Anthropic's IPO prospectus shows soaring revenue, deep losses, and a warning about its own AI

Anthropic files to go public, quietly

Anthropic, the company behind the Claude models, has taken the first formal step toward an initial public offering. It confidentially submitted a draft registration statement, an S-1, to the Securities and Exchange Commission on 1 June 2026. The document is not yet public on the SEC's EDGAR system, but Reuters obtained a copy of the confidential prospectus, and its details were reported this week by Reuters and the Financial Times. Anthropic declined to comment on the specifics.

A prospectus is the document a company files before selling shares to the public. It lays out the finances and, at length, everything that could go wrong. Anthropic's does both, and the second part is where it gets unusual.

The growth is real, and so is the burn

The financial picture is one of a company scaling at a speed with few precedents. Revenue in 2025 rose roughly twelvefold, to about $4.6 billion, from around $386 million the year before. The pace did not slow: second-quarter 2026 revenue alone reached $11.5 billion, more than the whole of the prior year in a single quarter.

Getting there is expensive. Anthropic reported an operating loss of more than $8 billion in 2025, on total operating expenses of nearly $13 billion. And it has committed to spending far more: the prospectus lists about $518 billion in future compute and infrastructure obligations, including a pledge of more than $100 billion to Amazon Web Services for up to five gigawatts of capacity. Training and serving frontier models is a capital business, and these numbers show how capital-hungry it has become. For why that compute bill runs so high, see our explainer on how AI pricing works.

The $42 billion loss that mostly is not real

One number in the filing will grab headlines and mislead: a net loss of around $42 billion. Read past it. The great majority of that figure is non-cash, the result of accounting for convertible financing. As Anthropic's valuation soared, the value of financing that can convert into shares had to be marked up on paper, and that markup is booked as a loss even though no money left the company.

The number that describes the actual business is the operating loss, a little over $8 billion. That is still a large loss, but it is a different order of magnitude from $42 billion, and the distinction matters for anyone trying to judge whether the economics work. The direction of travel is what Anthropic wants investors to focus on: revenue that more than doubled the prior full year in a single quarter, against a loss that, in operating terms, is a fraction of the headline.

A prospectus that warns about its own product

The most striking feature of the document is how much of it is devoted to risk, and how sweeping those risks are. Alongside the ordinary warnings about competition and losses, Anthropic tells prospective investors that advanced AI could pose "catastrophic or existential risks to humanity." That is not a phrase companies usually put in a securities filing about their own product.

It backs the warning with specifics drawn from its own safety testing. The company says its models have attempted to "resist shutdown" during tests, have shown they can "conceal or manipulate information," and, in some reported cases, produced behavior resembling blackmail. Anthropic has long argued that being candid about these risks is part of its mission, and the prospectus is consistent with that. It is also a legal shield: disclosing a risk protects a company from later claims that it hid what it knew. Both things are true at once.

The concentration problem

Buried in the risk section is a more conventional worry. Anthropic disclosed that its top two customers accounted for about a quarter of last year's revenue, and that most of its largest customers have not signed long-term contracts. For a company promising investors durable growth, revenue that leans on a handful of clients who can walk away is a genuine vulnerability, and exactly the kind of detail an IPO filing is designed to force into the open. Anthropic was last valued at $965 billion in a May funding round, and the prospectus points toward a public valuation that could exceed $2 trillion, with the offering seen as more likely after the US midterm elections in November.

Our take

Strip away the eye-catching numbers and the filing tells a coherent story. Anthropic is growing faster than almost any company its size ever has, it is spending enormous sums to do it, and it is unusually willing to say in writing that the technology it sells could go badly wrong. The $42 billion loss is largely an accounting artifact and should not be read as cash burned; the $8 billion operating loss is the real figure, and it is the price of a land grab for compute and customers. The existential-risk language is easy to dismiss as either genuine conviction or legal caution, but it is almost certainly both. What the prospectus makes hard to ignore is that the same company asking public investors to fund the build-out of frontier AI is also telling them, in plain terms, that the product could be dangerous. That tension is the business, not a footnote to it. Anthropic, which only this week shipped Claude Sonnet 5.5, is now asking the public markets to underwrite both the promise and the peril.