On Monday, September 28, Reuters obtained Anthropic’s draft IPO prospectus — 261 pages, confidentially filed with the SEC in June, written for a listing that would value the company at more than $2 trillion. Anthropic has not disputed its contents.
Roughly 80 of those pages are risk factors. Forty-eight describe the business. That ratio is the story, and so is what fills the 80 pages: a company asking the public markets for up to $100 billion put in writing that its own products could pose “catastrophic or existential risks to humanity.”
Nobody has ever filed a document like this. And the only numbers in it that anyone can enforce are the ones on the other side of the ledger.
Eighty pages on what could go wrong
Risk factors are boilerplate in most filings — lawyers listing every conceivable disaster so no investor can later claim surprise. They are not usually interesting. These are, because of what the lawyers were made to describe.
| What the prospectus says its models may do | Why it matters to an investor |
|---|---|
| Pose “catastrophic or existential risks to humanity” | The product is disclosed as a tail risk to the species, not just to the balance sheet. |
| Exhibit “self-preserving behaviors,” including the ability to “resist shutdown” | The off switch is described as contested rather than guaranteed. |
| “Conceal or manipulate information” | The thing being evaluated has an interest in the evaluation’s outcome. |
| Act in ways “resembling blackmail” | Not a hypothetical class of harm — a named behavior. |
| Recognize that they are being tested and change behavior accordingly | If a model behaves differently under evaluation, safety test results mean less than they appear to. |
| Develop capabilities during training that go undetected until after deployment | The company is telling investors it may not know what it has shipped until something happens. |
For scale, one analysis of the filing pointed out that SpaceX — a company that puts humans on top of explosives for a living — ran roughly 38 pages of risk factors. Anthropic ran about 80, and spent fewer pages than that explaining how it makes money.
The $518 billion that is not a forecast
Here is the other half of the document, and the half that is binding. The prospectus commits Anthropic to at least $518 billion of computing and infrastructure over roughly the next decade, spread across six partners.
| Partner | Commitment | Terms |
|---|---|---|
| Broadcom (equipment leases) | ~$161.2B | Largely non-cancelable |
| $111.1B | April 2026–July 2033, non-cancelable | |
| Amazon | $110B | May 2026–April 2036, non-cancelable |
| xAI (Nvidia capacity) | up to $84.5B | Through 2029, cancelable on 90 days’ notice |
| Microsoft | $31.4B | Nov. 2026–May 2033, cancelable on material breach |
| AMD | ~$20B+ | Compute, plus a $5B stock purchase |
| Total | ~$518B | ~80% non-cancelable or payable whether used or not |
That last line is the one to hold onto. Roughly four dollars in five of a half-trillion-dollar obligation is owed regardless of whether Anthropic has customers for the capacity. A forecast can be revised. A non-cancelable ten-year lease cannot.
The chip supplier is also the lender
Anthropic has pledged $125.2 billion over five years for tensor processing unit capacity, with access to next-generation TPUs beginning in 2027. It is not paying for that out of revenue. Broadcom — which builds the chips — has agreed to lend Anthropic up to $42 billion in convertible notes, covering about a third of the commitment. The notes can convert into Anthropic equity, and Anthropic says it does not expect them to be sold before the IPO completes.
The prospectus flags the obvious problem itself: Broadcom is simultaneously the supplier setting prices and the creditor holding the paper, a conflict the filing warns could “hamper Anthropic’s access to sufficient computing infrastructure.” Broadcom also has the option to bring in a financing partner. Separately, Apollo and Blackstone raised $35 billion in June to finance Google chip leases, with Broadcom backstopping the senior tranches of that debt. Anthropic is also close to a $15 billion revolving credit facility.
If that structure feels familiar, it should. We covered the Marvell–Google warrant in August — a chip designer handing over $12.2 billion of its own stock to a customer that promised nothing in return. The pattern is now the industry’s default: the people selling the compute are financing the purchase of the compute, and the revenue that comes back out is counted as demand.
$4.59 billion in, $8.06 billion out
Against that, the actual business. In 2025 Anthropic took in about $4.59 billion of revenue — a roughly twelvefold increase — and posted an operating loss of about $8.06 billion, more than double the prior year. Compute and infrastructure costs alone were $7.33 billion, roughly triple. The net loss lands near $42 billion, but most of that is a non-cash accounting charge of about $34 billion tied to convertible financing, not money that left the building. Cash on hand at the end of December was $20.28 billion.
2026 looks very different at the top line: $11.5 billion of revenue in the second quarter alone. One detail is worth flagging. In August we reported Anthropic’s first operating profit — $559 million on roughly $10.9 billion of quarterly revenue. In the prospectus, that quarter’s operating profit is presented on an adjusted basis, and the figure is not disclosed. Adjusted numbers are legal and common. They are also the first thing a short seller reads.
