Google Shifts TPU Distribution Strategy, Selling Custom Chips Directly to Select Customers

Summary

Google announced during its Q1 2026 earnings call that it will begin selling its Tensor Processing Units (TPUs) directly to select customers for installation in their own data centers, marking a significant departure from its previous rental-only model where customers accessed TPU capacity through Google's cloud services. The move comes as part of Google's broader effort to expand its addressable market and compete directly with Nvidia's dominance in AI chip supply.

The announcement follows Google's recent introduction of two new TPU variants—the TPU 8t optimized for AI training and the TPU 8i designed for inference workloads. Google has already committed to major capacity agreements with Anthropic (expected to begin in 2027) and a reported multibillion-dollar chip deal with Meta, signaling strong enterprise demand for custom AI silicon beyond cloud rental models.

Google is not alone in challenging Nvidia's market leadership. Amazon has similarly monetized its custom chips (Graviton, Trainium, and Nitro) with an estimated annual run rate exceeding $20 billion, and has signed capacity agreements with both Anthropic and OpenAI. While Nvidia has publicly dismissed these competitive threats, citing superior flexibility and developer ecosystem strength, the shift toward alternative AI chip suppliers reflects enterprises' growing demand for ownership, cost control, and reduced cloud vendor lock-in.

• This announcement intensifies competitive pressure on Nvidia, with both Google and Amazon now offering custom AI chips as viable alternatives
• The shift reflects broader enterprise demand for hardware ownership and independence from cloud vendor lock-in