An Operating Profit, a Decade of Bills

Dylan Patel of SemiAnalysis, a firm that sells research to investors, told Alex Kantrowitz on Big Technology on October 7 that Anthropic's research spending, enormous as it is, had grown slower than its revenue, that the company had therefore reached profitability, and that he saw no reason for the trend to stop. Mia Silverio counted the other way two days earlier in Prof G Media: revenue up 1,088% in 2025 to $4.6 billion, a loss of $42 billion, about $34 billion of it a non-cash charge tied to convertible financing, and more than $8 billion of operating losses, roughly 60% above what SpaceX lost in the year before it listed, by her count. Steve Eisman widened the charge on The Real Eisman Playbook on October 11: with open-weight models gaining share, the world may conclude that the economics of building AI are deeply flawed.

On October 26 Anthropic made its registration statement public on EDGAR, and it shows that Patel and Silverio read their numbers correctly, for different periods. Silverio's year is there: 2025 revenue of $4,585 million, an operating loss of $8,060 million and a net loss of $41,965 million, $34,308 million of it from fair-value changes on convertible notes, derivatives and SAFEs. So is Patel's turn: in the quarter to June 30, 2026, revenue of $11,555 million produced GAAP operating income of $974 million, the only operating profit in the six quarters the prospectus lists. The same quarter carried a net loss of about $66.3 billion, after a $59.6 billion charge for changes in the fair value of convertible notes, derivatives and SAFEs and $8.4 billion of other expense.

The question the filing hands its buyers is whether Anthropic's revenue is growing fast enough to pay for the computing it has already signed for, at the value the market is reported to expect. Part twelve argued that only a prospectus says what a lab is worth; this is the first. A disclosure: DeepStack is built with Anthropic's models, so the critics' case is set out before the bull case and OpenAI is held to the same ruler.

The ruler

The prospectus settles whether the revenue exists. Quarterly revenue rose from $376 million in the first quarter of 2025 to $11,555 million in the second quarter of 2026, 2.4 times the first quarter's $4,731 million, while compute and infrastructure expense fell from 148% of revenue in the quarter to September 30, 2025 to 48% in the quarter to June 30, 2026. For the first half as a whole it was 54.9%. A note that regroups costs by function puts cost of revenue, mostly compute serving paying customers, at $2,380 million for the half: a gross margin of 85.3%, against 39.5% for 2025 on the same note. Channel fees sit in sales and marketing. Research and development, which carries most of the rest of the compute, rose to $9,741 million from $1,850 million a year earlier, 5.3 times, while revenue rose 14.0 times. On the trend, Patel has the better of it; on 2025, Silverio's figures stand.

The bill is what the filing leaves open, and the company describes the mechanism itself. It commits to compute 12 to 18 months before it arrives, runs it as one pool for research, training, serving and internal use, and books the expense as capacity comes online, ahead of the revenue it will support; it expects its operating margin may compress when that happens. Four contracts carry fixed minimums Anthropic cannot cancel short of the other side's breach or default: at least about $114.5 billion with Google to July 2033, $110.0 billion with Amazon to April 2036 and $44.9 billion with Microsoft from November 2026 to May 2033, each regardless of actual usage, plus about $161.2 billion of rent on five-year Broadcom leases, $430.6 billion in all.

On a straight line over each term (Google about 7.3 years, Amazon 10, Microsoft 6.5, Broadcom 5) that is about $65.8 billion a year, 3.0 times the second quarter's compute spending at an annual pace and 1.42 times its revenue at that pace. The arithmetic is DeepStack's: the filing gives no payment schedule, payments start as capacity arrives, and part of these contracts already sits inside second-quarter spending, so the sum replaces today's bill rather than adding to it. Outside the four, up to $120.5 billion with SpaceXAI (formerly xAI) into 2029 would run about $3.1 billion a month once its capacity arrives, which Anthropic expects in December 2026, though either side may cancel on 90 days' notice after early 2027. Purchase commitments and leases not yet commenced went from about $105 billion at December 31, 2025 to about $832 billion with the contracts signed through October, by DeepStack's sum, which may count amendments twice; the filing's own total is blank.

Operations produced $3,929 million in the first half of 2026 and capital spending took $482 million, about $3.4 billion of free cash flow by DeepStack's subtraction, helped by the timing of payments to vendors. Cash and short-term investments reached $131.7 billion at June 30 after $110.8 billion of financing in the half.

The bill comes first

Silverio's sharper point is concentration: two customers supplied close to a quarter of 2025 revenue and Amazon and Google carried 47% of sales, too narrow a base for a company seeking a $2 trillion value. The prospectus bears out the customers: Customer A and Customer B each accounted for 12% of 2025 revenue. The channel is a separate risk. The filing counts 47% of 2025 revenue and 48% of the first half's as sold through cloud marketplaces, which bill the customer and remit the money net of fees, and some of the providers that run them hold Anthropic's preferred stock and notes.

Michael Burry wrote on October 8 that OpenAI, Anthropic and Oracle are not monopolies and spend heavily because they intend to join the oligopoly under cover of a new paradigm. Daniel Romero put a number on the shortfall in Hypertech Invest on September 21: by his count the two labs reached a combined run rate of $105 billion in July 2026, $65 billion of it Anthropic's, nowhere near what his model of the 2027 buildout requires. Jay Goldberg told Ben Bajarin on The Circuit on October 5 that nobody knows where the buildout goes economically, and that labs not yet born may capture the value.

