Anthropic says it is profitable on an adjusted operating basis — while committing extraordinary sums to future compute

Anthropic has told investors it expects positive adjusted operating income for a second consecutive quarter. The FT says its gross margins are above 80% before revenue-sharing payments to distributors such as Amazon and before model-training expenses. That is a significant milestone for a frontier lab because inference economics have been one of the biggest unanswered questions around the industry. Anthropic is simultaneously preparing for a possible Nasdaq IPO.

But “profitable” needs a large asterisk. The Information’s excellent catch-up analysis calculated that Anthropic has signed agreements covering at least 14.8GW of potential compute capacity, with maximum contractual value of roughly $517bn over coming years. Amazon and Google alone account for 11GW and more than $300bn of that estimate; Microsoft adds roughly 1GW, while SpaceX, Lambda, Nscale and others provide additional capacity. These figures are contract ceilings and long-term reservations, not $517bn already spent, and some agreements have conditions or cancellation rights.

This creates an interesting financial structure. Claude can have excellent gross margins on today’s inference while Anthropic simultaneously locks itself into enormous future capacity because being unable to serve demand would be strategically catastrophic. The economic question therefore becomes less “is each token profitable?” and more “can demand grow fast enough to keep 15GW of pre-booked infrastructure highly utilized?”

Frontier AI is beginning to look like a hybrid of software economics and an airline-style capacity business — enormous margins when infrastructure is full, enormous forecasting risk when it is not.