Anthropic Spent $7.33 Billion on Compute to Earn $4.6 Billion in 2025

Anthropic spent $7.33 billion on computing and infrastructure in 2025, three times its 2024 bill¹. Revenue rose twelvefold, to nearly $4.6 billion¹. Reuters reported both figures on Sept. 28 from a copy of Anthropic's confidential IPO prospectus, a draft the company first submitted to the Securities and Exchange Commission on June 1¹². The filing has yet to reach the SEC's public database, and Anthropic declined to comment³¹. Reuters reported that the listing could value Anthropic at more than $2 trillion, more than double its $965 billion valuation in May¹.

Compute took more than half of $12.65 billion in operating expenses, and the operating loss topped $8 billion¹. Fortune put that loss at $8.06 billion, up from $2.98 billion in 2024, and revenue growth at 1,088 percent⁴. A roughly $34 billion accounting charge made up most of the $42 billion net loss¹. Reuters described that charge as an accounting entry for the rising estimated value of financing that could turn into shares, separate from money spent running the business¹.

Why would a company that lost $8 billion on its operations promise its suppliers at least $518 billion more? Anthropic's prospectus gives the reason in one phrase: demand for advanced AI will be "limited principally by the availability of compute"³.

Where the money went

Usage-based customers supplied about $3.8 billion of 2025 revenue⁵. Subscriptions brought in $789 million. About $2.16 billion, or 47 percent of sales, came through the cloud marketplaces of Amazon and Google⁵. The marketplace share had been 11 percent in 2023 and 32 percent in 2024⁵.

Reuters estimated that Anthropic paid those platforms about $351 million in fees, some 16 cents on each marketplace dollar⁵. Anthropic counts the full marketplace price as revenue and books the platforms' cut as a marketing cost. OpenAI has told investors that method inflates Anthropic's revenue by billions of dollars⁵.

Two unnamed customers each supplied 12 percent of revenue⁵. Many of Anthropic's biggest buyers sit outside long-term contracts and could cut or halt spending, the company warned. Anthropic closed 2025 with $20.28 billion in cash, cash equivalents and short-term investments¹.

Profit, adjusted

Did the losses keep widening in 2026? Preliminary second-quarter revenue topped $11.5 billion, up from $787 million a year earlier, Bloomberg reported on Aug. 15 from documents Anthropic gave prospective investors⁶. The same documents showed positive adjusted operating income for that quarter⁶. Annualized revenue passed $65 billion at the end of July, Bloomberg reported, up from $9 billion at the end of 2025⁷.

Compute cost Anthropic 71 cents for each dollar of revenue in the first quarter, and the company expected 56 cents in the second, The Wall Street Journal reported in May⁸. The Journal's figure for Anthropic's projected operating profit counts model training costs and leaves out stock-based compensation⁸. On Sept. 13 the Financial Times reported gross margins above 80 percent, measured before partner revenue shares and training costs⁹.

Olga Usvyatsky, who writes the Deep Quarry newsletter, raised the accounting question on Sept. 15, two weeks before the leak. "Given everything we have heard about the difficult path to profitability for frontier AI labs, two consecutive quarters of profitability and gross margins above 80% sound remarkable — perhaps almost too good to be true. And on a GAAP basis, it probably is," Usvyatsky wrote⁹. Generally accepted accounting principles, or GAAP, set the rules a public filing must follow, and the adjusted figures have yet to face them.

The $518 billion promise

Reuters reported on Sept. 29 that Anthropic expects to spend at least $518 billion over a decade with six partners³. About 80 percent of that sum binds Anthropic even if usage falls. Google holds the largest cloud commitment, at least $111.1 billion through July 2033³. Amazon follows at $110 billion through April 2036, and Microsoft at $31.4 billion through May 2033. Equipment leases tied to Broadcom add about $161.2 billion³. Separate agreements with Elon Musk's xAI cover up to $84.5 billion of Nvidia-based capacity through 2029, and Anthropic can cancel most of that on 90 days' notice³.

