Google is reportedly building a 'Frozen v2' chip that bakes Gemini into silicon

Google is developing a new AI inference chip, internally called “Frozen v2,” that hardwires portions of its Gemini model architecture directly into silicon, according to reporting published July 20, 2026. Google has not confirmed the project.

The design would still let engineers refresh model weights, but fixing parts of the architecture in hardware could reportedly deliver six to ten times more AI tokens per unit of power than Google’s current tensor processing units, or TPUs. Deployment is said to be targeted around 2028.

The motivation, per the reporting, is an internal compute crunch. Google Cloud has reportedly been turning away outside business and paying SpaceX nearly $1 billion a month for capacity, an unusual arrangement for a company that runs some of the world’s largest data centers. A chip tuned narrowly for Gemini inference could stretch each watt further at a moment when power, not silicon supply alone, is the binding constraint on AI expansion.

There are limits to the ambition. Frozen v2 is described as a complement to TPUs rather than a replacement, and it would reportedly stay useful only if Google keeps the same underlying Gemini architecture across future models, a real constraint given how fast frontier designs change. People describing the effort framed it as a trial run rather than a TPU-scale production program.

The specifics are unconfirmed, and Google has not commented on the chip. Alphabet shares rose about 3 percent after the report, a sign investors read even an unconfirmed efficiency edge as material. The test is whether Frozen v2 reaches deployment near its 2028 target, or stays a hedged experiment that never leaves the lab.