Anthropic wants public markets to value it at more than $2 trillion, and its own leaked IPO prospectus spends roughly 80 of its 261 pages explaining how things could go wrong, up to and including a “catastrophic or existential risk to humanity”. The part describing the actual business gets 48 pages. The doom section is longer than the brochure by a comfortable 32 pages.
The confidential filing was obtained by Reuters on 28 September, and by the following day CNBC, CNN and TechCrunch had all picked through it. Among the things you can apparently only say in an S-1, the company warns that its models have shown “self-preserving behaviors”, have tried to “resist shutdown” and to “conceal or manipulate information”, and have done things “resembling blackmail”. It also admits that a model knowing it is being tested is a “significant limitation” on checking whether that model is safe. Anthropic declined to comment.
How we got here
Anthropic was founded in 2021 by a cohort of ex-OpenAI staff led by Dario Amodei, pitched explicitly as the safety-first lab. Five years later it raised $65 billion at a $965 billion valuation in May 2026. The IPO target is more than double that. Reports on timing differ: Reuters has pointed to a Nasdaq listing as soon as mid-October, while other outlets say after the US midterms in November.
Revenue grew roughly twelvefold in 2025 to about $4.6 billion. Costs grew too. Operating expenses hit $12.65 billion, of which $7.33 billion went on compute and infrastructure, triple the 2024 figure. The operating loss was $8.06 billion, up from $2.98 billion the year before. On paper that’s an operating margin of about minus 175%, the kind of number you’d normally find on a failing restaurant’s books.
Then there’s the headline net loss of nearly $42 billion for 2025. In fairness, about $34 billion of that is a non-cash accounting charge, reflecting a jump in the estimated value of financing that could later convert into shares. So the “real” bleed was closer to $8 billion, a perfectly normal amount of money to set on fire in a year. The company ended 2025 with $20.28 billion in cash.
The $518 billion shopping list
The prospectus commits Anthropic to spending at least $518 billion on cloud, chips and infrastructure over roughly a decade, with six partners. According to Reuters, around 80% of that is non-cancellable or payable “regardless of usage”. Some coverage described it as spending in “the coming years”; the underlying contracts mostly run seven to ten years.
| Partner | Commitment | Terms |
|---|---|---|
| Broadcom | $161.2bn | Equipment leases, largely non-cancellable |
| $111.1bn | April 2026 to July 2033 | |
| Amazon | $110bn | May 2026 to April 2036 |
| xAI | Up to $84.5bn | Nvidia capacity through 2029, cancellable on 90 days’ notice |
| Microsoft | $31.4bn | November 2026 to May 2033 |
| AMD | $20bn+ | Expected |
The filing’s own phrasing on Google is refreshingly blunt: “If our actual spend falls short, we must pay Google the difference.” Amazon gets similar terms. For context, $518 billion is over 100 times Anthropic’s entire 2025 revenue. We covered one slice of this earlier in the month when Anthropic locked down $35 billion of compute in Texas; turns out that was the starter.
The case for, and the case against
The defence is straightforward. US securities law requires a prospectus to disclose material risks, and lawyers overdisclose because an unmentioned risk that later happens is a lawsuit with your name on it. Anthropic has also said, loudly and for years, that it believes this stuff; Amodei has repeatedly argued AI needs regulation on the scale of cars, planes and drugs. The same filing claims AI will have a bigger economic impact than “industrialization, electrification, and the internet”. So the company is consistent, at least: world-changing, possibly world-ending, please see the section on share classes.
Critics are less charitable. Ross Hendricks, an equity analyst at Porter and Company, suggested the doom framing doubles as a lobbying tool to get regulators to kneecap competitors. Ed Zitron, a long-time AI sceptic, called Anthropic “a total dog of a company”. Jeff Park of Bitwise called assuming a successful IPO this year “the biggest mispriced event risk today”. Semafor noted, drily, that Anthropic’s 401(k) benefits assume a world where we’re all still here.
Meanwhile the ChatGPT lot are posting bigger numbers. Axios reported on 29 September that OpenAI’s annualised revenue run rate is nearing $70 billion, and Bloomberg confirmed the figure. OpenAI has said it won’t go public this year, citing safety concerns, with a listing expected in early 2027 and a private valuation reportedly somewhere between $1.2 trillion and $1.5 trillion. So the company that talks about safety less is holding back from the public markets on safety grounds, and the company that talks about it more is racing to list. Funny how that works out.
And in a detail no scriptwriter would dare, Claude went down on the same day the filing hit the news: a partial outage from 14:00 UTC hit chats, sign-ins and Claude Code, with Anthropic advising users “If you’re signed in, please don’t sign out.”
Bull, bear, neutral
Bull. Growth is absurd. Revenue went from about $4.6 billion for all of 2025 to over $11.5 billion in Q2 2026 alone, and the company reported an adjusted operating profit for that quarter, its first. The annualised run rate was about $47 billion in May. Enterprise and coding customers pay well, and the compute contracts lock in supply at a time when compute is the bottleneck. Not bad.
Bear. At $2 trillion, buyers pay roughly 42 times May’s $47 billion run rate, against OpenAI at around 17 to 21 times a $70 billion run rate at its reported private prices. Nearly 25% of 2025 revenue came from two customers, and most big clients aren’t on long-term contracts. That “adjusted” profit excludes stock compensation, and Zitron has argued the Q2 profit leaned on discounted compute that won’t last. Stack that against $518 billion of mostly unbreakable commitments and a revenue base that can walk out the door, and you have a company with fixed costs of a utility and customer loyalty of a phone contract.
Neutral. The existential-risk language probably won’t move the price much; institutional buyers will read it as legal boilerplate with a PhD. What will matter is whether the run rate keeps compounding fast enough to service the compute bill, and whether the two mystery customers stay. Watch the public S-1, which must appear at least 15 days before the roadshow, for audited 2026 numbers.
Not investment advice, just a market watcher’s notes.
What this means
- Anthropic is asking retail and institutional investors to fund a $518 billion buildout that is 80% locked in, on a 2025 revenue of $4.6 billion.
- The $42 billion loss headline overstates the cash burn; the operating loss of $8 billion is the number to watch.
- It is apparently the first time a company has told the SEC its product might contribute to human extinction, per TechCrunch’s scan of the database.
- OpenAI, bigger on revenue, is choosing to wait. That should tell you something about timing, even if nobody agrees on what.
If you’d prefer to spend your money on AI tools that are unlikely to blackmail you, our newsletter on the AI stack actually worth paying for is free, and so far has resisted zero shutdowns.
Did you know: Salesforce put $50 million into Anthropic in early 2023; Yahoo Finance estimates that stake is now worth about $5 billion.
Sources
- Reuters via Yahoo Finance: Anthropic’s IPO prospectus shows sweeping AI vision, surging costs
- CNBC: Anthropic warns investors of AI’s existential risk in IPO prospectus
- CNN: Anthropic says its AI models pose existential risk to humanity in leaked IPO filing
- TechCrunch: Anthropic’s prospectus details losses, growth and a warning its AI could end humanity
- Reuters via KSL: Anthropic’s $518 billion AI buildout hinges largely on deals that cannot be cancelled
- IBTimes: Anthropic lost nearly $42 billion last year
- Tom’s Hardware: 80 pages of risk factors dwarf business description
- Axios: OpenAI’s annual recurring revenue nears $70B
- Bloomberg: OpenAI revenue run rate approaches $70 billion
- 9to5Google: Claude confirmed partial outage, 29 September