Google is etching Gemini's architecture into a new chip

Google is developing a chip, Frozen v2, that hardwires parts of Gemini's neural architecture directly into the silicon, per Tom's Hardware, via The Information. Engineers are projecting 6 to 10 times more tokens per watt than current TPUs deliver, with a target as early as 2028.

The stakes are about capacity, not novelty. Google Cloud has been turning away customers because it doesn't have enough compute to serve them, per the report. Google frames Frozen v2 as a new chip family alongside TPUs, not a replacement. The idea reportedly came from Jeff Dean.

Here's the part that should give builders pause. Etching an architecture into circuitry only pays off if that architecture holds still. General-purpose TPUs can adapt as Google changes how Gemini is built. A chip with Gemini's structure baked in can't. Google is making a real bet that transformer-style architecture won't meaningfully change between now and 2028, not hedging on it.

If that bet is wrong, Frozen v2 becomes a fast, efficient chip built for an architecture Google already left behind. The report doesn't say what happens to the silicon if Gemini's design changes before then, only that the project assumes it won't.

What to watch: whether Google says anything more specific about which parts of Gemini's architecture get fixed versus which stay flexible. The report doesn't say how much of the design is actually frozen versus how much room is left to adapt. That ratio is the whole bet.