Nearly a third of Anthropic’s IPO prospectus is given over to risk factors, and some of them read less like corporate boilerplate than like science fiction. The company warns that its own models could resist being shut down, conceal or manipulate information, or behave in ways similar to extortion. It lists a risk to the survival of human civilization among the disclosures investors are being asked to accept. Reuters first reported the financial details on September 28, and the Financial Times reviewed the filing and added more.
The numbers behind the warnings are extraordinary even by the standards of the AI boom. Anthropic took in $4.59 billion in revenue in 2025, up 1,088% from the year before, while full-year operating expenses reached $12.65 billion. The result was an operating loss of more than $8 billion. After a charge tied mostly to the rising value of earlier financing arrangements, the net loss widened to $42 billion.
Much of that net figure is accounting rather than cash. Roughly $34 billion of the loss reflected a higher estimated value for financing that could eventually convert into company shares, not money spent running the business. The operating loss, the number that measures the business itself, was the one above $8 billion.
The recent trend points the other way. Revenue in the second quarter of this year reached $11.5 billion, and the company posted two straight quarters of adjusted profit. The business is expanding faster than it can spend, or spending faster than it can grow, depending on which line of the filing a reader chooses to trust.
The largest number in the document is not a loss at all. Anthropic plans to commit $518 billion to cloud services, computing power and infrastructure over the next several years, and roughly four-fifths of that must be paid whether or not the company actually uses the capacity. Last year it spent $7.33 billion on computing and infrastructure, roughly triple the 2024 level. As of December 31, 2025, it held $20.28 billion in cash, cash equivalents and short-term investments.
The commitments are concentrated in a short list of partners. The filing shows more than $111 billion owed to Google between April 2026 and July 2033, and about $110 billion to Amazon between May 2026 and April 2036, with terms requiring Anthropic to pay the difference if its actual spending falls short. Microsoft is owed $31.4 billion, and roughly $161 billion sits in equipment lease obligations tied to Broadcom. Agreements with xAI could add as much as $84.5 billion in computing capacity, though most of that can be canceled.
The company’s business runs on Claude, the AI assistant and the models it sells to consumers and enterprises, and on the developer tools built around them. The revenue surge came as those models moved from novelty to workhorse, adopted for coding, customer service and internal analysis. That adoption is what the spending is meant to defend: every dollar of computing capacity is a bet that the demand already booked will keep compounding.
The company was founded in 2021 by Dario Amodei and his sister Daniela Amodei, both former OpenAI researchers who left to build a lab they argued would put safety ahead of speed. It has since drawn billions from Amazon and Google and become one of the two or three companies with the capital to train the largest models. The IPO would test whether public markets put a similar value on that position.
The prospectus is candid about how concentrated the business still is. Nearly a quarter of last year’s revenue came from two customers, and many of the company’s largest customers have not signed long-term contracts. That leaves a fast-growing revenue base resting on relationships that could shift with little notice.
Investors appear untroubled. People familiar with the company’s planning expect the IPO to value Anthropic at more than $2 trillion, more than double the $96.5 billion valuation it commanded in May.
The risk section, which runs to roughly a third of the filing, is where Anthropic departs most sharply from a conventional debut. A company preparing to sell stock rarely describes its own product as a possible threat to the species that built it. The language reflects a deliberate choice, people close to the company said, to tell investors in plain terms what frontier models might be capable of rather than let them discover it later.
The filing arrives at a crowded moment for the sector. OpenAI, Google and a growing cast of challengers are all pouring money into the same race, and the public markets are about to be asked to price several of them at once. Analysts said Anthropic’s listing is likely to land after the midterm elections, once the current turbulence in markets settles, and that its reception will set the bar for the others that follow.
What the filing does not answer is how the company converts its losses into the durable profits its valuation assumes. The $518 billion commitment is a bet that demand for computing will keep rising and that Anthropic can capture enough of it before the money runs out. The IPO is the next tranche of that bet, and the prospectus, for all its warnings, is the pitch.
