Google Developing Model-Specific 'Frozen' AI Chip to Cut Gemini Inference Costs, Report Says

Google is reportedly building a new class of server chip, code-named "Frozen," that hardwires parts of its Gemini models directly into silicon to run inference far more efficiently, according to a report from The Information. The design, referred to as "Frozen v2," would form a new chip line distinct from Google's existing Tensor Processing Units and could enter deployment as early as 2028.