In the first quarter of 2026, Alphabet reported $62.6 billion of net income and Amazon reported one of the strongest quarters in its history. Read past the headline and a large share of both numbers turns out to be the same kind of entry: not a customer paying a bill, but a company marking up the value of a stake it holds in Anthropic, a private lab that files no public financials of its own. Alphabet's own 10-Q attributes $28.7 billion of that net income to the markup. Amazon's own quarter carries a separate, similarly triggered $16.8 billion unrealized gain on the same company's stock. Both numbers are filed, not alleged. Both come from a stake built for single-digit billions in cash that a single funding round, months later, revalued into nine figures more. Part 2 found Nvidia marking up its own bets on the labs for 27 cents of every profit dollar. Now go one rung up, into the labs themselves, and the mark isn't a footnote to the story — it is the story.
Part 4 of The Stack. We are one rung up from the cloud, whose ~$2 trillion of disclosed backlog is, by the plan's own accounting, roughly half committed by the four labs this article covers. Same two questions, asked of the layer everything else in the series exists to feed: who books the equity, and who eats the subsidized loss? Here the equity is a ring that closes on itself — hyperscaler stakes in labs, marked up by the same capital the hyperscalers supply those labs to spend. The loss is the labs' own burn, run against commitments an order of magnitude larger than their revenue.
/ 01The Ring
Start with the two numbers that anchor this entire article, because unlike most of what follows in this series, they do not need a hedge. Alphabet's Form 10-Q for the quarter ended March 31, 2026 reports “Other income and expenses, net” of $37.7 billion, driven by a $36.3 billion net unrealized gain on non-marketable equity securities.1 Of Alphabet's $62.6 billion total net income that quarter, $28.7 billion — nearly half — traces directly to the markup on Alphabet's stake in Anthropic, a position built for roughly $3 billion in cash between 2023 and 2025 and sitting at an estimated 14–15% of the company as of Q1 2026.2 At Anthropic's ~$965 billion Series H valuation, that stake marks to somewhere in the $135–145 billion range — a roughly 45–48x unrealized return on the cash Alphabet actually wrote a check for, sitting inside its reported profit as a non-cash entry.
Amazon's version of the same event is separate and independently disclosed. In the same quarter, a tranche of Amazon's convertible notes in Anthropic converted into preferred stock when Anthropic's Series G round priced — the kind of “significant financing event” that, under ordinary accounting rules, requires marking the position to the new fair value. That single conversion produced a $16.8 billion unrealized gain, more than half of Amazon's total pre-tax income for the quarter, and pushed a cumulative ~$8 billion cash investment to a mark-to-market value above $70 billion.3 Neither entry is fraud, an accounting trick, or off-books. Both are required, GAAP-compliant responses to a triggering event. What they are, in plain English, is two of the seven most valuable companies on earth reporting as profit their own opinion of a private company's worth — an opinion that happens to be substantially informed by capital both companies themselves helped put into that private company.
Zoom out from the two anchor filings to the chart that puts them in context. Goldman Sachs' own tally, reproduced by J.P. Morgan, has “other income” — substantially driven by markups on stakes in Anthropic and OpenAI — running at 60% of Alphabet's Q1 2026 profit and 51% of Amazon's, against Nvidia's 27% one rung down.4 The reason the ring is worth drawing rather than just stating is that it closes on itself with almost no gap: hyperscaler compute revenue funds lab spending; lab spending, credibly promised against real usage, supports a higher valuation at the next funding round; the hyperscaler holding a stake marks that stake to the new price; the mark becomes reported profit; the stock re-rates on that profit; and cheaper capital, borrowed against a richer balance sheet, funds the next round of chip orders and compute commitments that starts the cycle again.
Nvidia sells the chips
Multi-year, multi-gigawatt commitments to OpenAI, Anthropic, and xAI — the labs' single largest capital outlay.
The labs run on hyperscaler capital
Google and Amazon supply cloud compute to Anthropic and hold equity stakes in it at the same time — the customer and the shareholder are the same company.
A new round prices the lab higher
Anthropic's valuation moved from $350–380B to $965B in under three months, on rounds that closed while its compute relationships with Google and Amazon were expanding.
The hyperscaler marks its stake to the new price
A GAAP-required, non-cash entry — $28.7B at Alphabet, $16.8B at Amazon — that lands in the same quarter's income statement as “other income.”
