Google May be Building a New AI Chip With Gemini Baked Directly Into the Hardware

Google is reportedly developing a new server chip that would integrate elements of its Gemini AI model directly into the hardware, potentially allowing the company to run its artificial intelligence services more efficiently while addressing growing constraints on computing capacity.

According to The Information, the new chip is informally known as Frozen v2 and is being developed as part of Google's efforts to more tightly integrate its AI software and underlying hardware infrastructure.

Google reportedly plans to deploy the chip as early as 2028, although engineers are still finalising its design and determining how much information from its AI models will be hardwired into the chip.

Gemini Could Be Baked Into Google's AI Hardware

The most significant aspect of Frozen v2 is Google's reported plan to incorporate elements of its Gemini models directly into the hardware.

AI models typically run on general-purpose accelerators that must repeatedly access and process enormous amounts of model data. By hardwiring certain model information into specialised silicon, Google could potentially reduce some of the computing and memory requirements involved in serving Gemini responses to users.

The approach could make AI inference, the process of running a trained AI model to generate responses, considerably more efficient.

Frozen v2 could reportedly deliver six to 10 times greater efficiency than Google's latest custom AI chips when measured by the number of AI tokens served for each unit of power consumed.

If achieved, such efficiency gains could allow Google to serve significantly more Gemini users without requiring a proportional increase in electricity and computing infrastructure.

Google Faces AI Computing Capacity Crunch

The development comes as Google reportedly faces constraints on the amount of AI computing capacity available across its infrastructure. The capacity crunch has reportedly created internal tensions and even prompted Google Cloud to turn down some deals with external customers as the company balances demand for computing resources against the requirements of its own AI products.

The rapid expansion of generative AI has forced technology companies to spend heavily on data centres, processors and electricity to train and operate increasingly sophisticated models.

More efficient chips could therefore help Google increase its available AI capacity without relying exclusively on building additional data centres and deploying more conventional accelerators.

Frozen Chips Will Not Replace Google's TPUs

Google already develops its own Tensor Processing Units, or TPUs, which are specialised chips designed for artificial intelligence and machine learning workloads.

The Frozen project is reportedly intended to create a separate family of custom chips rather than replace Google's existing TPUs.

This could give Google multiple types of specialised hardware optimised for different AI workloads, with Frozen chips potentially designed specifically around serving the company's own Gemini models.

"Our teams are constantly researching and experimenting with new innovations. By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimised," a Google Cloud spokesperson said.

Google's Growing Push for Custom AI Silicon

The project highlights how the AI race is increasingly moving beyond simply developing the most capable models.

Companies including Google are also attempting to control more of the infrastructure required to run those models, potentially reducing their dependence on third-party chipmakers while lowering the enormous costs associated with AI computing.

Designing hardware specifically around Gemini could give Google another advantage in this area, particularly as the number of people using AI services continues to grow.

The development also comes amid challenges surrounding Google's next-generation Gemini model. The company reportedly delayed the launch of its latest Gemini AI model after it failed to meet internal performance targets, with engineers working to improve its capabilities, particularly in coding.

While that work focuses on making Gemini more capable, the Frozen project appears to tackle the other side of the equation: making increasingly powerful AI models cheaper and more efficient to operate at massive scale.