Anthropic’s S-1: The Real Opportunity May Not Be Anthropic

Most readers will come to Anthropic’s S-1 looking for a valuation answer: how much is the company worth, and can it justify the market’s expectations?

We think the more useful question is simpler: if Anthropic’s plan works, who gets paid first?

The headline $80.7 billion first-half net loss is eye-catching, but it is not the heart of the filing. A large part of that number reflects non-cash fair-value movements in convertible notes, derivative liabilities, and SAFEs.

The more consequential disclosure sits in the infrastructure section: a set of long-term compute arrangements that reads less like a conventional software procurement plan and more like an industrial order book.

Adding the seven named arrangements with Broadcom, SpaceXAI, Google, Amazon, Microsoft, Nscale, and Lambda produces roughly $637.0 billion of nominal commitments.

That is GoAI’s sum of the disclosed line items, not a total reported by Anthropic. Of that figure, approximately $430.6 billion sits in Broadcom, Google, Amazon, and Microsoft arrangements with minimum-spend or largely non-cancellable features.

The total excludes Google’s separate $125–150 billion estimate for additional TPU capacity, which depends on capacity and deployment conditions, as well as AMD’s potential initial capacity spend of more than $20 billion.

That distinction matters. A model company earns revenue one token, one API call, and one agent workflow at a time. Its infrastructure partners are increasingly paid through multi-year commitments, equipment leases, and take-or-pay provisions.

Anthropic’s filing is…