Anthropic lost $42 billion last year. About $34 billion of that never left the building.
The headline number from Anthropic's IPO filing is mostly an accounting charge. The numbers underneath it, compute bills and two big customers, tell you more about the company.
Forty-two billion dollars. That’s the net loss Anthropic reported for 2025, according to the IPO prospectus Reuters reviewed and first reported on September 28. On revenue of nearly $4.6 billion, it reads like a company setting fire to nine dollars for every one it takes in.
It isn’t quite that. Reuters, in the same story, says roughly $34 billion of the loss was an accounting charge. The estimated value of financing that could later turn into Anthropic shares went up, because Anthropic itself got more valuable, and that increase gets booked as a loss. No cash went anywhere.
So the headline is technically true and mostly misleading. I’ve seen it quoted bare in a lot of places since Monday, which is why I wanted to pull it apart.
Take the $34 billion out
Strip the charge away and you’re left with a loss of about $8 billion, which lines up with what Reuters calls the operating loss: more than $8 billion, excluding writedowns of liabilities tied to earlier fundraising. Heise, reading the same figures, points out that the operating loss more than doubled from almost $3 billion in 2024.
That’s the number I’d actually watch. Here’s how the pieces Reuters reported fit together:
| 2025 figure (per Reuters) | Amount |
|---|---|
| Revenue | about $4.6 billion |
| Total operating expenses | $12.65 billion |
| Of which compute and infrastructure | $7.33 billion |
| Operating loss | more than $8 billion |
| Net loss | about $42 billion |
| Cash and short-term investments, December 31 | $20.28 billion |
The line that jumps out at me is compute. Anthropic spent $7.33 billion on it, three times its 2024 figure, and that’s more than half of everything it spent all year. Put another way, the compute bill alone was about $1.60 for every dollar of revenue. You don’t have to be a finance person to see that the business, as it stood in 2025, didn’t cover the cost of running its own models.
And that’s before the big one. The filing lists about $518 billion in cloud, computing and infrastructure obligations over the coming years. I couldn’t find a clean year-by-year breakdown in any report I read, so I can’t tell you how front-loaded it is.
Two customers, a quarter of the money
The other number that matters is smaller. Nearly a quarter of 2025 revenue came from two customers, Reuters and the Financial Times both report, and the filing warns that many of its biggest clients aren’t locked into long-term contracts and could cut back or stop.
The prospectus, as reported, doesn’t name them. Heise, citing VentureBeat, says they’re Cursor and GitHub Copilot, the two big coding tools. That fits VentureBeat’s 2025 reporting that those two drove about $1.2 billion of Anthropic’s run-rate at the time. But that’s an inference from last year’s sources, and nobody I read has it from the filing itself. Treat the names as likely, not confirmed.
If they’re right, it’s an awkward pair. GitHub belongs to Microsoft, OpenAI’s biggest backer. Cursor, per heise, was bought by SpaceX this summer, which also owns xAI. Anthropic’s two largest customers would then be owned by companies with their own model businesses.
Why 2025 may already look old
There’s a catch in all of this, and it cuts in Anthropic’s favor. The prospectus covers calendar 2025. The company has grown fast since then. AFP, citing the Wall Street Journal, reports Anthropic brought in $11.6 billion in the second quarter of 2026 alone, against $6.7 billion for OpenAI. That single quarter is more than twice last year’s full revenue.
So an investor reading the 2025 numbers is looking at a company that’s already changed shape. Whether the compute bill has grown even faster than that revenue is the part I can’t answer from public reporting, and I’d want to see it before believing the $2 trillion valuation Reuters says Anthropic is aiming for. That’s more than double the $965 billion it was valued at in May.
The 80 pages
Most of the coverage has gone to the risk section, and it’s long. Reuters counted roughly 80 pages of risk factors in a 261-page main body, against 48 pages describing the business. SpaceX’s prospectus, for comparison, spent about 38 of its 277 pages on risk.
Some of it is standard lawyer caution. Some of it isn’t. The filing warns its models could show “self-preserving behaviors,” including attempts to “resist shutdown,” and says that “potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety.” That’s a company admitting in a legal document that its tests might not catch what its models do. It also says, per Reuters, that the return on its safety spending is unclear, and it didn’t disclose how much that spending is.
The full prospectus hasn't been made public. Every figure here comes from outlets that say they reviewed it, mainly Reuters and the Financial Times. Anthropic declined to comment to Reuters. Numbers could shift when the public filing appears.
Where I land
I think the $42 billion headline gives Anthropic’s critics an easy stick that isn’t really there, and it hides the harder question. The accounting charge will mostly vanish from the conversation once analysts get the real document. The compute ratio won’t. A company whose main cost grew threefold in a year, with a quarter of its sales riding on two customers who can walk, is a real bet. It might be a good one, given that second quarter. But I’d read those numbers before the existential-risk pages, not after.
The listing is expected after the November 3 midterms, per Reuters. The public filing, whenever it lands, is the thing to wait for. It’ll show whether 2026’s revenue has caught up with the compute bill or just chased it.
Related: what Anthropic shipped with Claude Sonnet 5.5, and the White House AI accord Dario Amodei signed the same week.
Sources
- Reuters via U.S. News: Exclusive: Anthropic’s IPO prospectus shows sweeping AI vision, surging costs, September 28, 2026
- Reuters via CNA: Anthropic warns AI may pose ‘existential risks to humanity’ in IPO filing, September 28, 2026
- heise online: Anthropic: Net loss of 42 billion US dollars in 2025 alone, September 29, 2026
- The Verge: Anthropic warns of ‘catastrophic’ AI risks in its own IPO filing
- VentureBeat: Anthropic revenue tied to two customers as AI pricing war threatens margins (2025 report on Cursor and GitHub Copilot)
- AFP via Dawn: Trump says AI bosses have signed ‘morally binding’ accord to self-regulate (Q2 revenue figures citing the Wall Street Journal), September 30, 2026