In the quarter that ended in late April 2026, Nvidia — the most valuable company in the history of money — booked a profit, and roughly twenty-seven percent of it did not come from selling anyone a chip. It came from writing up the value of stakes it holds in the very companies that buy its chips. Not cash a customer paid. A number Nvidia wrote about something Nvidia owns. In the same three months, its share of the accelerator market it is supposed to dominate outright fell again. Part 1 followed the token into the building. Now we follow it into the silicon — the one node the state loved enough to exempt from tariffs — and find that even here, at the richest rung on the ladder, a good chunk of the richest company's profit is a mark.
Part 2 of The Stack. We are one rung up from the shell. Same two questions, asked of the chip layer: who books the equity, and who eats the subsidized loss? Here the equity story is Nvidia's eroding moat and its increasingly circular profit; the loss is a depreciation assumption that decides whether hyperscaler earnings are real, and a rental market quietly repricing the risk onto whoever doesn't own the chip.
/ 01Twenty-Seven Percent of the Profit Is a Mark
Start with the number, because it is filed, not alleged. In Nvidia's own 10-Q for the quarter ended April 26, 2026, the line called “Other income (expense), net” swung on the back of $13.4 billion of unrealized gains on publicly-held equity securities and $2.6 billion on non-marketable equity securities — stakes in AI companies whose value went up on paper.1 Goldman Sachs, quoted by Cembalest, put that “other income” at 27% of Nvidia's Q1 profit — against 60% at Google and 51% at Amazon in the same quarter.2 This is the exact fact the Computedollar's sixth essay built a wall around, and it is worth restating in the tape-reader's register: a large share of the most profitable companies' profit is one company marking up its bet on another.
The circle is the point. Nvidia sells chips to the labs and the clouds. The clouds and sovereign funds pour capital into the labs. The labs spend that capital buying Nvidia chips and renting cloud compute. And Nvidia's stake in the labs gets marked up as each new funding round — partly enabled by the compute relationship — validates a higher valuation. Everyone's revenue is someone else's capex. Everyone's profit is partly a mark on everyone else's private valuation.
Nvidia saw this critique coming and, to its credit, argued its own book in a November 2025 note to analysts: it invested $4.7 billion in private companies in 2025, about 3% of revenue; those companies “primarily raise capital from third-party sources other than Nvidia” and “predominantly generate revenue from third-party customers rather than Nvidia”; days-sales-outstanding is a stable 53 days; there is no “extensive vendor financing.”3 This is precisely the kind of source Keene-style reading exists to isolate: a company defending its own accounting. Take it seriously and stress-test it, don't repeat it. The 10-Q confirms the shape of the thing — the gains are real and SEC-disclosed — while also confirming what they are: predominantly unrealized marks on holdings in other AI companies, not cash collected from an arm's-length buyer.1
An unrealized gain on a rising asset is still value; if OpenAI and Anthropic are worth what their latest rounds say, Nvidia's mark is conservative, not fictional. Buffett himself carries billions in unrealized equity gains and no one calls Berkshire a mirage. The distinction that matters is auditability and liquidity: Berkshire's marks are on public tickers you can sell into a deep market tomorrow; Nvidia's largest AI marks are on private companies that file nothing, priced by the same rounds the marks are meant to validate. That is a real difference in the quality of a profit dollar, and it is the difference this section is flagging — not that the dollar is fake, but that it is softer than the word “profit” implies.
/ 02The Moat Drains in the Report That Prices It Permanent
Here is the genuinely external thing in Nvidia's story — its pricing power on the chip itself — and by Cembalest's own chart it is leaking. Nvidia's share of accelerator revenue slips from about 85% in 2023 to an estimated 75% in 2026, as the hyperscalers build their own silicon and report total-cost-of-ownership reductions of 30–40% against merchant GPU fleets.4 Every point of share that migrates from Nvidia's 75%-plus-gross-margin GPUs to a hyperscaler's in-house chip is margin the hyperscaler keeps instead of paying as what amounts to a merchant tax. And the workload mix is pushing the same way: inference now exceeds training as a share of AI compute, and inference is exactly the high-volume, well-characterized job that custom chips are built to do cheaply.
Follow the money into the picks-and-shovels beneficiary and you meet Broadcom, the leading merchant designer of these custom chips. Its AI-semiconductor revenue hit $10.8 billion in a single quarter, up 143% year over year, with guidance to triple sequentially and a reiterated 2027 target above $100 billion; it names six core custom-chip customers, reportedly including Google, Meta, OpenAI, and Anthropic.5 Google's Ironwood TPU, Amazon's Trainium, Microsoft's Maia, and Meta's MTIA are the in-house line. Anthropic has committed to up to a million Google TPUs — reported as the largest deal in Google Cloud's history — and to running Claude on AWS Trainium for a decade, which Cembalest calls “the strongest third-party endorsement of ASICs to date.”6
Notice a tell inside the tell. Broadcom recently shifted its stated model from selling full integrated AI systems to selling “chips only” — stepping back from taking system-level inventory and integration risk on its customers' capex bets.5 And Google and Amazon rent their ASICs out as a cloud service; Microsoft and Meta keep theirs entirely captive, no public instance type.7 Everyone in this layer is quietly arranging to hold the margin and hand off the risk — the same instinct we watched Jacobs act on one rung down.
