Expanding our use of Google Cloud TPUs and Services

Source: Anthropic Date: 2025-10-23 URL: https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services

Summary

Anthropic announced plans to expand Google Cloud infrastructure by up to one million TPUs, valued at “tens of billions of dollars,” bringing over 1 GW of capacity online in 2026. Purpose: serving 300,000+ business customers and supporting research/alignment work. Anthropic noted a diversified compute strategy across Google TPUs, Amazon Trainium, and NVIDIA GPUs, with Amazon remaining “primary training partner” via Project Rainier.

Implications

• Capital / compute infrastructure thread. “Tens of billions” in Google Cloud TPUs alongside the $8B Amazon commitment means Anthropic’s total committed compute infrastructure is now in the $20B+ range. At this scale, compute is the binding constraint — everything else (product, sales, research) scales around available inference and training capacity.
• 1 GW capacity in 2026. A gigawatt of AI compute is enormous — comparable to a major hyperscaler’s dedicated AI infrastructure. This is the supply-side answer to the OSTP submission’s “50 GW by 2027” demand-side ask.
• Multi-cloud diversification. Explicitly maintaining Google TPUs + Amazon Trainium + NVIDIA GPUs is strategic: vendor lock-in to any single hyperscaler would be existentially risky at Anthropic’s scale. The three-vendor approach creates negotiating leverage and redundancy.
• Project Rainier. Amazon’s “Project Rainier” is Anthropic’s primary training cluster — this is the first public name for what was previously described generically as AWS training infrastructure. Worth tracking as Rainier’s scale and timeline become clearer.
• Watch: when the 1 GW Google capacity comes online; how TPU v5/v6 performance compares to Nvidia H100/H200 for Anthropic’s training workloads; whether the Google commitment shifts the AWS/Google balance.