Google Reportedly Designs a More Efficient Gemini Chip

Google may be trying to bind its AI models and hardware more closely together with a server chip designed specifically around Gemini.

What happened

Google is reportedly developing a new server processor internally called Frozen v2, according to The Information and subsequent Reuters reporting. The project is intended to run Gemini models more efficiently by incorporating elements of the model directly into the hardware.

The chip is reportedly targeting a possible 2028 introduction and could process substantially more tokens for each unit of electricity than Google’s current AI hardware. Development remains early, and Google has not confirmed that Frozen v2 will enter production. The company said it continually experiments with new hardware.

Why it matters

Running AI models at enormous scale is becoming an energy and cost problem. A processor optimised jointly with Gemini could reduce the amount of computing required for inference, lower Google’s operating costs and ease pressure on limited data-centre capacity.

It could also reduce dependence on Nvidia, although Google already designs its own Tensor Processing Units. Frozen v2 would deepen that vertical integration by tailoring the silicon even more closely to one model family.

The bigger picture

The AI infrastructure race is shifting from buying the fastest general-purpose accelerator to designing complete systems in which models, chips, networking and software are developed together. That approach can produce large efficiency gains, but it is expensive and inflexible: a chip designed years in advance risks being overtaken by changes in model architecture.

For now, Frozen v2 should be treated as a reported development programme, not a confirmed commercial product.

Source: Reuters