Google Develops Frozen v2 AI Chip to Cut Computing Costs, Targeting 6-10x Efficiency Boost by 2028

According to The Information, Google is developing an experimental AI chip codenamed Frozen v2, targeting deployment by 2028, with plans to embed Gemini AI architecture directly into the processor to reduce computational overhead and data transfer. The efficiency gain could reach 6 to 10 times compared to Google's current latest self-developed AI chip in terms of AI tokens processed per unit of power.

The project comes as Google faces significant AI computing resource constraints, forcing Google Cloud to decline some customer requests. The company spent approximately $1 billion monthly purchasing AI computing resources from SpaceX last month to meet enterprise client demands. Unlike existing Tensor Processing Units, Frozen v2 is designed to complement rather than replace TPUs, providing Google with additional AI computing flexibility as it reduces reliance on external chip suppliers.