, three counts, no disclosure.

Two of the three arrived within a day: a named customer and testimony in place of silence, both wearing the same name. The share count still gates the price, and the relationship between $21 billion claimed and $242.16 traded is an open question the public record does not answer, and Etched’s legal framework keeps open. The percentages survive either way, as ceilings: Sequoia’s 2.9 percent, Jane Street’s 3.3 percent, one division each, and every unanswered structural question can only shrink them. Everything above is annotation.

How to verify

If you want to check the work: progress page | terms of service | patent policy | Hiive | funding record | investor page | code org | job board with 108 roles | wage database. Start with the napkin: $300 million into $10.3 billion is roughly 2.9 percent of the company.

Every number traces to a source. The analysis is ours. - The SOMEN Desk