Google just bet its inference future on a chip built for one model - The New Stack
Google's reported "Frozen v2" chip would hardwire Gemini's architecture into silicon, potentially delivering 6–10x more tokens per watt and reshaping AI inference economics.