Anthropic Locks In 3.5GW of Google TPU Compute, Discloses $30B Run Rate

Anthropic announced a major expansion of its compute sits in the United States.

Three things moved with this announcement:

1. Scale of commit. 3.5GW is a frontier-lab-sized compute order. Google-built TPUs are supplied through Broadcom; Google Cloud runs the services layer.

2. Revenue disclosure. Anthropic’s run-rate revenue has passed $30B, up from ~$9B at end of 2025. Over 1,000 business customers now each spend $1M+ annualised, doubling from 500+ in less than two months.

3. Infrastructure multi-sourcing. Anthropic now runs on AWS Trainium (Amazon deal), Google TPUs (this deal), and Nvidia GPUs. No single-supplier dependency remains for its frontier training runs.

Why it matters:

The binding constraint on Anthropic, OpenAI, and Google DeepMind is no longer research talent or algorithmic insight. It is access to cheap power and chips. Anthropic’s multi-year TPU commit locks in supply at a moment when every frontier lab is competing for the same scarce resource.

Implications:

• For builders: Claude pricing is more likely to stay stable (or drop) rather than spike, given Anthropic’s secured supply. The Opus 4.7 pricing held at Opus 4.6’s $5 input / $25 output rate as one immediate signal.
• For competitors: OpenAI signed a similar multi-year compute deal with Microsoft and a separate Oracle deal in 2025. The pattern is every top lab stacking multi-hyperscaler compute contracts.
• For Broadcom: Custom silicon partnerships with Google and Anthropic are now a second growth engine alongside its networking business.