Anthropic strikes 3.5GW TPU pact with Google, Broadcom for 2027

Anthropic has reached an agreement with Google and Broadcom to secure 3.5 gigawatts of next-generation Google Tensor Processing Unit capacity, with initial deployments starting in 2027. The company expects most of the buildout to be located in the United States.

The compute will support training and inference for Anthropic's Claude models and enable greater use of custom silicon, reducing reliance on general-purpose graphics processors.

Broadcom disclosed in an SEC filing that the arrangement covers 3.5 gigawatts and is contingent on Anthropic's continued commercial performance. The filing also notes the parties are discussing operational and financing support for the rollout.

The partnership expands Anthropic's work with Google Cloud and Broadcom and follows additional TPU capacity announced in October. Anthropic characterizes the plan as a major extension of its November 2025 pledge to invest $50 billion in U.S. computing infrastructure.

Anthropic continues to train and run Claude across AWS Trainium, Google TPUs and NVIDIA GPUs, aligning workloads to the most suitable chips. Amazon remains the company's primary cloud provider and training partner, and the two continue to collaborate on Project Rainier.

Claude is accessible on Amazon Web Services through Bedrock, on Google Cloud through Vertex AI, and on Microsoft Azure through Foundry.

In 2026, Anthropic reported accelerating demand. Run-rate revenue has surpassed $30 billion, up from about $9 billion at the end of 2025. The number of business customers spending more than $1 million annually has risen to over 1,000, from more than 500 in February.

“We are building the capacity necessary to serve the exponential growth we have seen in our customer base while also enabling Claude to define the frontier of AI development,” wrote CFO Krishna Rao. “We are making our most significant compute commitment to date to keep pace with our unprecedented growth.”