US Markets
Home›US Markets›Sectors›Google develops chip for Gemini AI models, codenamed F…
Google develops chip for Gemini AI models, codenamed Frozen v2
The chip is expected to go live by 2028 and could cut inference power needs by about 83% to 90% for the same token volume, according to an analysis citing insider expectations.
Google is working on a purpose-built chip for its Gemini AI models, an effort reported by Yahoo Finance citing The Information, with the project described as codenamed Frozen v2.
Insiders believe Frozen v2 could be roughly six to ten times more efficient than Google’s TPUs, which are developed in collaboration with Broadcom. The reporting says the chip would embed parts of Gemini’s architecture into silicon and focus on inference efficiency, with model weights loaded when needed rather than stored permanently.
The timeline in the report points to a potential rollout by 2028 and frames the chip as a way to improve Gemini’s competitiveness, particularly because Google’s models were not described as being near the top 10 on an industry leaderboard at the time.
The analysis also links the efficiency goal to Google’s scale of usage, noting that in June 2026 Alphabet said its model interfaces processed about 19 billion tokens per minute, with total token usage across products reaching 3.2 quadrillion per month. The report estimates that at six times efficiency power required for a given volume of tokens could drop by about 83%, and at ten times efficiency it could fall about 90%, though it cautions that overall savings would be lower after accounting for memory, networking, cooling, and other infrastructure costs.