Google builds chip to run Gemini AI with 10× efficiency

Google is developing a new custom server chip designed to run its Gemini AI model with dramatically greater efficiency, according to multiple reports. The project, informally called "Frozen v2," aims to integrate the neural network architecture of Gemini directly into the hardware, potentially making the chip six to ten times more efficient than the company's current custom AI processors. The company has not publicly confirmed the initiative, which remains in development with a targeted deployment as early as 2028.

The reported design represents a significant shift from general-purpose AI chips. Instead of loading a model into memory, the Frozen v2 chip would have Gemini's fundamental blueprint etched into its circuitry, a process described as "hardwiring" or "freezing" the architecture. This approach is intended to drastically reduce the power and time required for processing AI queries by eliminating much of the data shuttling overhead. Engineers are reportedly still finalizing the design, including determining how much of the model's information will be permanently embedded, though the plan allows for the model's "weights" or learned parameters to be updated via software.