Google is developing a server chip that would etch part of its Gemini model architecture directly into silicon, targeting six to 10 times more tokens per watt.
Key Points:
- Engineers on the project estimate the processor, known internally as Frozen v2, could serve six to 10 times more tokens per unit of power than Google's newest Tensor Processing Units.
- Deployment is targeted for as soon as 2028, and the company has not confirmed the project exists.
- Alphabet shares rose about 3% on Monday, a day before the company reports second-quarter results.
Frozen v2 Chip Hardwires Gemini Into Silicon
News of the effort surfaced Monday, when two people with direct knowledge of the work described a processor known internally as Frozen v2. Engineers on the project estimate it could serve six to 10 times more tokens per unit of power than the newest generation of Google's Tensor Processing Units. Deployment is targeted for as soon as 2028, though the team has not settled on how much of the model the hardware will carry.
Google has neither confirmed nor denied that the project exists.
A spokesperson said its teams constantly research and experiment with new ideas, and that not every effort moves into production.
The chip would open a separate line rather than replace the Tensor units that already run Gemini and Google Cloud workloads. Production volumes are expected to fall well short of TPU levels. The company still treats the work as exploratory, and several design questions remain open.
Also Read: Dow Jones Industrial Average Volatility Tests Bitcoin's Decoupling Story
Gemini Architecture Trades Flexibility For Power
Standard AI accelerators load a model into memory, then shuttle its data back and forth to answer each query, a pattern that burns both power and time. Frozen v2 would instead lock Gemini's structural blueprint into the circuits, cutting the calculations and the data movement that every single response requires.
The model's weights, meaning the numbers produced during training, would remain updatable.
An earlier version of the project would have burned those weights into the hardware as well. Because the architecture stays fixed, though, the chip can support later Gemini releases only if Google keeps that foundation intact.
Alphabet Stock Reacts Before Q2 Earnings
Alphabet shares climbed about 3% on Monday, trading near $356 intraday before the gain faded the following session. Investors read the report as a partial answer to a capacity squeeze severe enough that Google Cloud has turned away some outside business.
The company reports second-quarter results on Wednesday after the U.S. market close, with analysts looking for earnings of $2.95 a share on revenue near $117 billion.
Alphabet lifted its 2026 capital spending guidance to as much as $190 billion earlier this year, and first-quarter revenue rose 22% to $109.9 billion on cloud growth of 63%. Google has since absorbed a delay to its next Gemini Pro release, while Chinese models now account for roughly 45% of token use at U.S. businesses.
Read Next: Meteorite That Hit A Home In New Jersey Fuels AI-Driven Hunt For Life