Google reportedly building 'Frozen v2' chip that hardwires Gemini into silicon

Google is developing an inference chip codenamed 'Frozen v2' that bakes elements of Gemini's architecture directly into the silicon to cut calculations and data movement, The Information reported. Engineers on the project expect it to process roughly 6–10x more tokens per watt than Google's latest TPU, with deployment targeted around 2028, as surging Gemini demand strains internal compute and has led Google Cloud to hold back from some external customer deals. Google did not confirm or deny the report, and Alphabet shares rose on the news.