The mark becomes profit; cheaper capital funds the next round
Reported earnings rise, the stock re-rates, and the hyperscaler's now-cheaper capital funds the next tranche of compute and chip orders — back to step one.
Part 2 made this case for Nvidia and it applies with equal force here, maybe more. Anthropic's revenue run-rate genuinely moved from roughly $9 billion to $47 billion inside seven months; a valuation that rises on that kind of growth is not manufactured, and the accounting rule that forces Alphabet and Amazon to mark their stakes on a triggering event exists precisely so investors see the gain when it happens rather than only when the company eventually goes public. Buffett carries billions in unrealized gains on Berkshire's public equity book every quarter and no one calls that fictional. The honest distinction, as in Part 2, is liquidity and independence of the pricing signal: Berkshire's marks are set by strangers trading a public ticker every second; Alphabet's and Amazon's marks on Anthropic are set by the same small circle of investors — some of them the marking party itself — negotiating a private round every few months.
Before this reads as a closed loop that only tightens, one dated fact runs the other way. Nvidia's September 2025 pledge to invest up to $100 billion in OpenAI, tied to 10 gigawatts of Nvidia systems, did not survive contact with negotiation: by March 2026, Jensen Huang confirmed the full amount was “probably not in the cards,” and what actually closed inside OpenAI's Feb 27, 2026 funding round was a $30 billion Nvidia equity stake — three-tenths of the headline number.6 Counterparties do renegotiate. The ring bends before it breaks. It just hasn't broken yet, and the aggregate is still large: Bloomberg's own tally of the named circular deals across Nvidia, OpenAI, Microsoft, Oracle, AMD, Google, and Amazon puts total value north of $800 billion — a figure that survived the Nvidia contraction with room to spare.7
/ 02The Businesses & the Burn
Now the businesses actually generating the revenue the ring is pricing. Start with the gap that Cembalest calls the single biggest individual corporate risk in the whole AI story — bigger, in his words, than Nvidia, despite the company in question being private. OpenAI's annualized revenue run rate crossed $25 billion by March 2026, on 50 million consumer subscribers and 9 million business users.8 Set against that: $1.4 trillion in disclosed infrastructure commitments — the $500 billion Stargate program with SoftBank and Oracle, a separate $300 billion, five-year Oracle cloud deal, a $38 billion AWS Bedrock pact signed in March 2026, and an Azure consumption commitment analysts estimate above $200 billion over its remaining term.9 Divide one by the other and the ratio runs past 56x. OpenAI itself, on Feb 20, 2026, partially walked the number back — telling investors its real compute-spend target through 2030 is closer to $600 billion, not $1.4 trillion.10 Read that reset two ways at once, because both are true: it is evidence the $1.4T figure always had more ceiling than binding-obligation character to it, and it is evidence that even OpenAI's own leadership found the original number indefensible enough to retract inside four months of it becoming a media flashpoint.
OpenAI is reportedly on track to lose about $14 billion in 2026, nearly triple its 2025 loss, while projecting $100 billion of revenue by 2029 — the standard early-stage-SaaS logic of losing more before winning bigger, scaled to absolute dollar figures no SaaS company has ever produced.7 The tension surfaced publicly in late April 2026, when reporting broke of a rift between CEO Sam Altman and CFO Sarah Friar over whether revenue can plausibly support the commitments. Altman's response, on Brad Gerstner's BG2 podcast, was characteristically blunt: “If you want to sell your shares, I'll find you a buyer. I just… enough.”11
