Alphabet Google develops Frozen v2 chip hardwiring Gemini model architecture into silicon for six to ten times more efficient AI inference
Alphabet $GOOGL is developing a chip called Frozen v2 that hardwires its Gemini model architecture into silicon, projected to be six to ten times more power-efficient per token than current chips, per The Information. Release slated for 2028, aimed at easing Google Cloud's AI com
Alphabet $GOOGL is developing a new AI chip internally known as Frozen v2, designed to make its Gemini models dramatically more efficient to run, according to The Information. The approach etches parts of the Gemini model's architecture directly into the chip's silicon, and Google projects it will be six to ten times more power-efficient per token generated than its current AI chips, one of the largest single-generation efficiency gains reported in AI inference hardware. The chip is slated for release in 2028 and would sit alongside Google's existing tensor processing units rather than replace them. The project is reportedly aimed at addressing an AI computing capacity crunch inside Google that has caused internal tension and led Google Cloud to decline some outside customer deals in order to preserve capacity for internal use. The story underscores a shift across the AI industry: compute availability, not model capability, has become the binding constraint on how much AI companies can actually deploy, pushing even the largest cloud providers toward custom, model-specific silicon.