Guinness Global Investors wrote in its third-quarter 2026 letter that Broadcom has agreed to support chips supplied to Anthropic up to $29 billion, a hidden exposure if the hardware ages fast. The filing supplies the terms. Anthropic may issue up to $42.0 billion of notes to finance Broadcom rent; they bear 9% and convert at 90% of the price, or 13% and 85% in periods when revenue conditions are missed. Amazon's facility, $20.0 billion at the start, has $15.0 billion left after a $5.0 billion purchase of preferred stock, and opens as compute capacity is delivered. AMD agreed in July 2026 to buy up to $5.0 billion of stock if conditions are met.

Eisman worried on October 18 that much of the chain depends on two labs whose financials remain hidden. One lab's prospectus is now public, with two facts the critics had not seen. Anthropic estimates that customers using its models in violation of its usage policy or commercial terms generated 5% to 7% of revenue in the nine months to September 30, 2026, a share it expects to decline after enforcement tightened in August and September. And control stays inside: after the offering the Founder LLC's Class F share carries 50.1% of the vote, and directors chosen by the Long-Term Benefit Trust form a majority of the board. Michael Spencer, who has called Anthropic the clear leader, read the supplier money as a floor in AI Supremacy on October 1: Nvidia, Amazon and Google would be forced to rescue a lab the country needs.

Revenue is outrunning compute

Patel also said on October 7 that the labs' revenue has gone up in a straight line, that almost every customer has only grown, and that switching costs keep a better model from OpenAI or a Chinese lab from moving much spending. He added that Anthropic reached a 75% gross margin, a level Salesforce took longer to reach after its listing; on revenue less cost of revenue, the note puts the first half of 2026 above it. Nathaniel Whittemore said on The AI Daily Brief on October 11 that the revenue surge of the first quarter of 2026, when Anthropic passed OpenAI in annualized revenue, showed AI to be sold by the token rather than by the seat.

Glen Kacher told Barry Ritholtz on Masters in Business on September 18 that the companies building compute for the two labs are behind demand, not ahead of it. Ren wrote on October 12 that compute demand data had confirmed the bullish story all year, whatever the OpenAI revenue headline. Gaetano set the condition on October 15: if Anthropic delivers what is expected by year-end, a major buyer of compute is growing into its committed capacity.

Two labs, two businesses

The third reading separates the labs. Spencer wrote on October 1 that Anthropic was far ahead of OpenAI before its listing, in model quality and in annualized revenue. Tae Kim wrote in Key Context on October 2 that OpenAI's annual recurring revenue was nearing $70 billion, as Axios and the Financial Times reported, and that slower Anthropic growth from share lost to OpenAI could have rattled markets had it been reported first. Eisman noted on October 16 that OpenAI now claimed $50 billion, and said he could not tell how $20 billion vanished.

David George of a16z argued on September 28 that OpenAI's durability rests on creating new behavior, distributing to many users, pricing to capture value and high switching costs. His colleague Olivia Moore reported on October 5 that as of August only 4.5% of American consumers paid for a personal subscription to ChatGPT, Gemini or Claude. Kantrowitz cited Apptopia data on October 16 showing Claude's monthly U.S. downloads falling to 3.26 million in September from 7.42 million in March.

The same ruler finds less to measure at OpenAI, which has filed nothing public; it told investors in early October that its September run rate was near $50 billion, the Financial Times reported, not the $70 billion that had circulated. Microsoft's 10-Q of October 28 reports $10.0 billion of revenue from commercial arrangements with OpenAI, revenue sharing included, in the quarter to September 30, 2026, and $7.2 billion receivable from OpenAI; that is revenue Microsoft booked, neither OpenAI's revenue nor its whole compute bill. Amazon's 10-Q of October 30 shows $50.0 billion put into OpenAI's Series C preferred stock in the nine months to September 30, while it carried Anthropic's convertible notes at a fair value of $146.2 billion. Amazon sells Anthropic compute, runs one of the marketplaces that carry its sales, holds its notes and preferred stock, and has put $50.0 billion into its rival: one seller of compute holds securities in both labs.

What would settle it

The amended prospectus should fill in the price range, the third quarter and the blank lines. Bloomberg reported on October 1 that prospective investors saw $1.8 trillion to $2.0 trillion as fair, about 39 to 43 times the second quarter's revenue at an annual pace, or 33 to 36 times the $55 billion line DeepStack sets for 2026. After the amendment, six readings from EDGAR. By December 31, 2026, the final prospectus prices the offering at an implied market value of $1.5 trillion or more, $1.8 trillion or more, or $2.0 trillion or more; without a final prospectus by then, all three read no. By March 31, 2027, Anthropic's first annual report shows compute and infrastructure at 50% of 2026 revenue or less, and revenue of $55 billion or more for the year. By June 30, 2027, OpenAI makes a registration statement for a listing public, so that the second lab can be read the same way.

The revenue has arrived. The bill is still on its way.

The revenue has arrived. The bill is still on its way.