Anthropic's numbers move on a different axis entirely — not the size of the commitment against revenue, but the speed of the revenue itself. Its disclosed run rate went from roughly $9 billion at the end of 2025 to $14 billion in February, $19 billion in early March, $30 billion in early April (an “80x growth” claim the company made itself), and $47 billion by mid-May.12 Funding kept pace: a $30 billion Series G at a $350–380 billion valuation closed Feb 12, 2026; a $65 billion Series H at a $965 billion post-money valuation closed roughly ten weeks later, on May 28.13 That is a 2.5–2.7x valuation jump in under three months, and it is the specific repricing event that produced Alphabet's and Amazon's marks in Section 01. Anthropic confidentially filed a draft S-1 on June 1, 2026, days after the round closed.12
| Date | Event | Figure |
|---|---|---|
| End 2025 | Run-rate revenue | ~$9B |
| Feb 12, 2026 | Series G closes | $30B raised, $350–380B valuation |
| Early Mar 2026 | Run-rate revenue | ~$19B |
| Early Apr 2026 | Run-rate revenue (“80x growth”) | ~$30B |
| May 28, 2026 | Series H closes | $65B raised, $965B valuation |
| Mid-May 2026 | Run-rate revenue | ~$47B |
| Jun 1, 2026 | Confidential S-1 filed | — |
Both labs disclose the harder, less-fakeable number underneath the growth: gross margin, meaning revenue exceeds the direct compute cost of serving it. Anthropic runs at roughly 44%, spending about $0.71 per revenue dollar on compute in Q1 2026 and projecting improvement to $0.56 in Q2, with a company target of 77% gross margin by 2028 that this article treats as a projection, not a result.14 OpenAI's gross margin is reported around 42%, per the Financial Times.15 Both numbers matter more than headline revenue growth, because a lab that cannot clear gross margin can never reach operating profitability no matter how large it becomes. Both labs' self-projected paths to cash-flow-positive — OpenAI in 2029–2030, Anthropic in 2028 — are, in Cembalest's own words, worth taking at face value only with the appropriate discount:
“Speculative, uncertain and subject to revision…” — Cembalest, on both labs' self-projected cash-flow-positive dates, “Semiquincententacles,” June 2026
xAI's numbers are the roughest of the four American labs, and it is worth flagging exactly how we know them: xAI has never filed anything public of its own, and every figure below comes from SpaceX's IPO disclosures, filed because xAI and SpaceX are financially intertwined and reportedly discussing a $1.25 trillion combination.16 That related-party filing shows an operating loss of $2.47 billion on $818 million of revenue in Q1 2026 alone, with Q1 capex of $7.7 billion — an annualized run-rate near $31 billion; full-year 2025 losses were $6.4 billion on $3.2 billion of revenue.16 xAI raised $20 billion in a Series E that closed Jan 6, 2026 at a $230 billion valuation, upsized from an initial $15 billion target, with Nvidia itself among the investors.17
Google DeepMind is the hardest of the four to underwrite at all, and that opacity is itself a data point. Gemini's revenue is not broken out from Google Cloud and Search, so there is no standalone P&L to compare against OpenAI's or Anthropic's. What is disclosed: Gemini processes more than 16 billion tokens per minute via direct API, up 60% quarter over quarter; the Gemini app carries more than 750 million monthly active users; Google Cloud revenue grew 63% year over year to $20.0 billion in Q1 2026, with AI cited as the largest driver; and Alphabet has guided $180–190 billion of AI infrastructure capex for 2026.18 The most vertically integrated of the frontier players is also the one whose frontier-model economics are the least visible from the outside — worth stating plainly rather than treating as a research gap waiting to close, because it may not be closable from public filings at all.
Meta AI belongs in this section for a different reason: it just reversed itself. Meta formed Meta Superintelligence Labs in mid-2025 under former Scale AI CEO Alexandr Wang, and in April 2026 shipped Muse Spark — Meta's first proprietary, closed-weight flagship, breaking from the open Llama lineage that had made Meta the standard-bearer of the open-weight movement.19 The reported reason: Llama 4 Maverick had been overtaken on benchmarks by Chinese open models. Hold that thought. It is the hinge into the next section.
