Claude runs simultaneously across AWS Trainium, Google TPU, and Nvidia GPUs. Anthropic has live gigawatt-scale commitments across all three. Amazon remains the primary training partner. Anthropic also committed $30B to Azure with 1GW of Nvidia Grace Blackwell and Vera Rubin coming online. The 3.5GW TPU deal doesn't deepen Anthropic's dependency. It deepens Google's dependency on Anthropic as its largest external proof-of-concept customer. The moat narrative has the direction inverted.
The new 3.5GW is in addition to 1GW already coming online in 2026 under the Google Cloud agreement, and the filing states that Anthropic's use of the expanded capacity is contingent on its continued commercial performance. That contingency clause is the most underreported detail in the coverage. Mizuho projected $21B in Broadcom revenue from Anthropic in 2026 and $42B in 2027. Both figures are backstopped by Anthropic's revenue trajectory, which has surpassed $30B annualized, more than tripling from roughly $9B at end of 2025.
The numbers currently support the thesis. But the structure means Broadcom and Google carry more concentration risk here than Anthropic does.
Broadcom also locked in a supply assurance agreement with Google for networking and components through 2031 which is the less-discussed part of the filing and probably the stickier commercial relationship. The TPU-Anthropic piece is headline-grabbing. The Google-Broadcom networking lock-in is quieter and arguably more durable.