Google’s Frozen v2 Chip Could Hardwire Gemini Architecture for AI Efficiency Gains

Google is reportedly developing an internal server chip called Frozen v2 that embeds Gemini’s AI model architecture directly into silicon. The chip is planned for deployment starting in 2028 and is described as a test run for specialized chips, with a smaller production volume than Google’s TPU line.

According to sources cited by The Information, Frozen v2 could be 6 to 10 times more efficient at serving AI responses than Google’s current TPU chips. The immediate takeaway: Google appears to be exploring tighter links between AI model design and the hardware that runs it.

Unlike Google’s TPUs, which are designed to work with many models, Frozen v2 would bake parts of Gemini’s model structure into hardware. The approach is compared to “freezing” parameters in AI models, where values are locked so they stop changing.

The reported design does not hardcode the model weights themselves. That matters because new weights could still be loaded onto the chip, making Frozen v2 more flexible than an earlier idea that would have worked only with a single Gemini version.

The chip is aimed at AI inference: the process of serving responses after a model has been trained. If Frozen v2 delivers the reported efficiency gains, it could lower Google’s internal cost of running powerful AI models.

That would matter commercially because AI companies increasingly compete on how cheaply and efficiently they can serve model outputs. The Decoder’s report says the chip could help Google offer lower prices and gain an advantage over OpenAI and Anthropic.

Frozen v2’s advantage depends on Google continuing to use the same underlying model architecture. Because of that constraint, the chip probably will not become a product for outside customers, unlike Google’s TPUs, which are leased or offered through cloud and on-premises programs.

Another open question is how much of Gemini’s architecture will actually be hardcoded. That decision has reportedly not been finalized, making Frozen v2 a notable but still developing bet on specialized AI hardware.