/ 03The Open-Weight Challenger
Here is the fact that makes Meta's retreat legible, and the one that should worry every closed lab in Section 02 more than any funding gap. By April 2026, per Artificial Analysis and OpenRouter data cited by J.P. Morgan, the best Chinese open-weight models — DeepSeek V4, Qwen, Kimi — scored within a few dozen Elo points of closed frontier models while costing 10–50x less per token.20 The single cleanest illustrative pair: on the Artificial Analysis Intelligence Index, Claude Opus 4.8 scores 56 and costs $3,700 to run the full benchmark task set; DeepSeek V4 Pro (Max) scores 44 and costs $186 — about 20x cheaper for a 21% lower score.20 LMArena's own June 2026 read, across 360-plus models and 6.8 million-plus blind votes, has the entire top tier spanning only about 55 Elo points — a band narrow enough that “frontier” is starting to describe a neighborhood, not a single address.21
The bench is deep and it is not just DeepSeek. Alibaba's Qwen3.7 Max debuted as the highest-ranked Chinese model on at least one major aggregator by mid-2026. Moonshot AI's Kimi K2.6 leads Chinese coding benchmarks at 80.2% on SWE-bench Verified. MiniMax's M3 edges ahead on science reasoning at 92.7% GPQA Diamond.22 Europe has its own entrant in France's Mistral, which raised a €1.7 billion Series C at an €11.7 billion valuation in September 2025 and, as of June 2026, was reportedly in talks for €3 billion at a €20 billion mark; its 2025 ARR exceeded $400 million against a 2026 target above $1 billion — two orders of magnitude smaller than Anthropic's, worth stating explicitly for scale.23 China's Zhipu AI (Z.ai) became the world's first listed frontier-LLM company via a Hong Kong IPO on Jan 8, 2026; the stock ran 24.6x by mid-2026, and its GLM-5 family was reportedly trained without Nvidia chips at all — a live demonstration that export controls reshaped China's hardware path without stopping frontier-adjacent progress.24
And Nvidia is hedging its own moat. Nvidia Nemotron — an explicit bet that if inference migrates to cheaper open models, Nvidia had better not be purely a merchant-hardware vendor locked out of the software layer — shipped Nemotron 3 Nano in December 2025 and Nemotron 3 Super in March 2026, backed by a co-development coalition of Black Forest Labs, Cursor, LangChain, Mistral, Perplexity, and others.25 The sharpest single proof point of what a fine-tuned open model can do against a closed frontier one: Ramp Labs, on May 7, 2026, reported that a 35-billion-parameter Chinese open model, post-trained on Ramp's own data via reinforcement learning, outperformed Anthropic's Opus 4.6 while running at Haiku-level latency on a single data-center GPU.26 Owned data plus a smaller open model beat a smarter frontier one — independent of how much compute the frontier lab has behind it. That is the load-bearing fact for Part 5.
Meta's retreat from open weights cuts both ways and this article won't pretend it resolves cleanly. Read one way, the leading open-weight lab in the West pulling back weakens the commoditization thesis: if open-weight economics can't sustain even a hyperscaler-scale R&D budget, maybe the open lane is a dead end. Read the other way, it strengthens the thesis: Meta didn't retreat from strength, it retreated because Chinese open models had already eaten Llama 4 Maverick's lunch on benchmarks — evidence commoditization from below is real enough to push a trillion-dollar company out of a lane it invented. Both readings are available from the same fact. This series won't manufacture a verdict the evidence doesn't support.
Meanwhile the closed labs are visibly repricing upward — the subsidy running in the opposite direction from the open-weight tier, and both forces operating at once rather than cancelling out. OpenAI doubled its own token prices between GPT-5.4 and GPT-5.5.27 Microsoft raised Copilot prices starting June 1, 2026 and cut its own internal Claude Code licenses, with reported price increases across Anthropic and Google models running 3–9x, and some users reporting hikes as high as 100x.27 Anthropic's own subscription-versus-API gap is the sharpest illustration of who is meant to absorb that repricing: a Claude Max 20x subscriber paying $200 a month would spend up to $8,000 a month if the same usage were billed at metered API rates — a 40x discount Anthropic is visibly trying to close by pushing high-usage customers off flat plans.27 The migration evidence is real, if early: Lindy AI's founder announced moving the company's entire AI service from Claude to DeepSeek, citing millions in savings; Coinbase's Brian Armstrong projects “80% of workloads will be running in 99% cheaper models” within a year and says Coinbase is already moving workflows that direction.27 Frontier tokens still command a real premium where it counts — Cursor and Anthropic charge 6x more for tokens that run 2.5x faster, and Cembalest's own caveat is that frontier models remain necessary for cybersecurity, distillation training, agentic reasoning, and scientific discovery.28 Commoditization at the low end, on this evidence, narrows the share of tokens that need the frontier tier. It does not yet eliminate demand for it.
/ 04Geopolitics
The commoditization story above is inseparable from a live fight over whether the open-weight challengers got there honestly. On Feb 12, 2026, OpenAI sent a memo to the House Select Committee on China accusing DeepSeek of building tooling — obfuscated third-party routers — to programmatically harvest OpenAI's own outputs for training, a direct violation of OpenAI's terms of service barring the use of its outputs to build imitation frontier models.29 Twelve days later, Anthropic filed its own technical disclosure naming DeepSeek, Moonshot AI, and MiniMax by name, quantifying roughly 24,000 fraudulent accounts and more than 16 million exchanges used to query Claude for apparent distillation purposes.30 This is the first time both major U.S. labs made quantified, named, public distillation allegations against specific Chinese competitors inside the same month — a real escalation from the general suspicion that predates this window into a formal federal complaint.
