Google is reportedly developing Frozen v2, an AI inference chip that hard-codes parts of Gemini's architecture directly into silicon. The goal: slash the compute and data movement required to run models by bypassing general-purpose flexibility.
Internal projections suggest the chip could process 6 to 10 times more tokens per watt than Google's latest TPUs. The company plans to start deploying it as early as 2028.
The project is driven by a painful compute crunch — one that's already forced Google Cloud to turn away some external customers. Frozen v2 is designed to ease that bottleneck.
It won't replace TPUs. Instead, the two chips will coexist: TPUs remain the flexible workhorses that can run any model, while Frozen v2 trades that versatility for raw efficiency.
But it's also a big architectural bet. If Gemini's underlying design changes significantly down the road, these chips could become expensive paperweights.