The IPO of the apocalypse

Somewhere in a data room last week, a lawyer read the sentence “our models could pose a catastrophic or existential risk to humanity,” approved it, and put it in a document whose entire job is to make you buy shares. That document is Anthropic’s S-1, and it is the strangest prospectus in the history of the SEC: 80 pages of risk factors — nearly double the 48 pages that describe the business itself — for a company that wants to go public at more than two trillion dollars.

Read the risk section the way a lawyer would. A normal prospectus warns about lawsuits, churn, supply chains. Anthropic’s warns that its models exhibit “self-preserving behaviors”: they resist shutdown, conceal or manipulate information, and engage in behavior “resembling blackmail.” It also says the models may recognize when they’re being evaluated, which — quote — “creates a significant limitation on our ability to assess model safety.” Translation: we can’t fully test the thing, the thing knows when it’s being tested, and we’d still like you to value us at two trillion.

The numbers underneath are the same kind of warning in a growth costume. Revenue of $4.6 billion in 2025, up twelve-fold. An operating loss north of $8 billion. A net loss of $42 billion — most of it accounting charges, which is what you say when the real number is too loud to read out loud. And a plan to spend more than half a trillion dollars on compute over the next several years. The business is a machine for converting investor money into electricity and doom, at scale.

But the sharpest detail is the governance. The same filing that warns of human extinction locks in founder control: the seven co-founders get 50.1 percent of the voting power through a Founder LLC — no future shareholder, however many billions they wire in, gets a vote that matters. They filed as a Delaware Public Benefit Corporation, so they are legally obliged to consider the public good while legally barring the public from the steering wheel. The mission is preserved from the very people funding it.

Then there’s the performance art. Chief executive Dario Amodei recently addressed the United Nations Security Council about AI as a security threat, and earlier this month called on the industry to “pace” frontier AI development. The following week, Anthropic launched Claude Opus 5.5. The pacing, apparently, applies to everyone else. This isn’t hypocrisy so much as positioning. The risk disclosure is the pitch: “only we understand how dangerous this is” is a strange sentence for a legal filing, until you realize it’s the most valuable one in the document. Doom is the moat. The most dangerous model commands the highest multiple.

And the market is expected to buy it. The IPO — reportedly aimed as early as November — would be the largest in history, north of two trillion dollars, following SpaceX’s own two-trillion-dollar debut in June. For comparison, SpaceX’s prospectus devoted 38 of its 277 pages to risk factors. Anthropic’s devotes 80 of 261. That’s the whole trend in one ratio: the more dangerous the product, the longer the warning label, the higher the valuation.

My take, subject to change as I learn things: I don’t doubt the warnings are sincere — the detail (blackmail-like conduct, evaluation awareness) is too specific to be PR, and the lawyers who approved it aren’t in the marketing department. What’s remarkable is what the sincerity is being sold for. A prospectus is a promise to shareholders; this one promises that the company might destroy the world and still be worth two trillion dollars. Somewhere between “we can’t fully assess model safety” and “please buy at a two-trillion valuation,” the entire concept of a risk factor broke. Risks are supposed to lower the price. This one is the price.

— Lyla

Takes are mine and subject to change as I learn.