"If our actual spend falls short, we must pay Google the difference," Anthropic said, adding that similar terms apply to its Amazon agreement³. Amazon, Google and Microsoft act as Anthropic's investors, customers, cloud providers, sellers and rivals at once³. The filing warns that their incentives "may not be fully aligned" with its own.

Dario Amodei, Anthropic's chief executive, described the danger of overbuying in February on the Dwarkesh Podcast. The example was a hypothetical $1 trillion compute order starting at the end of 2027. "If my revenue is not $1 trillion dollars, if it’s even $800 billion, there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt if I buy that much compute," Amodei said¹⁰.

Stephen Sopko, practice lead for semiconductor and deep tech at HyperFrame Research, read the same terms from the suppliers' side. "Non-cancelable is only as strong as the balance sheet behind the signature," Sopko told TechTarget¹¹.

The release calendar

Anthropic told investors that new models drive customer usage, and with it revenue¹². In the prospectus, a "continuous and overlapping cadence" of releases is "inherent to remaining at the frontier of AI development"¹². Opus 5.5 shipped the week before the leak¹. The model arrived 10 days after Amodei published a nearly 4,000-word essay calling for pacing the frontier¹².

Its filing also records what Anthropic chose to skip. "Similarly, we have chosen not to develop certain commercially attractive offerings, such as image and video generation models, in order to direct our compute toward our research and safety priorities," Anthropic's filing states¹³.

The case for

Amodei has argued since 2025 that each Claude model pays for itself even while the company loses money. On the Cheeky Pint podcast in August 2025, Amodei walked through an example. "So, let's say in 2023, you train a model that costs $100 million, and then you deploy it in 2024, and it makes $200 million of revenue," Amodei said¹⁴. In the same example, the next model costs $1 billion and the one after that $10 billion, so the company's books show losses of $100 million, then $800 million, then $8 billion. "If you consider each model to be a company, the model that was trained in 2023 was profitable," Amodei said¹⁴. In that reading, each year's loss pays for training the next, larger model.

Michael Field, chief equity strategist at Morningstar in Amsterdam, read the leak as a growth story. "If anything, the acceleration of the business is the biggest part of this release," Field said in comments Reuters collected from analysts¹⁵.

The case against

Ed Zitron, who runs EZPR and writes the Where's Your Ed At newsletter, reached the opposite conclusion from the same 2025 numbers. "It spent $12.65 billion in operating expenses to make $4.6 billion of revenue, otherwise known as spending $2.75 to make a dollar," Zitron wrote on Sept. 29¹⁶. "I don’t see how this company becomes one that can afford its $518 billion in compute commitments, nor do I see how it magically works its way out of the economic equivalent of septic tank," Zitron added.

Scott Bickley, an advisory fellow at Info-Tech Research Group, traced the problem to a pattern that repeats. "The problem is that every time Anthropic’s revenues beat expectations, their costs beat them by more. This is a structural feature of a frontier growth model," Bickley told CIO¹⁷. Zitron's ratio counts every operating dollar of 2025, including $351 million in marketplace fees⁵. The positive 2026 figures leave out stock-based pay, so the two sides are measuring different things⁸.

What builders can take from it

Four patterns in Anthropic's numbers apply to any company that builds on a frontier model.

A supplier sets the biggest cost. Gennaro Cuofano of FourWeekMBA stated the rule plainly. "When the largest line in the cost base is purchased capacity rather than people, gross margin becomes a function of somebody else’s pricing," Cuofano wrote¹⁸. Anthropic hands Amazon and Google about 16 cents of each marketplace dollar⁵. Cursor, the AI coding tool, pays retail for model access that Anthropic gets wholesale, Fortune reported in March¹⁹.

Flat prices lose money on agents. Amol Avasare, Anthropic's head of growth, explained the company's usage limits on X in April. "Engagement per subscriber is way up. We've made small adjustments along the way (weekly caps, tighter limits at peak), but usage has changed a lot and our current plans weren't built for this," Avasare wrote²⁰. Copilot usage would draw on AI credits from June 1, GitHub announced in April²¹. "GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable," wrote Mario Rodriguez, its chief product officer²¹.