The often-cited TPU shipment ramp (4.3 million units in 2026 → 10 million 2027 → 35 million-plus 2028) is an analyst/press projection, not a Google-disclosed figure, as is AMD's ~$7.2 billion 2026 MI400 revenue.7 Useful as directional texture; not to be printed as company guidance. AMD's filed numbers are strong on their own — Q1 2026 data-center revenue $5.8 billion, Q2 guided to $11.2 billion, with Meta committing to install up to 6GW of Instinct GPUs.8
/ 03The Assumption Nobody Outside the Company Can Audit
Now the sharpest “who eats the risk” number in the entire stack, and it hides in a place most readers never look: the useful-life assumption a hyperscaler picks for its chips. Depreciate a $40,000 GPU over three years and it costs you ~$13,000 a year in reported expense. Depreciate it over six and it costs ~$6,700. Same chip, same cash out the door — but the six-year choice roughly halves the annual hit to reported profit. Over the last several years, every hyperscaler quietly stretched the assumption.
| Assumed life (yrs) | 2020 | 2022 | 2023 | 2025 |
|---|---|---|---|---|
| Meta | 3.0 | 5.0 | 5.0 | 5.5 |
| 3.0 | 4.0 | 6.0 | 6.0 | |
| Microsoft | 3.0 | 6.0 | 6.0 | 6.0 |
| Amazon | 4.0 | 5.0 | 5.0 | 5.0 |
| Oracle | 5.0 | 5.0 | 5.0 | 6.0 |
Cembalest modeled what happens if you reverted to a three-year schedule for the GPU and networking gear bought since ChatGPT launched. The answer, built on public 10-K data, not a leak:
| Amazon | Meta | MSFT | Oracle | ||
|---|---|---|---|---|---|
| Δ operating margin | −9% | −8% | −6% | −7% | −14% |
| Δ pro-forma EPS | −7% | −7% | −7% | −6% | −17% |
Seventeen percent of Oracle's earnings per share turns on where it draws a line nobody outside Oracle can independently verify. Six to seven percent for the others. This is not an accounting technicality; it is the difference between a reported number and an estimate wearing a reported number's clothes.
Cembalest supplies the counter-argument himself, and it is a real one: A100-class chips still run at high utilization and positive margin well past three years; older GPUs get repurposed for inference or resold into emerging markets; swapping them out early is expensive because power and cooling are engineered to a generation. Trump's decision to re-permit H200 sales to China even props up the resale value — and therefore the depreciation assumption — of older chips.9 So the six-year life may be economically honest. But that is exactly the trap: it is a judgment call that determines whether reported EPS is real, made by the only party who can't be neutral about the answer. Buffett's whole objection to “adjusted” earnings is this: depreciation is a real cost, and the moment its size becomes a management choice, the earnings number becomes an opinion.
/ 04The CoreWeave Tell
If depreciation is where the risk hides on the balance sheet, the rental market is where it shows up in daylight. Cembalest's December 2025 table of GPU spot-rental rates has H100 neocloud pricing down 26% year over year, A100 down 22–26%, hyperscaler H100 down 21%.9 Prices for renting the chip are softening even as everyone insists the chip is scarce.
And the clearest tell is CoreWeave, the neocloud that rents GPUs it mostly financed with debt. It fell more than 50% from its June 2025 IPO peak by December, and by July 1, 2026 traded around $86.72, down roughly 48% from its high — with a 13% single-day drop tied to a report that Meta is building competing capacity, a securities class action alleging it overstated its ability to meet demand, and the CEO selling $32.9 million of stock the week before.10 The important discipline here: CoreWeave's business is still growing fast — revenue up 112% year over year to $2.08 billion in Q1 2026.10 This is not a revenue collapse. It is the market repricing the risk of being the party that owns a depreciating chip on leverage while its rental price drifts down. The risk transfers to whoever doesn't own the chip outright — which is the neocloud, not the chip designer.
A separate mid-2026 source reports that one-year reserved-contract H100 pricing rose about 40% (Oct 2025 → Mar 2026) — the opposite direction from the spot table above.11 These may both be true: oversupplied on-demand/spot rates can fall while scarcity pricing for guaranteed one-year capacity rises. But I could not reconcile them to a single primary index this pass, so I am not going to hand you a clean “rental rates are up/down” sentence. When a trend has two honest directions depending on the contract, saying so is the reporting.
/ 05Eleven Days of Gas
Underneath every chip in this essay is a single building on a single island, and it is worth pausing on how much of the world economy balances on it. TSMC posted Q1 2026 revenue of $35.9 billion, up 40.6%, at a 66.2% gross margin, with high-performance-computing and AI now 61% of its revenue, and raised full-year guidance above 30% growth.12 Eight of the ten most valuable companies on Earth depend on it. Semiconductor trade has overtaken crude-oil trade as a share of global GDP.
