Anthropic Goes Fully Vertical: 1 Million TPUs, Owned Infrastructure, and the End of Cloud Dependency

Anthropic is going fully vertical on the AI stack. The company will directly purchase close to 1 million Google TPUv7 units and deploy them in facilities it controls—a significant diversification away from pure NVIDIA dependency. Broadcom will sell the systems directly to Anthropic, while TeraWulf, Hut8, and Cipher Mining provide the economics of AI compute infrastructure — -arms-race/”>data center infrastructure.

Owned Infrastructure, Not Cloud Rental

Anthropic is purchasing chips directly and deploying in facilities it controls—a fundamentally different model from renting compute through AWS, Google Cloud, or Microsoft Azure. Ownership changes the economics and strategic positioning entirely.

This is vertical integration at unprecedented scale for an AI lab — as explored in the intelligence factory race between AI labs — .

Crypto Miners Become AI Infrastructure

TeraWulf, Hut8, and Cipher Mining—all Bitcoin mining companies—are pivoting their data center expertise to AI workloads. They have power contracts, cooling infrastructure, and real estate that AI companies desperately need.

The crypto infrastructure buildout now serves AI’s compute hunger.

Broadcom as TPU Supplier

Google designs TPUs; Broadcom manufactures them. This deal positions Broadcom as a direct supplier to AI labs, not just a component vendor to hyperscalers. The custom silicon market is fragmenting away from NVIDIA’s grip.

The Scale Is Massive

1 million TPUv7 units represents one of the largest single chip deployments outside the hyperscalers themselves. At projected TPUv7 performance levels, this gives Anthropic compute capacity rivaling the major cloud providers.

The model is emerging: AI labs that want ownership without operational burden outsource physical deployment while maintaining strategic control of the compute layer.

For deeper analysis of AI infrastructure strategy, subscribe to The Business Engineer.