Google AI chip project Frozen v2 could make the company’s Gemini models significantly more efficient by incorporating parts of the AI system directly into the hardware.
The project, informally known as “Frozen v2,” could help Google respond to the growing pressure on its AI computing infrastructure. Demand has reportedly become so high that Google Cloud has had to turn down business from outside customers because it does not have enough available capacity.
The new chip is still under development and may not be deployed before 2028, but its design shows how closely Google is beginning to connect its AI models with the hardware running them.
They are not alone in developing custom hardware, as OpenAI is also working with Broadcom to build its own AI chip.
Google AI Chip Could Be Up to 10 Times More Efficient
According to a report by The Information, the new Google Gemini chip could be between six and 10 times more efficient than the company’s latest custom AI chips.
That estimate is based on how many AI tokens the chip could process for each unit of electricity it consumes. Tokens are the small pieces of text and data that AI models process when answering questions or completing other tasks.
Improving this ratio matters because running models such as Gemini for millions of users requires enormous amounts of computing power and electricity. Even a modest efficiency improvement could reduce costs and allow Google to handle more requests using the same amount of energy.
A six-to-10-fold improvement would be far more significant.
Parts of Gemini Could Be Built Directly Into the Hardware
The most unusual part of the project is Google’s plan to hardwire certain elements of Gemini into the chip itself.
Engineers are still deciding how much information from the model should be placed directly into the hardware. The final design has not been completed, and details could change before the chip enters production.
Building parts of Gemini into the chip could make the system faster and more efficient because the hardware would be designed specifically around the way Google’s AI models operate.
“Our teams are constantly researching and experimenting with new innovations,” a Google Cloud spokesperson said, adding that designing hardware and software together allows the company to create more integrated and optimized systems.
Frozen v2 Would Not Replace Google’s TPUs
Google has been building its own AI processors for years. Its tensor processing units, better known as TPUs, are already used to train and run artificial intelligence models across the company’s services and cloud infrastructure.
The Google AI chip would create a separate family of internally developed processors rather than replace the company’s existing TPUs.
This would give Google another type of processor optimized for serving Gemini to users. It could also reduce some of the pressure created by the rapid growth of generative AI services.
The company has not announced when Frozen v2 will enter production, but deployment could begin as early as 2028.
Google Is Facing an AI Computing Capacity Crunch
The project comes as Google struggles to secure enough computing capacity to meet demand for AI.
According to The Information, the shortage has created tensions inside the company and forced Google Cloud to reject some potential deals with external customers. That suggests Google must balance the computing needs of its own Gemini products with demand from businesses that want to use its cloud infrastructure.
Alphabet shares rose 3.3% in early trading following the report, reflecting investor interest in technology that could improve Google’s position in the AI infrastructure race.
Efficiency has become increasingly important as technology companies invest heavily in chips, data centers and electricity. Building larger facilities is one solution, but designing hardware that can produce more AI output using the same resources could be just as valuable.
Why Google’s New Gemini Chip Matters
Google’s plan highlights a broader change in the AI industry. Competition is no longer focused only on building the most capable model. Companies must also find affordable and energy-efficient ways to make those models available to millions of people.
By developing hardware specifically around Gemini, Google could gain more control over the cost and performance of its AI services.
The project also shows how the line between AI software and hardware is beginning to disappear. Instead of designing a model first and choosing the chips later, Google is attempting to develop both parts as one integrated system.
However, Frozen v2 remains several years away, and Google is still working on its design. The company must also improve Gemini itself after reportedly delaying its latest model because it failed to meet internal targets, particularly in coding.
If successful, the Google AI chip could make Gemini faster, cheaper and considerably more energy-efficient.
Source: Reuters, citing reporting by The Information.
