Google Integrates Gemini Into Hardware: The New Frozen v2 Chip Is 6–10x More Efficient Than TPUs

Google is building a server chip called Frozen v2 that has the Gemini architecture embedded onto silicon. It’s claimed to be 6–10x more efficient than existing TPUs at handling AI workloads. It can also load new weights independently, meaning it won’t be rendered useless by future updates. It will be released in small numbers initially in 2028.

Embedding the architecture onto silicon makes each Gemini request much cheaper, which allows Google to undersell OpenAI and Anthropic. The AI chip arms race continues as Meta releases their MTIA in September, Microsoft moves to AMD, and now Google hardwires their own Gemini.