Google is working on a new Frozen v2 chip for servers that would integrate part of Gemini’s architecture directly into the hardware. The project aims to run artificial intelligence models faster and reduce the energy consumption associated with processing.
The semiconductor is internally codenamed Frozen v2. Its potential rollout would begin in 2028, although the design is still under development. Engineers are also assessing how much of the model’s information would be physically embedded in the chip.
Frozen v2 chip would reduce data movement
Currently, artificial intelligence systems move large volumes of information between memory and processors. This movement consumes energy and can increase the time needed to generate a response. For this reason, Frozen v2 would incorporate certain elements of Gemini within the semiconductor itself. This configuration would make it possible to process requests with fewer data transfers and would reduce pressure on Google Cloud’s infrastructure.
According to the published information, the chip could deliver between six and ten times more efficiency than Google’s latest custom accelerators. The estimate is based on the number of artificial intelligence tokens processed per unit of energy. However, Google is keeping the project in an experimental phase. A Google Cloud spokesperson explained that its teams constantly test new technologies and noted that some developments may not reach production.
The new chip would operate alongside Google’s TPUs
Frozen v2 would be independent of Google’s tensor processing units; these TPUs are chips designed to train and run machine learning applications in data centers. Therefore, the new semiconductor would not directly replace TPUs; its role would be to complement the existing infrastructure with a design more closely tailored to Gemini’s characteristics.
This strategy reflects a growing trend among major technology companies: developers of AI models are seeking to control a larger share of their infrastructure to improve performance and reduce their dependence on external providers. Likewise, co-designing hardware and software makes it possible to tune each component to a specific workload. In Google’s case, that integration could enable more efficient execution of Gemini within its own services.
Google seeks to ease pressure on its computing capacity
The development of Frozen v2 would also be linked to the high demand for artificial intelligence computing capacity. The growth of Gemini and Google Cloud services requires more servers, energy, and specialized accelerators. According to the original report cited by Reuters, an internal capacity shortage would have created tensions and led Google Cloud to turn down some deals with external customers.
In this context, a chip optimized for Gemini could increase the number of queries processed with the same electrical infrastructure. It would also make it possible to handle more workloads without expanding data center energy consumption at the same pace.
Alphabet shares reacted to the report
The market responded positively to the information: Alphabet shares rose as much as 3.7% in New York after the semiconductor’s development became known. The reaction reflects investors’ interest in companies that design their own chips for artificial intelligence. This capability can reduce operating costs and improve the availability of computing resources.
Amazon, OpenAI, and other companies are also developing custom processors or maintaining agreements with semiconductor manufacturers. This competition is driving a new phase in which AI models and hardware are designed in a coordinated manner.
Frozen v2 still must clear the design phase
Google anticipates a possible introduction of the chip starting in 2028; however, the timeline will depend on technical results, manufacturing tests, and the final definition of the Gemini components that would be integrated into the hardware. Meanwhile, TPUs will continue to play a central role in the company’s artificial intelligence infrastructure. Frozen v2 would expand that platform with a specialized option to meet Gemini’s growing demand.
The project shows how Google is trying to improve the energy efficiency and speed of its models through closer integration between data centers, semiconductors, and artificial intelligence software.
Source: Reuters
Photo: Shutterstock