And the substrate is a hostage's balance sheet. Taiwan imports about 90% of its primary energy; it let nuclear generation fall from 50% of electricity in the 1980s to 5% today, backfilled with imported LNG — of which it holds ten to eleven days.13 It imports 60% of its food and 67% of its calories. LNG suppliers in Singapore now write act-of-war clauses into Taiwan delivery contracts. Cembalest, who says he “almost didn't believe these figures,” calls it “the most blockade-sensitive advanced economy in the world.”13 When an index fund holds the Nasdaq 100, this is the object at the bottom of the position: a leveraged bet that an island with eleven days of fuel stays perfectly accessible forever.
The onshoring answer is real but partial. Commerce Secretary Lutnick set a goal of 40% of U.S. chip demand onshored by the end of 2028; Cembalest's own supply-side math has the U.S. reaching, generously, 30–35% of advanced-node production by 2028–2030 — still “highly reliant on Taiwan.”14 The reason it's slow is cost, and one number captures it: a J.P. Morgan table puts depreciation per wafer at $1,500 in Taiwan versus $7,289 in the U.S. — a nearly 5x penalty for making the same wafer domestically.9 That figure is presented without full methodology and should be treated as J.P. Morgan's own model, not a TSMC disclosure — but even discounted, it explains why the island stays load-bearing.
China is climbing the other side. Huawei's Ascend 910C improved foundry yield from ~20% to ~40% in a year and now accounts for 75%-plus of China's domestic AI-chip production.15 Per chip, Nvidia's B300 still beats it by 2.3–3.5x, and it takes roughly seven racks of 910C to match one rack of B300 — so Huawei is competing at the cluster level instead, trading efficiency for scale and cheap domestic power. And the export-control bite shows up in ASML's numbers in real time: China fell from 36% of ASML system sales in Q4 2025 to 19% in Q1 2026.16
/ 06Two Weeks
I want to show you one thing going stale in real time, because it is the whole ethic of this series compressed into a single example.
Cembalest's Outlook 2026 report is dated January 1, 2026. In it, the semiconductor Section 232 tariff investigation is described as “ongoing,” and the working assumption across the market was that chips were largely tariff-exempt. In fact, Commerce had already transmitted its final national-security finding to the President on December 22, 2025, and on January 15, 2026 — two weeks after the report's date — the White House imposed a 25% Section 232 tariff, Proclamation 11002, on advanced semiconductor articles, naming Nvidia's H200 and AMD's MI325X as covered examples.17 “Semis are tariff-exempt” was a primary-sourced, carefully hedged claim from one of the most data-literate people in finance — and it was out of date inside a fortnight.
“The most rigorous number in the world is still only true until the next filing. Date your facts or they will date you.”
This is not a gotcha at Cembalest's expense; his hedging (“a warning label,” “subject to revision”) is exactly right. It is the reason this series dates every claim, flags estimate-versus-reported on every contested figure, and would rather tell you a rental-rate trend has two honest directions than hand you a clean sentence that's wrong by Tuesday. The export-control picture proves the same point: Chinese firms ordered more than two million H200 chips, but the U.S. State Department reportedly stalled ~400,000 already-approved units, and Chinese customs separately told domestic firms not to import them — so on part of this trade, Beijing is now the binding constraint, not Washington.18 Both of those claims rest on secondary reporting of unnamed sourcing, and I'm telling you that too.
/ 07Follow the Equity, Follow the Token
The ledger for the silicon layer. Follow the equity: Nvidia's profit is increasingly a mark it writes on private companies it owns, even as its revenue share of the market it dominates erodes from 85 toward 75. Broadcom captures the same boom more cleanly — selling design and manufacturing to the hyperscalers at $10.8 billion a quarter, holding no stakes in its customers — but even Broadcom is retreating to “chips only” to avoid holding execution risk. The equity in this layer is real, and richer than the shell's fee; but a meaningful slice of it is paper marks and management assumptions, not cash from strangers.
Follow the token: the depreciation-life assumption is the single biggest lever on whether hyperscaler AI economics are as good as reported — a three-year reversion costs Oracle 17% of EPS on J.P. Morgan's own model. Softening spot-rental rates and CoreWeave's de-rating are the visible symptom of oversupply risk transferring onto whoever owns the depreciating chip on leverage. And all of it balances on an island with eleven days of gas that got a tariff exemption precisely because the state judged the silicon too strategically precious to tax — the exact opposite of the transformers one rung down, which got taxed and back-ordered for four years.
So the boom's two richest layers are now on the table. The shell books a fee and holds the obsolescence at arm's length. The silicon books real margin threaded with paper marks and unauditable assumptions. In both, the same instinct repeats: hold the upside, hand off the risk. The next rung up is where that instinct gets industrialized — the cloud, which rents you all of this on a contract, and carries roughly two trillion dollars of promised future revenue against counterparties whose revenue does not exist yet. Part 3 follows the token into the toll booth.