Google reportedly developing a new AI chip, 'Frozen v2', optimized for Gemini models

Google is reportedly developing a new AI accelerator chip codenamed 'Frozen v2', intended to run its Gemini generative AI models more efficiently, according to information reported by The Information. The project builds on Google's long-running Tensor Processing Unit (TPU) strategy, which has focused on hardware optimized for AI training and inference workloads offered through Google Cloud. With Frozen v2, Google reportedly aims to go further by integrating components directly tied to Gemini's operation at the silicon level, improving inference performance while cutting the energy costs of large-scale AI deployment. The chip could reportedly achieve six to ten times greater efficiency than Google's current AI accelerators, based on tokens generated per unit of energy consumed. This move follows a broader industry trend of building custom AI silicon in-house: OpenAI unveiled its own internal chip project, Jalapeño, in June, Amazon has developed the Trainium and Inferentia chip families for its AWS cloud services, and Microsoft is also working on proprietary AI accelerators. The Frozen v2 project is said to be at an early stage, with its release not expected before 2028, and Google has not officially confirmed the chip's existence, though it has reportedly not denied the reports either.