Google is reportedly developing a custom AI chip called Frozen v2 to run Gemini models more efficiently, signaling a shift toward model-specific hardware that could reshape the future of AI infrastructure.
 |
| Google's reported custom AI chip highlights a growing industry trend where hardware and AI models are designed together to improve speed, efficiency and scalability. Image: CH |
Tech Desk — July 21, 2026:
The biggest breakthroughs in artificial intelligence may no longer come from bigger models alone. They may come from the chips running them.
That is why Google's reported work on a new AI processor deserves attention. Instead of simply building faster hardware, the company is reportedly designing a chip around its Gemini models, bringing hardware and software closer together than before.
According to Reuters, citing The Information, Google is developing a server chip internally known as Frozen v2. The reported goal is to embed parts of the Gemini model directly into the hardware, making AI processing significantly more efficient.
It is a strategy that reflects a wider shift across the technology industry. For years, chipmakers produced processors that could handle almost any workload. Now, AI is pushing companies toward highly specialized hardware designed for specific tasks.
The idea is simple. If a chip already "knows" part of the model it will run, it does not have to perform every calculation from scratch. That reduces unnecessary computation, lowers power consumption and speeds up inference.
Those gains matter more than ever.
AI companies are facing soaring infrastructure costs as demand for generative AI continues to rise. Every chatbot response, image generation request or coding assistant consumes computing resources. At scale, even small efficiency improvements translate into billions of dollars in savings.
The reported numbers are therefore striking. Frozen v2 could process six to ten times more AI tokens per unit of power than Google's latest custom AI chips.
If achieved, that would represent more than an engineering milestone. It would become a competitive advantage.
Power efficiency is quickly becoming one of the most valuable currencies in artificial intelligence. Faster chips are important, but chips that deliver more performance while consuming less electricity may prove even more valuable as data centers expand worldwide.
The reported project also addresses an immediate business challenge.
Google Cloud has reportedly faced AI computing shortages that forced it to turn away some customer opportunities. Building a more efficient processor is not only about advancing technology. It is also about increasing capacity without endlessly adding more expensive hardware.
Interestingly, Frozen v2 is not expected to replace Google's Tensor Processing Units, or TPUs. Instead, it would work alongside them.
That suggests Google is building a layered AI infrastructure where different chips perform different jobs. General-purpose AI processors would continue handling flexible workloads, while highly specialized chips accelerate frequently used Gemini tasks.
This reflects a broader evolution across the semiconductor industry.
Companies are no longer competing only on model performance. They are competing on the efficiency of the entire AI stack, from silicon and networking to software and cloud infrastructure.
The reported 2028 deployment timeline also highlights another reality. Building advanced chips takes years. Engineers are reportedly still deciding how much of Gemini should be permanently embedded into the hardware.
That decision is crucial.
Hardwire too much, and future model updates become difficult. Hardwire too little, and the efficiency gains become less meaningful. Finding the right balance could determine whether the project succeeds.
The timing is equally notable.
The report follows recent news that Google's latest Gemini model was delayed after reportedly falling short of internal performance goals, particularly in coding. While software development continues, the company also appears to be investing heavily in the infrastructure that will power future AI systems.
That may ultimately prove just as important.
The next phase of the AI race is unlikely to be decided solely by who builds the smartest model. It will increasingly be shaped by who can run those models faster, cheaper and more efficiently.
If Google's reported Frozen v2 project delivers on its ambitions, it would signal that the future of artificial intelligence is not just about better algorithms. It is about designing the silicon beneath them with the same level of intelligence.