Woofun AI reports that Google is developing the Frozen v2 AI inference chip, which integrates specific Gemini architecture components directly into hardware to minimize computation and data movement. The company targets a deployment timeline as early as 2028, with internal projections indicating token processing per watt could reach six to ten times that of current TPUs.

This initiative aims to mitigate computing power shortages that have already compelled Google Cloud to decline certain external customer orders. While Frozen v2 prioritizes efficiency over flexibility compared to versatile TPUs, the strategy carries architectural risk; significant future redesigns of the Gemini model could render these specialized chips obsolete.