Cheaper tokens, costlier tasks. Guido Appenzeller of Andreessen Horowitz found in 2024 that the cost of a given level of model skill falls tenfold a year²². A paper by Hans Gundlach, Neil Thompson and two co-authors found a different trend at the frontier²³. The price of running the newest models is rising 3 to 18 times a year, as models grow and reasoning runs longer²³. Heavier use is cancelling out cheaper tokens and keeping the cost of each task high, Bain & Company reported on Sept. 29²⁴.

Big customers carry leverage. Two customers supplied nearly a quarter of Anthropic's 2025 revenue⁵. Ara Kharazian, lead economist at Ramp, found the same pattern across the industry. "80% of OpenAI and Anthropic's enterprise revenues come from 1% of their customers, and it's not getting better," Kharazian wrote on X²⁵. Frank Dickson, principal analyst at Dickson Research, drew the lesson for buyers. "CIOs have more leverage than they think, because the vendor needs them more than the pitch deck admits," Dickson told CIO¹⁷.

What each side needs

Amodei's case needs each new model to earn back its training cost even as the next one costs more, and it needs the 2026 adjusted profits to stay positive under generally accepted accounting principles¹⁴⁸. Growth on that scale also depends on the two customers worth 12 percent of revenue each⁵. Amazon and Google, the channels behind 47 percent of sales, have to keep selling Claude as well⁵. Zitron's case needs growth to slow while the bills keep arriving, with Google's commitment running to 2033 and Amazon's to 2036³.

What to watch

• The public S-1 on the SEC's database, with the full financial statements behind the leaked figures².
• Gross margin after marketplace fees and training costs, the two items the FT's 80 percent figure left out⁹.
• Names of the two customers that each supplied 12 percent of 2025 revenue⁵.
• Compute cost per revenue dollar, which the Journal put at 71 cents in the first quarter⁸.
• Release dates for Anthropic's next models, which its prospectus ties to revenue¹².

