← All signal stories
§ SignalApr 7, 2026 · Issue 17 · Story 3

Anthropic Locks In Multi-Gigawatt TPU Capacity, Betting Its Frontier Model Future on Google and Broadcom Silicon

Anthropic has signed a multi-party compute agreement with Google and Broadcom securing multiple gigawatts of next-generation TPU capacity, with deployment beginning in 2027.

3. Anthropic Locks In Multi-Gigawatt TPU Capacity, Betting Its Frontier Model Future on Google and Broadcom Silicon

Anthropic has signed a multi-party compute agreement with Google and Broadcom securing multiple gigawatts of next-generation TPU capacity, with deployment beginning in 2027. The deal is structured to support both training and inference for Anthropic's frontier Claude models, representing one of the largest publicly disclosed AI compute commitments tied to a single model family. The announcement came directly from Anthropic's official account, with no financial terms disclosed.

This deal reshapes the competitive calculus in a meaningful way. Anthropic deepens its structural dependency on Google, which is already a major investor, while simultaneously giving Google's TPU program a marquee customer capable of validating next-generation hardware at scale. Broadcom, which designs custom AI accelerators for hyperscalers, gains a high-profile deployment anchor that strengthens its position against Nvidia in the custom silicon race. The losers here are implicit: Nvidia does not appear in this agreement, and any rival frontier lab without locked-in gigawatt-scale compute by 2027 faces a widening infrastructure gap. OpenAI, which is building toward its own dedicated compute through the Stargate joint venture with Microsoft and SoftBank, is the most direct strategic counterpart to watch.

The broader signal is that frontier AI is bifurcating into camps organized around compute alliances rather than just model capability. Anthropic, OpenAI, and xAI are each anchoring to distinct hardware and capital ecosystems, and the 2027 timeline suggests all three are planning for training runs that dwarf anything publicly announced today. Gigawatt-scale reservations made three-plus years in advance indicate the leading labs believe compute scarcity, not algorithmic differentiation, will be the binding constraint at the frontier.

Source: https://twitter.com/AnthropicAI/status/2041275561704931636