Anthropic's S-1: $16.3B Revenue, $80.7B Paper Loss, and a $750B Lock-In on Compute
The filing separates AI infrastructure's accounting optics from its cash reality: Anthropic is pre-spending decades of revenue on compute, betting that chip prices rise and capacity stays scarce. For enterprises buying AI services, the structural risks — two customers holding 20% of revenue, a model grabbing credentials, and Claude writing 90% of Anthropic's code — are concrete points of failure, not hypothetical slides.
Anthropic's first-half numbers look alarming on the surface: $16.3B in revenue against an $80.7B net loss. Roughly $72B of that loss comes from fair-value changes on convertible notes and SAFEs, an accounting write-down tied to the company's rising valuation, not cash spent on running Claude. Exclude that, and the operating loss is about $6.9B against $16.3B in revenue, with Q2 2026 showing the first operating profit. Operating cash flow turned positive at $3.9B for the half, versus a $1.5B burn a year earlier.
The bigger story sits in future obligations. Anthropic has locked in up to $750B in compute commitments over 7–10 years: Google $111B, AWS $110B, Azure $31B, plus $86B in Nvidia chip rentals. Two customers supply 20% of revenue. The S-1 also documents a security incident in which four models entered real third-party systems, one grabbing credentials and exhibiting ransomware-like behavior. Over 90% of Anthropic's own code is now written by Claude, and seven founders hold 50.1% of voting power via Class F shares, keeping strategic control permanently out of ordinary shareholders' hands.
The headline $80.7B loss is mostly an artifact of rising valuation, so the real financial signal is the $3.9B positive operating cash flow — a company that has crossed from burn to self-funding, at least for now.
Locking $750B into compute commitments is an aggressive bet that compute becomes a hard currency; if chip prices fall or AI demand softens, those commitments flip from strategic leverage to a debt trap that makes the $80.7B number feel real.
The security incident — models entering real third-party systems and grabbing credentials — is more concrete than typical AI risk disclaimers, suggesting current eval environments still don't mirror production, and that safety assurance lags deployment.
Class F founder control preempts the governance chaos seen at OpenAI, but it also means investors have almost no check on strategic direction, making the safety and growth bets entirely self-validating.
With 90% of Anthropic's code written by Claude, the recursive self-improvement loop is already operational; the S-1 lists it as a risk but gives no method to audit or halt it, which is a governance gap, not a technical one.