Overview
- Multiple outlets reported Monday that Google is developing an internal server chip nicknamed Frozen v2 that would embed elements of its Gemini model directly into the hardware.
- Engineers quoted in the reporting estimate the design could serve roughly six to ten times more AI tokens per watt than Google’s current TPU inference chips, though those claims come from internal projections and are unconfirmed by the company.
- Google is said to be finalizing how much of Gemini’s architecture would be hardwired and is targeting possible deployment around 2028 if the design moves forward into production.
- The project is described as a specialized complement to Google’s TPU line intended to ease an internal AI compute crunch and reduce reliance on third‑party accelerators, but it would limit flexibility if future Gemini architectures change.
- Investors reacted positively to the reports with Alphabet shares rising about 3%, and the move follows broader industry efforts by competitors to build custom inference silicon to cut data‑center power and costs.