Google hardwires Gemini into silicon as the AI compute crunch gets a 10x plot twist 🧊

Google is developing a custom chip that bakes part of Gemini's model architecture directly into its circuitry, a design shift aimed at running the AI more efficiently and easing a capacity crunch that already forced the company to turn away business. The chip, codenamed Frozen v2, was reported by The Information on Monday. Rather than acting as another Tensor Processing Unit upgrade, the chip locks in elements of Gemini's structural blueprint so the hardware can skip redundant calculations on each query, while the model's trained weights remain updatable. Engineers project a six to ten times improvement in tokens generated per watt of electricity consumed.

The move comes as Google struggles to serve the AI demand it has already created. In March, Google told Meta it could not fill the volume of Gemini compute Meta wanted to purchase, prompting Meta to instruct employees to ration their AI usage. Google is spending up to $190 billion on AI infrastructure this year, yet still lacks enough servers to meet customer requests, and competitors such as OpenAI, Anthropic and Chinese labs already account for up to 45% of U.S. company AI token usage in part because they run 60–90% cheaper.

Alphabet shares climbed roughly 3% during Monday's session on the news, touching $356 intraday, though the gains receded in Tuesday's session as investors awaited Q2 2026 earnings due Wednesday, July 22. Frozen v2 is the latest step by a major AI company to reduce reliance on Nvidia, which controls roughly 85% of the GPU market for AI; Meta, Amazon, Microsoft and OpenAI all maintain custom silicon programs of their own for the same reason.