Google Is Developing Frozen v2 to Make Gemini 6-10 Times More Efficient

Google is planning a custom server chip for Gemini, and the move could give the company far more control over the hardware behind its AI products. According to The Information, the chip, called Frozen v2, is expected to make Gemini inference 6-10 times more efficient than Google's current AI chips, with a release window set for 2028.

The report says the efficiency estimate is based on tokens generated per unit of energy, and that it applies to inference rather than training. In other words, Frozen v2 is meant to help an already trained model handle user requests more efficiently.

Google is also reportedly pushing the design closer to the model itself by embedding part of Gemini's architecture directly into the physical chip. The idea is to remove unnecessary computation and reduce the constant movement of data between servers, which can drain resources in large data centers.

The project also has a strategic side. A custom chip would reduce Google's dependence on suppliers such as Nvidia, while helping the company control more of the hardware and software stack behind Gemini. Google says that joint hardware-and-software design from the ground up is key to maximum system performance.

The market reaction was immediate: after the leak, Alphabet shares rose 3%. That suggests investors are paying close attention not just to Gemini's progress, but also to how far Google can go in building its own infrastructure around it.

If Frozen v2 delivers on that promise, the bigger question is how much of Google's AI future it will bring in-house.

Would a custom chip that cuts Gemini's energy use by up to 10x reshape Google's AI strategy?