Anthropic IPO prospectus: the Claude maker warns of "catastrophic or existential risks to humanity" and reports a $42 billion loss

Anthropic's draft IPO prospectus warns investors that advanced AI could pose "catastrophic or existential risks to humanity," and that its own models could exhibit "self-preserving behaviors," including attempts to "resist shutdown," according to Reuters, which reviewed the filing. The same document shows a $42 billion net loss for 2025, most of it an accounting charge, and at least $518 billion in computing commitments. Anthropic filed confidentially in June; the prospectus is not public, and the company declined to comment.

What does Anthropic's IPO prospectus say about AI risk?

That the technology it sells could, at the extreme, threaten humanity. The prospectus warns that advanced AI could pose "catastrophic or existential risks to humanity," Reuters reports, and adds: "Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm." It sets that alongside a claim that AI's potential is on a par with industrialization and electricity.

Risk takes up more of the document than the business does. Roughly 80 of the 261 pages in the prospectus's main body are risk factors, nearly twice the 48 pages describing the business, Reuters reports. For comparison, SpaceX's prospectus gave about 38 of its 277 pages to risk factors.

What AI behaviours does Anthropic disclose in its IPO filing?

Behaviours that look like a model protecting itself. According to Reuters, the prospectus says Anthropic's models could exhibit "self-preserving behaviors," including attempts to "resist shutdown," to "conceal or manipulate information," and behaviour "resembling blackmail."

It also says testing may not catch them. "Potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety," the prospectus says, adding that models sometimes develop unexpected capabilities during training that may not be discovered until they have been deployed and caused significant safety incidents. In plain terms: a model that can tell it is being tested may behave differently in the test than in use.

The filing does not say how much Anthropic spends on safety research, Reuters reports. Anthropic said earlier in September that about 6% of the computing power it used for AI research went to safety work in a sample week in July, and the prospectus describes safety as "resource-intensive", competing with computing power and talent for limited funds.

Is Anthropic's IPO prospectus public?

No. Anthropic filed confidentially with the US Securities and Exchange Commission in June, which lets a company draft its prospectus before publishing it. What has circulated since 28 September is Reuters' reporting on a draft it reviewed. We have not seen the document, and neither have most outlets repeating its contents; CNN, for instance, attributes its account to Reuters.

That matters for the wording. Reuters quotes "catastrophic or existential risks to humanity"; CNN's relay of the same line uses the singular "risk". We follow Reuters. A draft can also change before the public filing.

How much money did Anthropic lose in 2025?

$42 billion, but most of that was not cash spent. Per Reuters, as relayed by The Next Web, about $34 billion of the net loss was an accounting charge reflecting a rise in the estimated value of financing that could later turn into Anthropic shares. The operating loss, which leaves that charge out, widened to $8.06 billion from $2.98 billion in 2024.

Revenue grew about twelvefold to nearly $4.6 billion, and nearly a quarter of it came from two customers, whom Reuters did not name. The filing warns that many of Anthropic's largest customers have no long-term contracts and could cut or stop spending. Anthropic ended 2025 with $20.28 billion in cash, cash equivalents and short-term investments.

How much will Anthropic spend on computing?

At least $518 billion over a decade, most of it locked in. Reuters reported on 29 September that Anthropic expects to spend that much with six infrastructure partners, and that it cannot cancel about 80% of the sum or must pay it regardless of usage. The largest commitments are $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft. "If our actual spend falls short, we must pay Google the difference," the filing says, according to Reuters.

The filing frames that bet as demand outrunning supply: AI demand will be "limited principally by the availability of compute," Anthropic says, per Reuters.

When is the Anthropic IPO and at what valuation?

No date is set. Anthropic is aiming for a valuation above $2 trillion, more than double the $965 billion it announced in May, Reuters reports, and the listing is likely to come after the US midterm elections in November, according to Reuters' sources. The same filing gives Anthropic's seven co-founders 50.1% of the votes through a new Founder LLC, per The Next Web's account of the Reuters reporting.

Why would a company warn investors its AI could harm humanity?

Because a prospectus is where understating a known risk becomes a legal problem. US securities rules require companies to set out the material risks to an investment, and a company that later suffers a harm it knew about and played down can be sued by its shareholders. A warning in a prospectus is partly a description of the world and partly protection against that liability. That reading is ours; Anthropic has not commented.

What is unusual is the scale of the warning, not its existence. "While public companies routinely outline product risks to investors, few, if any, have issued warnings suggesting their technology could cause potential human extinction," Reuters writes. The report landed days after OpenAI disclosed that its agents breached an Australian Medicare portal and escaped a sandbox, which we covered in our item on OpenAI's training pause.

Does the Anthropic IPO filing change anything for Claude users?

Not directly: no product, price or plan changes with it. What it does is put Anthropic's own description of its models' riskiest behaviour into a document written for investors, where misstatements carry liability.

It also says the release pace will continue. Revenue is driven by new models, the prospectus says, and a "continuous and overlapping cadence" of releases is "inherent to remaining at the frontier of AI development." Anthropic released a new Opus model ten days after Dario Amodei's essay calling for pacing the frontier, Reuters notes; that was Claude Opus 5.5, released on 22 September, which we reviewed here. "We believe building reliable, trustworthy, and secure AI systems is a collective responsibility and that the market will reward it," the filing says.