What Google's "Frozen v2" AI chip means for you

Google is reportedly developing a new type of computer chip designed specifically to run its Gemini AI models more efficiently. The Information reported on 20 July 2026, citing two people familiar with the project, that the chip is internally called "Frozen v2". Google has not confirmed the report, though Alphabet's shares closed 1.51% higher the day the story broke.

Normal AI chips are general-purpose: they can run many different AI models. Frozen v2 takes a different approach. According to the report, it builds parts of Gemini's underlying architecture — the structural design that shapes how the model processes information — directly into the chip itself, rather than just storing the model's learned data on it. The idea, described by The Decoder, is that this cuts down on the calculations and data-shuffling needed to answer a query, so each chip can do more with less power.

The name refers to "freezing" parts of the model in place. The concept reportedly came from Google engineer Jeff Dean. An earlier version of the idea went further, building the model's actual learned weights into the chip, but Google dropped that approach because such a chip would only work with one specific version of Gemini and become outdated too quickly.

Engineers estimate Frozen v2 could serve six to ten times more tokens (chunks of text an AI model reads or writes) per unit of power than Google's newest TPUs, its current AI chips. That figure is still an internal estimate, and the design is reportedly unfinished. The chip would supplement, not replace, Google's general-purpose TPU line, which already includes separate chips for training and for answering queries, as reported by CNBC.

If the reports pan out, cheaper, more efficient chips could eventually help Google offer AI services at a lower price, because a compute shortage — a shortfall of processing power — has reportedly strained Google's teams internally and forced Google Cloud to turn away some business. Lower running costs could let Google price Gemini more aggressively against rivals such as OpenAI and Anthropic. None of this is guaranteed yet, and the notes don't say whether any savings would reach consumers directly.

There's a catch built into the design: because Frozen v2 has Gemini's architecture baked in, it would only keep working well if future Gemini models keep that same underlying structure. A major redesign of Gemini could make the chip obsolete.

Google is reportedly planning to deploy Frozen v2 from 2028, initially as a trial run in smaller volumes than its mainstream TPU chips. Nothing about the project is confirmed by Google, and the design is still being developed.