Google is designing a server-side AI chip, codenamed Frozen v2, that would implement parts of the Gemini model architecture directly in hardware, according to The Information. The approach aims to deliver 6–10x better energy efficiency than the latest TPUs and significantly more tokens processed per watt, potentially easing Google’s compute deficit — with a possible deployment as early as 2028. The trade-off is architectural rigidity: key model elements locked in silicon would require future Gemini versions to maintain compatible designs, tying model and infrastructure closer together. The project is still in early stages and has not been officially announced.

The Information — Google Plans New Frozen Chip to Run AI Models Efficiently