The Frozen v2 server chip targets a 6x to 10x efficiency increase over current TPUs by baking Gemini architecture directly into silicon. Google plans a 2028 release. This hardware integration aims to slash inference costs. Such a move gives Google a pricing advantage against competitors like OpenAI and Anthropic.