Google develops Frozen v2 chip built around Gemini's architecture

The Information reported on 20 July 2026 that Google is developing an internal server chip called "Frozen v2", designed to run its Gemini models more efficiently. The Information said the report cited two people familiar with the project; Google has not confirmed it.

Alphabet shares closed 1.51% higher on the Monday the report emerged.

The chip embeds parts of Gemini's architecture permanently into the silicon, which The Information said cuts the calculations and data movement needed to answer a query. The design is described as embedding model architecture rather than model weights.

The name refers to "freezing" model parameters into the hardware. The concept reportedly originated with Google engineer Jeff Dean, whose earlier Frozen design built the model weights themselves into the chip. Google dropped that version because it would only work with one Gemini model and would go out of date too quickly.

Google's engineering estimate puts the potential gain at six to ten times more tokens served per unit of power than its newest TPUs, according to The Information. The design is still being worked on, and the figure is not final.

Google plans to deploy the chip from 2028, trialling it in smaller volumes than its main TPU line, The Information reported. It would complement rather than replace the general-purpose TPUs; Google announced separate training and inference TPUs, the 8t and 8i, at its Cloud Next event.

The push comes amid a compute shortage that has strained internal teams and led Google Cloud to turn away some business, according to the report. Cheaper serving costs could let Google undercut OpenAI and Anthropic on price, The Information said.

The chip's usefulness depends on Gemini's architecture staying broadly stable; a major architecture change could make it obsolete, according to the report. Deployment is not expected before 2028, and Google has yet to confirm the project publicly.

CNBC and The Decoder both covered the report.