Google is developing a custom server chip called “Frozen v2” that will permanently embed parts of its Gemini AI architecture directly into silicon, a design approach that engineers project could deliver six to ten times more tokens per unit of power than the company’s current tensor processing units.
According to a report from The Information, the chip is designed as a specialized addition to Google’s custom silicon portfolio rather than a replacement for its general-purpose TPUs. By hardwiring Gemini’s architecture into the chip itself, Frozen v2 reduces the number of calculations and data movement required to process AI queries.
The project targets 2028 for deployment and is driven by a critical internal compute shortage that has forced Google Cloud to turn down outside business. Last month, Google agreed to pay SpaceX nearly $1 billion a month for compute capacity to bridge the gap.
The trade-off is reduced flexibility — Frozen v2 will work with future Gemini models only if Google maintains the same underlying architecture. The company reportedly views the project partly as a trial run and does not plan to produce it at TPU scale. Alphabet shares rose 1.51% on the news.
Google’s broader AI efforts face headwinds including a delayed Gemini Pro release, senior researcher departures, and Chinese AI models now accounting for 45% of U.S. enterprise token usage.