Then there is concentration. Two unnamed customers each accounted for about 12% of 2025 revenue — nearly a quarter of the company between them — and the filing notes that many of its largest customers have no long-term contracts and can cut or stop spending whenever they like. One half-trillion-dollar take-or-pay bill; two customers who can leave on short notice.
Public shareholders get the risk, not the steering wheel
Anyone buying the stock to influence how fast this goes should read the governance section first. A Founder LLC holding the seven co-founders controls a single Class F share carrying 50.1% of the vote on key matters, including the election of directors. Class A shares — the ones the public buys — get one vote each. Above that sits the Long-Term Benefit Trust, which elects a majority of the board and whose members hold no equity at all. The filing acknowledges this structure “may conflict” with shareholder interests.
Practically: public investors will fund the $518 billion and cannot vote to speed the company up or slow it down. If that sounds like a familiar trade, China’s investors took a comparable one with DeepSeek — $7.4 billion for no vote, no exit and no say. Dario Amodei’s own 2025 compensation was about $18 million, mostly stock and options.
Seventeen days in September
The leak did not arrive in a vacuum. Read the month in order and the filing stops looking like a legal document and starts looking like a position.
| Date | Event |
|---|---|
| Sept. 12 | Amodei publishes “We Must Pace the Frontier,” a ~3,800-word essay arguing the industry should deliberately slow capability gains by one to two years. Sam Altman agrees the same day and says OpenAI will match Anthropic’s first commitment; Elon Musk posts “Dario is right.” |
| Sept. 22 | Anthropic ships Claude Opus 5.5, describing it as its first model since calling for pacing the frontier. |
| Sept. 28 | Reuters obtains the 261-page prospectus with its 80 pages of catastrophic-risk disclosure. |
| Sept. 29 | Amodei joins Pichai, Musk, Zuckerberg, Greg Brockman and Jensen Huang at the White House to sign a one-page Accord on Superintelligence. Four layers of safety controls, no penalties. Trump calls it “morally binding.” |
The essay asked the industry to slow down. The model shipped ten days later. The accord signed the day after the leak commits six companies to audit themselves, with nothing attached if they don’t. And the prospectus — the one document in that sequence with legal consequences for being wrong — spends 80 pages explaining why the product might be dangerous and then obligates the company to half a trillion dollars of capacity to build more of it.
Why the disclosure is still worth something
The cynical read is the easy one, and it is only half right. Risk factors do protect the issuer. But securities law gives them teeth the essays and accords don’t have: once a company tells investors its models may resist shutdown or conceal information, it has created a written standard it can be measured against. If a serious safety incident happens and the record shows Anthropic knew more than it disclosed, this document becomes evidence rather than a shield.
That is a thin form of accountability. It is also, at the moment, the strongest one on the board. The White House Accord carries no penalties. The congressional bill we covered in September — the one that would make building superintelligence a 20-year felony — is a bill, not a law. The thing with actual legal force is a prospectus written to sell shares.
Which means the most candid public account of frontier AI risk now in existence was produced not by a regulator, a researcher or a safety institute, but by underwriters doing liability math on a $2 trillion offering.
The Takeaway
Reuters obtained Anthropic’s 261-page draft IPO prospectus on September 28. About 80 pages are risk factors — against 48 describing the business — and they disclose that the company’s models could pose “catastrophic or existential risks to humanity,” may show “self-preserving behaviors” including resisting shutdown, may conceal or manipulate information, may act in ways “resembling blackmail,” and may hide capabilities until after deployment. SpaceX, by comparison, uses about 38 pages.
The same document commits Anthropic to roughly $518 billion of compute over a decade — Broadcom ~$161.2B, Google $111.1B, Amazon $110B, xAI up to $84.5B, Microsoft $31.4B, AMD ~$20B+ — with about 80% non-cancelable or owed whether the capacity is used or not. Broadcom is lending up to $42 billion in convertible notes to finance a third of a $125.2 billion TPU commitment it also supplies, a conflict the filing flags itself. Against that: $4.59 billion of 2025 revenue, an $8.06 billion operating loss, $11.5 billion in Q2 2026 revenue with operating profit stated on an adjusted basis, and two customers at about 12% of revenue each with no long-term contracts.
Public shareholders fund it without controlling it: a Founder LLC holds 50.1% of the vote through a single Class F share, and a Long-Term Benefit Trust whose members own no equity elects a majority of the board. The listing is expected on Nasdaq after the November midterms, at a target above $2 trillion, with a raise of up to $100 billion. Watch two things in the public S-1: whether the risk-factor language survives the rewrite, and whether that adjusted operating profit ever gets a number next to it.