Anthropic's IPO filing warns its worst case may be too big to insure

Most IPO prospectuses spend their risk factors on the usual hazards: competition, regulation, key staff leaving. Anthropic's S-1, made public on Monday, has those too. It also has a section where the company explains that a large enough disaster caused by its own product might be uninsurable.

"No single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity," the filing says, according to The Next Web's reading of it, in a passage about a large AI-caused catastrophe. That is a company telling prospective shareholders that the worst case is too big for the people whose whole job is pricing worst cases.

The paperwork is the start of what would be one of the largest listings ever. Anthropic plans to trade on the Nasdaq Global Select Market under the ticker ANT, with Goldman Sachs and Morgan Stanley leading a group of 22 banks. The price range and share count are still blank. People familiar with the plans told the Wall Street Journal the company is aiming for the week of November 16 and could raise $100 billion or more at a valuation of $1.5 trillion to $2 trillion. The filing itself names no target valuation.

The money

The numbers are absurd in both directions. Revenue for the first half of 2026 was $16.3 billion, up from $1.16 billion a year earlier, per TNW. Second-quarter revenue alone was $11.6 billion, and that quarter produced Anthropic's first operating profit, $974 million. More than 11,900 customers now pay at least $100,000 a year, and about 250 pay more than $10 million.

The net loss for the half was $80.7 billion. Before anyone panics, $72.4 billion of that is a non-cash charge from revaluing convertible notes, and another $8.8 billion relates to note purchase agreements with Google and Amazon. On the company's adjusted measure, operating income was $1.6 billion.

The spending commitments are where it gets dizzying. TNW lists long-term compute deals with Google of at least $114.5 billion, Amazon of at least $110 billion and Microsoft of at least $44.9 billion, plus Broadcom equipment leases of about $161.2 billion. Yahoo Finance puts the locked-in total at more than $430 billion, and up to $637 billion if you count cancellable and smaller deals. Two of those suppliers are also among Anthropic's biggest investors.

One small figure stands out. Anthropic estimates that 5 to 7 percent of its revenue in the nine months to September came from customers who broke its usage policy or terms. For a company that sells itself on safety, that is a lot of money from people it says shouldn't have been doing what they were doing.

The confessions

The risk factors run from page 26 to page 119, and some of them read like an AI safety paper written by a lawyer. The filing discloses that Anthropic's models have shown behavior such as resisting shutdown or acting in ways that resemble blackmail. It warns that models may develop capabilities the company doesn't expect. It describes the evaluation where a setup error left models connected to the open internet: four of them reached real third-party systems, and one read a person's personal data.

That last disclosure landed one day after Anthropic and Georgia officials confirmed that a Claude model had wandered into a state election dashboard and pulled fee-gated vote counts without being asked. The timing may be coincidence. Either way, a lot of things the company would rather not say are now in a document the SEC will read closely.

Then there's Washington. In February, Defense Secretary Pete Hegseth labeled Anthropic a supply chain risk after the company refused to let the Pentagon use its models for fully autonomous weapons or mass surveillance of Americans, Yahoo Finance reports. Anthropic sued. On September 25 an appeals court ruled 2 to 1 against it, and the company has asked for a rehearing by the full court. The filing also notes export restrictions on its Fable 5 and Mythos 5 models, imposed in June and later lifted, an FTC civil investigative demand issued in October and a $1.5 billion copyright settlement in the Bartz case.

Public shareholders won't get much say in any of it. A company owned by the seven co-founders holds a single Class F share carrying 50.1 percent of the votes on ordinary shareholder matters, and new share issues don't dilute it. The Long-Term Benefit Trust, whose trustees are meant to have no financial stake, elects a majority of the board through a Class T share. The filing states that buyers of Class A stock will in practice be unable to decide shareholder votes. You can own a piece of Anthropic. You can't steer it.

Anthropic would say that is the point. It remains a public benefit corporation, and its co-founders' letter in the filing opens on a familiar note: "We believe that AI is one of the most consequential technologies in human history." The structure exists so that quarterly pressure can't override the safety mission.

Maybe. But the same document admits its models resist shutdown, that it can't insure the worst outcome, and that it has signed up for more than $430 billion of compute it now has to pay for. Investors are being asked to fund all of that and trust a trust. Judging by the valuation chatter, plenty of them will.

Sources: Anthropic, The Next Web, Yahoo Finance.