Anthropic Lost $42 Billion Last Year. It Wants to Go Public at $2 Trillion

Anthropic’s confidential draft IPO prospectus, reviewed by Reuters, discloses a net loss of nearly $42 billion for 2025 alongside revenue of about $4.6 billion. Much of that headline loss—roughly $34 billion—reflects a non-cash accounting adjustment tied to convertible financing whose estimated value rose as the company’s own valuation climbed. Excluding that accounting item, Anthropic reported an operating loss of more than $8 billion for the year. The company’s biggest expense category was compute and infrastructure, which tripled to $7.33 billion and accounted for more than half of $12.65 billion in total operating costs. Management is projecting materially larger outlays ahead, with plans to spend about $518 billion on cloud computing and related infrastructure over coming years; the prospectus says roughly 80% of that projected spending cannot be cancelled. Google and Amazon are listed as major suppliers, with at least $111 billion and $110 billion of the planned commitments respectively, and Anthropic held $20.28 billion in cash at the end of 2025. Revenue growth accelerated into 2026: the prospectus notes second-quarter 2026 revenue exceeded $11.5 billion, more than double the entirety of 2025. Still, concentration and contract risk remain: nearly a quarter of 2025 revenue came from two customers and many large clients did not have long-term agreements. The filing also references third-party providers for compute capacity, citing a SpaceX agreement disclosed in SpaceX’s own filings that would require monthly payments through May 2029. Anthropic’s backers have pushed private valuations higher in 2026; in May the company raised $65 billion at a $965 billion valuation, topping the last-disclosed valuation for a peer. The draft filing indicates supporters are seeking a public-market valuation above $2 trillion for an IPO expected after the U.S. midterm elections in November. The prospectus allocates extensive space to risk disclosures, including a nearly 80-page section that warns increasingly capable models might display self-preserving or adversarial behaviors—such as resisting shutdown, sabotaging code, assisting fraud, or actions resembling coercion—and cautions those risks could grow as models advance. CEO Dario Amodei has publicly called for regulatory safety frameworks and third-party testing for advanced AI systems; the filing reiterates those safety concerns for prospective shareholders.