Google is developing a new AI server chip, internally called Frozen v2, that is designed to run its Gemini models more efficiently by embedding parts of the models directly into the chip’s hardware, according to a report by The Information.
The chip is expected to process six to 10 times more AI tokens for the same amount of electricity than Google’s latest tensor processing units (TPUs), if internal targets are met.
Unlike Google’s general-purpose TPUs, Frozen v2 is being built specifically for inference, the stage where trained AI models generate responses. By hardwiring portions of Gemini’s neural network architecture into silicon, Google expects to reduce the amount of computation required to serve AI models while lowering power consumption. The chip is not intended to replace TPUs but to complement them for large-scale inference workloads.
The project comes as Google faces growing demand for AI compute across its own products and Google Cloud.
According to the report, shortages in AI computing capacity have forced the company to prioritise internal workloads and turn away some external cloud customers. Improving inference efficiency could allow Google to serve more AI requests without expanding data centre capacity at the same pace.
The report said Frozen v2 could enter production as early as 2028. The chip is expected to be used primarily for Google’s Gemini models, rather than supporting a broad range of AI workloads like existing TPUs.
Google has expanded its custom silicon efforts over the past decade to reduce reliance on third-party AI hardware. Its TPU lineup powers AI training and inference across Google services and Google Cloud, where the company also offers TPU access to external developers and enterprises.
ALSO READ: Alteryx Inspire 2026: Three Questions Every Data Leader Should Take to Orlando
