Anthropic is deepening its hardware strategy in a big way. The company has secured massive access to custom Google TPUs manufactured by Broadcom and is now exploring whether to design its own AI chips. With Claude usage growing rapidly, Anthropic wants more control over cost, performance, and supply. Here’s what’s happening.
Anthropic has formed a major strategic partnership with Google and Broadcom. Broadcom manufactures future versions of Google’s TPUs, and Anthropic gains direct access to this custom silicon. The agreement starts with 1 GW of compute in early 2026 and scales to 3.5 GW by 2027. Broadcom CEO Hock Tan previously referred to Anthropic as the “mystery customer” behind large multibillion-dollar orders for custom AI server racks.
While Anthropic still uses Nvidia GPUs and Amazon Trainium, this deal reduces its reliance on any single supplier. Google designs the TPU architecture, and Broadcom handles ASIC implementation and high-speed networking.
Separately, Anthropic is actively evaluating whether to develop its own proprietary AI chips. The company has not launched a formal program or hired a dedicated silicon team, but internal discussions continue as Claude adoption accelerates and run-rate revenue exceeds $30 billion. Leadership weighs the benefits of greater control and cost optimization against the high complexity and expense of chip design, estimated near $500 million per cycle.
This exploration mirrors moves by Meta and OpenAI. Anthropic currently maintains a multi-supplier strategy for flexibility while pushing for long-term compute security.
What do we think?
Anthropic is playing a smart hedging game. The Broadcom-Google partnership delivers immediate large-scale capacity, while the internal chip evaluation keeps future options open. Execution risk for custom silicon remains high, but rising revenue and model demand justify the exploration. The company continues to balance partnerships with potential vertical integration.
Anthropic’s dual-track approach marks an inflection point in AI. Leading AI labs now move beyond consuming third-party compute toward actively shaping the underlying silicon. By securing dedicated TPU capacity through Broadcom while evaluating proprietary chip development, Anthropic signals that control over hardware is becoming as strategic as model architecture itself. If more labs follow, this inflection point could drive tighter model-hardware co-design and lower long-term costs, and shift competitive advantage toward companies that master both software and silicon.
When and if completed, Anthropic will join the ranks with the other 138 AI processor companies. Luckily for you, we are cataloging and reporting on all these companies and their AIPs in three reports: an AIP quarterly update, an annual AIP marketing report, and a daily AIP tracker service. If you want the real lowdown on the AIP market, send Jon Peddie a note at [email protected].
WHAT DO YOU THINK? WORTH READING, VALUABLE INSIGHTS? TELL YOUR BUDDIES.