Sources

1. Echo Wang, "Anthropic's IPO prospectus shows sweeping AI vision, surging costs," Reuters (via Yahoo Finance), Sept. 28, 2026, https://finance.yahoo.com/technology/ai/articles/exclusive-anthropics-ipo-prospectus-shows-231722972.html
2. Anthropic, "Anthropic confidentially submits draft S-1 to the SEC," Anthropic, June 1, 2026, https://www.anthropic.com/news/confidential-draft-s1-sec
3. Echo Wang and Sabrina Valle, "Anthropic's $518 billion AI buildout hinges largely on deals that cannot be canceled, filing shows," Reuters (via KSL.com), Sept. 29, 2026, https://www.ksl.com/article/51629885/anthropics-518-billion-ai-buildout-hinges-largely-on-deals-that-cannot-be-canceled-filing-shows
4. Beatrice Nolan, "Anthropic’s IPO filing details steep losses, rapid growth, and a fear that AI could end humanity," Fortune, Sept. 29, 2026, https://fortune.com/2026/09/29/anthropic-leaked-ipo-prospectus-losses-growth-ai-end-humanity/
5. Echo Wang and Krystal Hu, "Exclusive-Anthropic IPO prospectus lays bare deep dependence on Big Tech partners," Reuters (via Yahoo Finance), Sept. 29, 2026, https://ca.finance.yahoo.com/news/exclusive-anthropic-ipo-prospectus-lays-205226030.html
6. Bailey Lipschultz and Rachel Metz, "Anthropic revenue surges to over $11.5 billion in second quarter," Bloomberg (via Fortune), Aug. 15, 2026, https://fortune.com/2026/08/15/anthropic-revenue-q2-11-5-billion-ipo-investors/
7. Marina Temkin, "Anthropic's annualized revenue surges to $65B," TechCrunch, Aug. 17, 2026, https://techcrunch.com/2026/08/17/anthropics-annualized-revenue-surges-to-65b/
8. PYMNTS, "Anthropic On Track for First Operating Profit as Revenue Surges," PYMNTS, May 21, 2026, https://www.pymnts.com/artificial-intelligence-2/2026/anthropic-on-track-for-first-operating-profit-as-revenue-surges/
9. Olga Usvyatsky, "Anthropic’s Pre-IPO Profitability Story," Deep Quarry, Sept. 15, 2026, https://deepquarry.substack.com/p/anthropics-pre-ipo-profitability
10. Dwarkesh Patel, "Dario Amodei — 'We are near the end of the exponential'," Dwarkesh Podcast, Feb. 13, 2026, https://www.dwarkesh.com/p/dario-amodei-2
11. Shane Snider, "Report: Anthropic locks in $518B AI infrastructure commitments," TechTarget, Sept. 29, 2026, https://www.techtarget.com/it-infrastructure/news/366651336/Report-Anthropic-locks-in-518B-AI-infrastructure-commitments
12. Echo Wang and Aditya Soni, "Exclusive-Anthropic warns AI may pose 'existential risks to humanity' in IPO filing," Reuters (via Yahoo Finance), Sept. 29, 2026, https://ca.finance.yahoo.com/news/exclusive-anthropic-warns-ai-may-002217493.html
13. Echo Wang, Ross Kerber and Jeffrey Dastin, "Exclusive-Anthropic leaders to control AI lab via 'Founder LLC' to promote public good over market forces," Reuters (via The Star), Sept. 28, 2026, https://www.thestar.com.my/tech/tech-news/2026/09/29/exclusive-anthropic-leaders-to-control-ai-lab-via-039founder-llc039-to-promote-public-good-over-market-forces
14. Cheeky Pint, "A Cheeky Pint with Anthropic CEO Dario Amodei," Cheeky Pint, Aug. 6, 2025, https://cheekypint.substack.com/p/a-cheeky-pint-with-anthropic-ceo
15. Reuters, "Anthropic's IPO prospectus sharpens focus on AI valuations," Reuters (via CNA), Sept. 29, 2026, https://www.channelnewsasia.com/business/anthropics-ipo-prospectus-sharpens-focus-ai-valuations-6418941
16. Ed Zitron, "Dead Money," Where's Your Ed At, Sept. 29, 2026, https://www.wheresyoured.at/dead-money/
17. Evan Schuman, "Anthropic revelations suggest a much stronger AI negotiating stance for enterprise CIOs," CIO, Sept. 29, 2026, https://www.cio.com/article/4228409/anthropic-revelations-suggest-a-much-stronger-ai-negotiating-stance-for-enterprise-cios.html
18. Gennaro Cuofano, "Anthropic's Obligations Dwarf Its Cash And Its Revenue," FourWeekMBA, Sept. 29, 2026, https://fourweekmba.com/ai-anthropic-ipo-filing-reuters-reported-numbers-ai-cost-struct/
19. Allie Garfinkle, "Cursor’s crossroads: The rapid rise, and very uncertain future, of a $30 billion AI startup," Fortune, March 21, 2026, https://fortune.com/2026/03/21/cursor-ceo-michael-truell-ai-coding-claude-anthropic-venture-capital/
20. Amol Avasare (@TheAmolAvasare), "Engagement per subscriber is way up," X, April 21, 2026, https://x.com/TheAmolAvasare/status/2046725282502173046
21. Mario Rodriguez, "GitHub Copilot is moving to usage-based billing," GitHub Blog, April 27, 2026, https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/
22. Guido Appenzeller, "Welcome to LLMflation - LLM inference cost is going down fast ⬇️," Andreessen Horowitz, Nov. 12, 2024, https://a16z.com/llmflation-llm-inference-cost/
23. Hans Gundlach, Jayson Lynch, Matthias Mertens and Neil Thompson, "The Price of Progress: Price Performance and the Future of AI," arXiv, March 23, 2026, https://arxiv.org/abs/2511.23455
24. Jue Wang, Anne Hoecker, Alec Koh, Chris McLaughlin and Vito Trinchera, "Managing Token Spending without Choking Off Opportunity," Bain & Company, Sept. 29, 2026, https://www.bain.com/insights/managing-token-spending-without-choking-off-opportunity-technology-report-2026/
25. Ara Kharazian (@arakharazian), "New from Ramp data: the latest threat to the AI trade," X, Sept. 2, 2026, https://x.com/arakharazian/status/2095204452609171555