Anthropic locks in $200B of Google TPU capacity

Anthropic has committed to a five-year, $200B purchase of Google TPU capacity, with multi-gigawatt deliveries beginning in 2027 [Kursol analysis][TechCrunch].

── What shipped ──

The agreement layers on top of Google's earlier $40B cash-and-compute investment in Anthropic, announced in April. The new commitment is purely infrastructure: TPU access at scale, rather than fresh equity.

Anthropic's annual recurring revenue is now reported at roughly $30B annualized, ahead of OpenAI's $24B. At a $40B/year compute commitment, Anthropic is signalling that almost every dollar of revenue routes back into training and inference capacity for several years.

── Why it matters ──

Three things are now true.

One — the cost basis of frontier models is public. A $40B/year run-rate on compute is a number competitors will be benchmarked against. Anyone shipping a frontier model without that order of capacity now has a quantifiable disadvantage — and a clear capital ask to make.

Two — multi-cloud is a fiction. Anthropic is on AWS (Amazon investment), Google (TPU), and lists Microsoft Foundry as a deployment surface. But the training side is now near-exclusively Google TPU at this scale. Calling that "multi-cloud" stretches the term. For application developers using Anthropic's API, the dependency chain runs through Google whether you use them or not.

Three — the optionality window has closed. Smaller labs without sovereign-class infra commitments are unlikely to catch frontier capability on training alone. The competitive surface shifts to inference cost, post-training, and verticalisation.

── Editor's take ──

The interesting story is not that Anthropic spent $200B. It is that Google now has a five-year guaranteed buyer for TPU capacity, which lets them justify expansion that was previously hard to underwrite. The deal is more useful to Google than to Anthropic, and the financial press will eventually price that in.