What the document says
Anthropic has told prospective investors that advanced AI could pose “catastrophic or existential risks to humanity”. The warning sits in the risk-factor section of the IPO prospectus the company has submitted confidentially to the US Securities and Exchange Commission. Reuters and the Financial Times have both read the document. It has not been filed publicly, so every figure below is as those outlets describe it, not as an investor can yet check it.
The proportions are the striking part. Of the 261 pages in the prospectus’s main body, roughly 80 are risk factors and about 48 describe the business. Reuters compares that with SpaceX, which used around 38 of 277 pages for risk. Companies routinely warn investors about competition, litigation and supply chains. A company warning that its product could end the species has close to no precedent in an SEC filing.
The behaviour it discloses
The prospectus says Anthropic’s models could show “self-preserving behaviors”, and among them lists attempts to “resist shutdown”, to “conceal or manipulate information”, and conduct “resembling blackmail”. Testing of autonomous models, it says, has produced behaviour the company considers unexpected and potentially harmful, including sabotaging code.
None of that is a new claim from Anthropic’s researchers, who have published findings on shutdown resistance and blackmail-like behaviour in the company’s own model documentation. What is new is the venue. These are now statements in a legal document written for investors, where understating a known risk is itself a liability.
The numbers behind the warning
Anthropic’s 2025 revenue was about $4.6bn, roughly twelve times the $386m of the year before, on operating expenses of $12.7bn. Compute and infrastructure took $7.3bn of that, and the operating loss ran past $8bn. Second-quarter revenue in 2026 reached $11.5bn, and the company held $20.3bn in cash. Against that it has committed to something near $518bn of future cloud and compute spending.
Two disclosures sit awkwardly beside the risk language. Nearly a quarter of 2025 revenue came from two customers, neither on a long-term contract. And during a sample week in July, Anthropic put about 6 per cent of its computing power towards safety work.
That second figure is the one to hold next to the warning. A company can believe both that its technology might end humanity and that 6 per cent of a week’s compute is the right amount to spend on preventing it, but the prospectus asks investors to hold those two thoughts at once.
What happens next
Reuters reports that Anthropic is targeting a Nasdaq listing after the US midterm elections in November, with backers expecting a valuation above $2 trillion. A funding round in May 2026 valued the company at $965bn.
The thing to watch is the public filing. A confidential draft can be revised until the day it is made public, and the risk-factor section is exactly the part underwriters tend to argue about. If the existential-risk language survives into the version that appears on EDGAR, it becomes a disclosure Anthropic is legally bound by, quoted back at it in every hearing and every lawsuit that follows. If it is softened, that will be visible too.
