
Google Reportedly Developing AI-Specific Chip "Frozen v2"—Aiming for 6-10x Inference Efficiency by Hardwiring Part of Gemini's Architecture
Reports indicate Google's "Frozen v2" hardwires part of Gemini's architecture, targeting 6-10x the inference throughput per watt compared to TPUs, with deployment possibly as early as 2028. Keeping pace with model updates remains a constraint.








