Google reportedly designing Frozen v2 chip with Gemini baked into silicon

AI chips may be about to change shape. The Information reported on July 20, 2026, that Google is developing “Frozen v2”, an inference processor with Gemini baked directly into the silicon.

What hardwiring a model means

According to the report, Frozen v2 hardwires parts of Gemini’s architecture into the circuitry itself. Engineers on the project estimate 6 to 10 times more token output per watt than Google’s latest TPUs. If that holds, it would be one of the largest single-generation efficiency jumps in AI hardware history.

The clever part is what gets frozen and what does not. The chip embeds the model’s architecture, not its weights. New Gemini versions can still be loaded, as long as Google keeps the same underlying blueprint. The tradeoff is equally real: change the architecture, and the chip is obsolete.

Born from a compute crunch

Context makes the project spicier. Google reportedly started it to relieve an internal compute shortage so severe that Google Cloud has been declining external customers. When a hyperscaler is turning away paying cloud business for lack of capacity, a 6 to 10x efficiency jump is not a research curiosity, it is an escape hatch.

Deployment is targeted for 2028, and Alphabet stock popped on the report. One caveat belongs in bold: Google has not officially confirmed the project.

The bet under the bet

Specialized silicon for one model family is a fascinating wager. It assumes Gemini’s core architecture is stable enough to be worth casting in hardware, in an industry where architectures have historically shifted every few years. If Google is right, it gets an efficiency moat no general-purpose chip can match. If the architecture moves on, Frozen v2 becomes a very expensive monument to a blueprint that did not last.

Genius efficiency play or costly bet on architectural stability? The receipts arrive in 2028.