Google is reportedly developing a new generation of custom artificial intelligence (AI) server chips designed to run its Gemini AI models more efficiently by embedding key elements of the models directly into the hardware. According to a report by The Information, the initiative is aimed at improving AI computing efficiency, reducing energy consumption and addressing the growing demand for AI infrastructure as generative AI adoption continues to accelerate.
The new chip, internally referred to as “Frozen v2,” is expected to become part of Google’s long-term AI infrastructure strategy. Unlike conventional AI accelerators that primarily execute software models, the proposed chip would integrate portions of Google’s Gemini model architecture into the hardware itself, allowing AI workloads to be processed with greater speed and significantly improved energy efficiency.
According to the report, Google plans to deploy the new chip as early as 2028, although engineers are still finalising its architecture and determining how much of the Gemini model should be permanently embedded into the chip’s design.
The project comes as technology companies face an unprecedented surge in demand for AI computing capacity. Rapid growth in generative AI applications has increased pressure on cloud infrastructure worldwide, creating shortages of high-performance AI processors and data centre resources.
The report suggests that Google’s AI infrastructure has experienced similar constraints, with growing demand reportedly contributing to capacity limitations within Google Cloud. According to the publication, these shortages have at times affected the company’s ability to accommodate certain enterprise cloud customers while balancing internal AI development priorities.
One of the primary objectives of the Frozen v2 project is to improve computational efficiency by increasing the number of AI tokens that can be generated for every unit of electricity consumed. The report estimates that the new chip could be six to ten times more efficient than Google’s current generation of custom AI processors when measured by AI token output per unit of power.
Improving power efficiency has become one of the industry’s biggest priorities as artificial intelligence models continue to grow in size and complexity. Large Language Models (LLMs) require enormous computing resources during both training and inference, resulting in higher operational costs and increased energy consumption for cloud providers.
A spokesperson for Google Cloud acknowledged the company’s continued investment in AI infrastructure innovation, stating that Google regularly researches and experiments with new technologies to optimise the performance of its systems.
The spokesperson noted that Google follows a co-design approach, developing both hardware and software together to ensure that AI systems are tightly integrated and highly optimised for performance, scalability and efficiency.
According to the report, the Frozen project is not intended to replace Google’s existing Tensor Processing Units (TPUs), which have powered many of the company’s AI services for years. Instead, the new chips are expected to complement Google’s current AI hardware portfolio by serving specialised workloads that require even greater efficiency.
Google has invested heavily in custom silicon over the past decade, designing processors for artificial intelligence, cloud computing and mobile devices rather than relying entirely on third-party semiconductor manufacturers. The company’s TPU architecture has become a cornerstone of Google’s AI ecosystem, supporting products including Search, Google Cloud and Gemini.
The reported development also comes amid intense global competition among major technology companies to build more powerful and energy-efficient AI infrastructure. Firms including NVIDIA, AMD, Microsoft, Amazon, Meta and OpenAI continue to invest billions of dollars in AI chips, hyperscale data centres and specialised computing hardware to meet rising demand for generative AI services.
Separately, reports last week indicated that Google had delayed the release of its latest Gemini AI model after internal evaluations reportedly found that the model had not yet achieved desired performance benchmarks, particularly in coding-related capabilities. The company is said to be refining the model before its wider rollout.
As artificial intelligence becomes increasingly central to cloud computing, enterprise software and consumer applications, advances in custom AI chips are expected to play a crucial role in improving performance while reducing operational costs and energy usage.
If successful, Google’s next-generation AI chip could significantly strengthen the company’s ability to deploy Gemini models more efficiently across its products and cloud platform, while supporting its long-term strategy of building an integrated AI ecosystem combining advanced software with purpose-built hardware.
Disclaimer: This report has been editorially prepared using publicly available information and media reports. Readers are advised to refer to official statements from Google for confirmed details regarding future product development.
Author
About the Author
