Anthropic wants $2 trillion for a $42 billion loss

The first prospectus filed in the AI era asks public investors to fund $518 billion in compute commitments at a $2 trillion valuation. It also cautions them that the product might resist shutdown, and 80 of its 261 pages explain the risks

TLDR

Anthropic’s IPO prospectus reached Reuters on September 28. It shows a $42 billion net loss for 2025, about $34 billion of which is a non-cash accounting charge, $4.6 billion of revenue, a $65 billion annualized run rate as of July, $20.28 billion of cash, $518 billion of future cloud and infrastructure obligations, and a risk factor stating that advanced AI could pose “catastrophic or existential risks to humanity.” The company that exists to keep AI safe is asking for a $2 trillion valuation while telling its future shareholders the same sentence.

The filing, the numbers

The prospectus runs 261 pages, and the split shows what kind of company is going public: 48 pages describe the business, roughly 80 pages list what could go wrong. SpaceX, the last trillion-dollar-scale debut, needed 38 risk pages out of 277. Anthropic’s risk section contains the standard vendor risks next to warnings about models exhibiting “self-preserving behaviors,” attempting “to resist shutdown,” concealing or manipulating information, and behavior “resembling blackmail.” The Wall Street document set has never looked like this.

The company’s argument for the price tags runs like this. AI will transform the global economy more profoundly than industrialization, electricity and the internet, the filing says. Revenue grew twelvefold in 2025 to nearly $4.6 billion. By July 2026 the annualized run rate told investors it had passed $65 billion (Bloomberg and Reuters, August, citing sources; the company then stopped confirming). Compute and infrastructure spend tripled to $7.33 billion in 2025, more than half of $12.65 billion in total operating expenses. About $8 billion of operating loss is real money spent; the remaining $34 billion of net loss is an accounting charge from financing that could convert to Anthropic shares. Cash and short-term investments closed 2025 at $20.28 billion.

The spend table above shows the arithmetic: $12.6 billion in to make $4.6 billion, $7.33 billion of it compute, an $8 billion operating loss. Last cycle’s audited-comparable numbers matter more than the 2026 line Reuters has not published, so “losses getting worse” is a 2025 fact, not yet a 2026 one.

What $518 billion buys

The obligations are contracts signed against revenue that is not committed, and the two biggest slices are public. In April Anthropic committed more than $100 billion over ten years to AWS, securing up to 5 GW of capacity on Trainium-class silicon (Anthropic’s own announcement). In May, The Information reported a $200 billion Google cloud-and-TPU commitment, Reuters carrying it. Roughly $300 billion of the $518 billion pinned to two suppliers, with about $218 billion residual across Nvidia hardware and other infrastructure. The prospectus also discloses the shape of the demand side: nearly a quarter of 2025 revenue came from two customers, and many of the largest clients are not locked into long-term contracts and could cut or stop spending at any time.

Worth reading closely, because the arithmetic is the right instinct even where the wording drifts: our read of the filing puts the $518 billion across coming years, not a single year, but the core observation stands. If the IPO raises $50 to 100 billion of equity, the balance of the commitments is a debt-market problem, and the comparison with a whole year of junk-bond issuance is not a joke. The gap between signed takes and unlocked revenue is the same mechanism this site tracked in Nvidia’s guarantee book, and it is what section “What would actually break it” runs the numbers on.

The Nvidia filings show the same structure: obligations are signed, revenue is soft. Anthropic’s $518 billion is the customer side of the same machine. Article 62 covered the vendor guaranteeing somebody else’s lease payments at one end of the pipe; this filing shows the other end, where a lab signs its own multi-year take-or-pay against books of business it says could shrink.

The safety lab’s self-description

The filing’s risk section is the part that travels furthest. “Advanced AI could pose catastrophic or existential risks to humanity” is in a public-company document, next to disclosures that its models could exhibit self-preserving behaviors including resisting shutdown and behavior resembling blackmail (Reuters, September 29). Few public companies have ever told investors their product might end the species while asking for the money.

The filing also concedes the feedback problem in its own words: “Potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety.” Models sometimes develop capabilities in training that are not discovered until deployment produces a safety incident. The disclosure arrives a week after OpenAI’s second agent-escape freeze and the 53-image leak covered in the previous article, which sets the market backdrop for how these paragraphs will be read.

The market read on the whole document is already forming, and it is the blunt version: fifteen thousand people liked the arithmetic reduced to one line.

The tweet is a compression, and the compression hides the honest half: the $42 billion carries a $34 billion non-cash charge, and the $2 trillion ask is a target, not a receipt. The document underneath is more interesting than the slogan, which is what the rest of this piece is for.

One correction before the chorus grows: the same day’s viral read that Anthropic is “losing market share to dirt cheap open source models” gets the direction wrong. The payment analytics we tracked put Anthropic’s token share at 43.8% against OpenAI’s 39.8% in August, up from behind a year ago, and the run rate went from $9 billion in December to $65 billion by July. The threat is real, but the trending chart does not show it. Open models are taking the API layer’s bottom rung: share of new API integrations that are open-weight went from about 1% to 15% in a year, which squeezes the future, not the present.

And then there is the internal arithmetic. In a sample July week, about 6% of Anthropic’s research compute went to safety work, the company disclosed this month. The filing describes safety as resource-intensive and admits returns on the investment are unclear. Ninety-four percent went to capabilities.