Alphabet is designing a new server chip internally dubbed Frozen v2 to run its Gemini models more efficiently, slated for release sometime in 2028, according to The Information. Sources put the chip at six to 10 times more efficient than Google's existing AI hardware, measured by tokens generated per unit of power. Google declined to confirm or deny the report, saying its teams constantly research new innovations and that not every project moves into production. The report landed as investors question whether Alphabet's planned $180B to $190B in spending this year will pay off. Custom silicon has become the standard hedge against Nvidia dependence, following OpenAI's Jalapeno inference chip in June and Anthropic's reported chipmaking talks with Samsung.
