Google is reportedly developing a new AI inference chip called Frozen v2 that hardwires elements of Gemini’s architecture directly into silicon, cutting down on computation and data movement when running the model.
Internally, the company expects it to process 6 to 10 times more tokens per watt than its latest TPU. Google plans to deploy the chip as early as 2028.
The goal is to alleviate a compute crunch that’s already forced Google Cloud to turn away some external customers.
Frozen v2 would exist alongside Google’s TPUs. While a TPU can run a variety of models, Frozen v2 trades flexibility for efficiency.
It’s also a big architectural bet: if Gemini’s underlying design changes significantly down the road, these chips could become obsolete.