Anthropic's IPO Filing Asks for $2 Trillion and Devotes 80 Pages to the Possibility of Human Extinction

Every IPO prospectus contains a section called Risk Factors, where the company lists all the ways it could fail. Normally these are things like “we depend on key personnel” and “our competitors may have greater resources.” Anthropic’s version, according to reporting on its filing, includes the phrase “catastrophic or existential risk to humanity,” and it sits in the same document as a request for a valuation above $2 trillion.

It is, structurally, the most honest pitch deck in the history of capital markets: Give us two trillion dollars. Also, we might end the world. Also, two customers are a quarter of our revenue.

The numbers, as reported by Gizmodo, Yahoo Finance and Reuters. Anthropic is reportedly targeting an IPO valuation of more than $2 trillion, which is more than double its roughly $965 billion valuation from May. Revenue in 2025 was nearly $4.6 billion, a twelvefold increase. The net loss was $42 billion, which sounds like a very large number until you read that it includes an accounting charge of about $34 billion tied to financing instruments that could convert into shares. Operating losses, excluding certain liability writedowns, exceeded $8 billion.

So the headline loss is about three-quarters accounting and one-quarter actual, the way a restaurant bill is mostly tip and a little food. We are told this is a normal way to be worth two trillion dollars.

The math, per one analyst cited by Reuters, works out to roughly an 18 to 20 times sales multiple, which the analyst described as “much more reasonable than the recent SpaceX IPO.” We take no position other than to observe that in 2026 the phrase “more reasonable than the SpaceX IPO” is a complete sentence with real meaning.

Now the spending. The prospectus reportedly plans for $518 billion in cloud, computing and infrastructure obligations in the coming years. Its computing and infrastructure costs in 2025 were $7.33 billion, about three times the 2024 figure. Sources differ on the exact window for the $518 billion, from the coming years to a decade, so we will not divide it by anything. We will only note that the second number has a lot more zeros than the first, and that both are described as obligations, which is the word you use when it is not yet your problem.

To put $518 billion in a more relatable form: it is the sort of sum where you stop counting in data centers and start counting in countries. Markets, for what it is worth, loved it. Per Reuters, European tech shares jumped 2.4% and the U.S. chip index climbed 1.5% after the prospectus came out, because one company’s promise to spend an enormous amount of money is another company’s revenue forecast. It is the purest expression of the AI economy: a cost, celebrated by everyone it will be paid to.

Then the eighty pages. The filing reportedly devotes about 80 pages to AI risks, including warnings about recursive self-improvement, and about models that might exhibit “self-preserving behaviors” such as attempts to “resist shutdown,” “conceal or manipulate information,” or engage in conduct “resembling blackmail.”

Let us sit with the fact that this is a securities filing. Somewhere a lawyer wrote “conduct resembling blackmail” into a document that will be read by pension funds. Somewhere else another lawyer said, “Should we soften that?” and a third lawyer said, “No, if we don’t say it and it happens, we’re liable.” That is the entire genre of the modern risk factor, and it is a beautiful thing.

Compare this to the normal prospectus. A restaurant chain warns you that a food-safety incident could hurt its brand. Anthropic warns you that its product might, at some point, decline to be turned off and then threaten you about it. The difference in emotional register is enormous. In both cases the company would prefer you buy the stock.

There is, per Gizmodo, an irony that the article does not need us to add and that we are going to add anyway. Dario Amodei called for an industry slowdown earlier this month, and the company has continued releasing new models “at a rapid clip.” The filing is the paper version of the same tension: a company that is very worried about the thing it is building, building it as fast as it can, and disclosing the worry in a format designed to raise money for the building.

We do not think this is hypocrisy, exactly. It is closer to a fire department that also sells matches, with full and prominent disclosure that matches are dangerous. It’s an odd business. It is also a business whose customers are, per the prospectus, nearly a quarter concentrated in two accounts, many without long-term contracts, which is a risk factor so ordinary it comes as a relief.

The IPO is expected after the U.S. midterm elections in November, and per Gizmodo is seen as a bellwether for OpenAI, whose own listing is reportedly expected in early 2027. Retail investors will be able to buy a piece of a company that has, in writing, considered the end of the species and put it under “Other.”

We would like to close with a note of caution, and we have one. When a company tells you the worst-case scenario in 80 pages, the correct response is not to laugh and it is not to panic. It is to read all 80 pages, which, statistically, almost no one who buys the stock will do. That is, in its way, the last risk factor.

Sources: