Google is reportedly developing a new artificial intelligence chip designed specifically for its Gemini models by embedding parts of the model's architecture directly into the hardware, according to a report by The Information. The project, internally codenamed "Frozen v2," could reportedly be deployed as early as 2028.
Unlike general-purpose AI accelerators such as Google's Tensor Processing Units (TPUs), the proposed chip is reportedly being designed around specific elements of Gemini's architecture. According to The Information, this approach would allow portions of the hardware to be dedicated exclusively to running Gemini models rather than supporting a wide range of AI systems.
The report, citing two people familiar with the project, said the chip could deliver six to ten times more AI tokens per unit of power than Google's current custom AI chips. While the underlying hardware architecture would remain fixed after manufacturing, Google would reportedly still be able to update Gemini's model weights.
Because the chip is tailored to Gemini's architecture, its effectiveness would depend on Google continuing to use a compatible model design in future versions. The report said Google views Frozen v2 as a complementary product rather than a replacement for its TPU lineup, with expected production volumes significantly lower than those of its existing AI chips.
The project comes as Google reportedly faces growing demand for AI computing resources. According to The Information, compute constraints have affected the company's cloud business, prompting investments in additional infrastructure. Google has not officially confirmed the existence of Frozen v2.
A company spokesperson told The Information that Google regularly explores new approaches to improving AI hardware efficiency, adding that not all experimental projects ultimately reach production.
The reported project comes amid intensifying competition in the AI sector, as major technology companies continue investing in custom AI chips to improve performance and reduce operating costs. Companies including Nvidia, AMD, Amazon, Microsoft and Meta are also developing specialised AI hardware for large language models. If deployed, Frozen v2 would represent a further move toward AI chips designed for specific model architectures rather than general-purpose AI workloads.