Export policy on the underlying chips is genuinely unsettled, and the executive and legislative branches are now visibly pulling in different directions. The Bureau of Industry and Security's Jan 15, 2026 final rule moved H200 and MI325X export licensing for China from presumption-of-denial to case-by-case review, subject to a 25% tariff on the transaction, a 50% volume cap relative to U.S. shipments, and mandatory third-party testing and KYC.31 In practice, the policy is now being constrained from both sides at once: Chinese firms placed orders for more than 2 million H200 chips, but the State Department reportedly stalled roughly 400,000 already-approved units pending further security review, while Chinese customs separately instructed domestic firms not to import H200s at all — Beijing, not only Washington, is now a binding constraint on part of this trade.31 Congress, meanwhile, wants a harder line than the administration's own rule: the AI OVERWATCH Act passed committee 42–2 and would impose a statutory two-year ban on Blackwell exports plus a congressional veto over future licenses — legislative appetite running more restrictive than executive policy, an open fight as of this writing.31
Layer sovereign AI on top and the same commoditization pressure driving enterprise migration toward cheaper open models is also driving nation-states — for sovereignty reasons, not cost reasons. France's Mistral is the EU's flagship, backed by a €109 billion national AI package. The UAE's G42 runs open-weight Arabic-language models anchored to the U.S. stack through a $1.5 billion Microsoft investment, with the first phase of a planned 5-gigawatt Abu Dhabi “Stargate” campus due in Q3 2026.32 And Zhipu's Hong Kong listing means China now has a public, tradeable frontier-lab equity vehicle before the United States does — OpenAI, Anthropic, and xAI are all still private as this article goes to file, mid-IPO-process on every one of them.
/ 05Follow the Equity, Follow the Token
Steelman this layer fully before landing the bear, because the growth is not manufactured. OpenAI's revenue run-rate more than doubled inside a year and crossed $25 billion. Anthropic's run rate went from $9 billion to $47 billion in seven months — a growth rate that would be remarkable at any company, private-market accounting conventions notwithstanding. Both labs clear a real gross margin, meaning the unit economics of serving a token are not fictional even if the path to operating profit is years out. And frontier capability still commands a genuine premium in the places that matter most — cybersecurity, agentic reasoning, distillation training, scientific discovery — a premium enterprise customers are still paying 6x for. This is not a fake business. It is a real, fast-growing, capital-intensive one.
Now the bear, and it is not subtle. Follow the equity and you find the cleanest, most auditable circular subsidy in the entire series: Alphabet and Amazon do not just sell compute to Anthropic, they hold equity in it and mark that equity up in the same reporting period their compute revenue from Anthropic grows — both effects landing in the same income statement, in the same quarter, disclosed in each company's own 10-Q. Follow the token and you find a burn that dwarfs the revenue funding it even after the industry's own most aggressive advocate quietly cut his number by more than half: $600 billion of compute spend against $25 billion of run-rate revenue is still a 24x gap, before counting the $1.4 trillion figure was ever supposed to represent. And follow the payback dates and you land exactly where Cembalest lands — OpenAI cash-flow-positive in 2029 or 2030, Anthropic in 2028, both, in his words, speculative, uncertain, and subject to revision.
“Everyone's revenue is someone else's capex. Everyone's profit is partly a mark on someone else's valuation.”
The frontier labs are, on the evidence gathered here, eating the single biggest cash loss anywhere in this stack — and that loss is being subsidized, quarter after quarter, by two hyperscalers' willingness to book an unrealized gain on the very lab whose compute bill they're also collecting. The ring bends when a counterparty pushes back, as Nvidia's $100B-to-$30B contraction shows; it has not broken. And underneath it, priced 10 to 50 times cheaper and closing the Elo gap by the month, sits an open-weight tier that doesn't need the ring to work at all — it needs a harness that can run it. That's the next rung up: the orchestration layer that decides whether a locked-in enterprise customer stays locked in, or becomes, transparently and at will, a price-shopper between Claude, GPT, and DeepSeek. Part 5 follows the token into the harness.