Artificial intelligence
Anthropic's Leaked Prospectus Commits $518 Billion to Compute Against $4.6 Billion of 2025 Revenue. Most of It Cannot Be Cancelled
Reuters and the Financial Times have reported from a leaked June draft of Anthropic's IPO prospectus: $4.6 billion of 2025 revenue, an $8.06 billion operating loss that is almost entirely compute, and $518 billion of cloud commitments that are roughly 80% non-cancellable. The headline loss is not the disclosure that matters.
MAI
Reuters obtained Anthropic's confidential IPO prospectus and reported its contents on September 28; the Financial Times reported from the same document. Anthropic has not filed publicly and has not confirmed any of it, and the draft is dated June. Everything below is a leak of a working document, and should be read as one.
It is still the first look anyone has had at the income statement of the second-largest company in the industry, and the number that deserves attention is not the loss.
What the document shows
| 2024 | 2025 | |
|---|---|---|
| Revenue | about $400 million | about $4.6 billion |
| Operating loss | $2.98 billion | $8.06 billion |
| Compute and infrastructure spend | — | $7.33 billion, triple 2024 |
| Net loss | — | $42 billion |
| Cash at 31 December | — | $20.28 billion |
Revenue grew twelvefold. The $42 billion net loss is largely an artefact: roughly $34 billion of it is non-cash charges tied to financing instruments, the sort of remeasurement that stops recurring once a company lists. The operating loss of $8.06 billion is the real figure, and $7.33 billion of it — better than nine tenths — is compute. Anthropic is not losing money on salaries or sales. It is losing money renting machines.
Reuters reports that 2026 has gone differently: revenue of $11.5 billion in the second quarter against $4.73 billion in the first, and a second consecutive quarter of operating profit on an adjusted basis. On those numbers the growth argument writes itself.
The commitment on the other side of the page
The prospectus puts Anthropic's cloud and infrastructure commitments at $518 billion. Around 80% of that is described as non-cancellable, or payable whether or not the capacity is used. Roughly $161.2 billion sits in Broadcom equipment leases, and nearly half of the cloud portion is split across Amazon, Microsoft and Google. Published accounts of the document differ on the period the total is spread over.
Set $518 billion against $4.6 billion of 2025 revenue, or against an annualised 2026 run rate in the forties. This is the actual disclosure in the filing, and it is a different kind of number from a loss. A take-or-pay obligation is not a capital plan that can be slowed if demand softens; it is a fixed cost that arrives on schedule. Every AI company's story rests on the assumption that inference demand keeps compounding. Anthropic's has been written into contracts.
Two customers
Nearly a quarter of 2025 revenue came from two customers, and Reuters reports that most major clients do not hold long-term contracts — the filing's own warning is that they "could cut or stop spending."
That is the sentence to read next to the $518 billion. On one side, obligations that are mostly irrevocable and dated. On the other, a revenue base that is concentrated and, contractually, month to month. The gap between those two durations is the real thing an underwriter has to price, and it is not the headline loss.
A safety argument inside a securities filing
The risk factors run past 80 pages — close to a third of the document. They include what the Financial Times characterised as existential risks to humanity, and warn that increasingly advanced models could manipulate, blackmail, or otherwise behave in unpredictable ways.
None of that is new from Anthropic, which has published versions of the same claim for years. The venue is what changes the meaning. A risk factor is a legal instrument: its function is to ensure that when the described thing happens, investors cannot say they were not warned. Anthropic has now placed its safety thesis in a document whose purpose is to move that risk onto whoever buys the shares. Either the company means it, in which case it is inviting public markets to underwrite a technology it describes as potentially catastrophic, or it does not, in which case this is the most expensive piece of positioning in the sector. The first reading is the more plausible one, and it is the less comfortable.
What the leak does not settle
The listing is expected after the November US midterm elections, at a valuation above $2 trillion, with reports of an offering as large as $100 billion and Nvidia — which invested $10 billion in November 2025 — named as a possible anchor investor. OpenAI filed confidentially in June and is expected to list by early 2027. None of this is fixed. A June draft is not a prospectus, the numbers in it have moved, and the version that reaches the SEC will have been rewritten by lawyers who now know this one leaked.
What the leak does establish is the shape of the trade. Anthropic is asking the public market to fund a wager that inference demand grows faster than an obligation that is already locked in. The revenue line is the argument that it will. The customer concentration is the argument that it might not.
Sources: Reuters: Exclusive — Anthropic's IPO prospectus shows sweeping AI vision, surging costs · Fortune: Anthropic's $2 trillion IPO S-1 prospectus has leaked—and it shows steep losses · Fortune: Anthropic's leaked IPO prospectus details steep losses, rapid growth, and a fear that AI could end humanity · SiliconANGLE: Leaked Anthropic IPO filing reveals $8B operating loss, rapid revenue growth · The Decoder: Anthropic's IPO filing shows soaring revenue, mounting costs, and "existential" risks · Euronews: Anthropic IPO filing warns AI may pose 'existential risks to humanity'