Google is reportedly developing a new server chip designed to run its Gemini artificial intelligence models more efficiently. Codenamed “Frozen v2,” the proposed chip may embed selected elements of an AI model directly into hardware, potentially reducing the computing power and energy required to generate AI responses.
According to a report by The Information, Google could begin deploying the chip as early as 2028. The design is reportedly still under development, and engineers have not yet finalised how much of the Gemini model’s functionality would be hardwired into the processor.
Although the technology remains at an experimental stage, Frozen v2 offers an important glimpse into how the next generation of AI systems may be built.
Moving AI Closer to the Hardware
Most modern AI models rely on a combination of software and specialised processors. The model is stored in memory, while chips perform the calculations required to process instructions and generate outputs.
Frozen v2 could change this relationship by placing certain stable components of Gemini directly into the chip.
This approach is similar to designing a machine for one highly specialised purpose rather than asking a general-purpose machine to perform every task. By reducing the movement of data between memory and processors, Google may be able to improve speed, lower power consumption, and serve more users with the same computing capacity.
The reported chip could be between six and ten times more efficient than Google’s latest custom AI processors when measured by the number of AI tokens produced per unit of power.
In generative AI, a token is a small unit of information, such as part of a word, a number, or a punctuation mark. Improving the number of tokens generated for every unit of energy could make AI services faster, more affordable, and more environmentally sustainable.
Frozen v2 Would Complement Google’s TPUs
Google has spent years developing Tensor Processing Units, or TPUs, for machine learning workloads. These processors power many of the company’s internal AI systems and are also available to customers through Google Cloud.
Frozen v2 is not expected to replace TPUs. Instead, it could complement them.
TPUs are designed to support a wide range of machine learning tasks. Frozen v2 appears to be more closely connected to the architecture of specific Gemini models. Google could therefore use different processors for different workloads, depending on the level of flexibility, performance, and efficiency required.
This reflects a broader shift in the technology industry. AI development is no longer focused only on creating larger models. Companies are also exploring how software, memory, networking, and processors can be designed together as one integrated system.
A Google Cloud spokesperson described this principle as co-designing hardware and software from the ground up so that systems can be highly integrated and optimised.
Why AI Companies Need More Efficient Chips
The rapid growth of generative AI has created enormous demand for data centres, advanced processors, electricity, and cooling systems.
Every time a user interacts with an AI chatbot, the request must be processed by powerful computing infrastructure. When millions of people use these services simultaneously, the cost becomes significant.
Limited computing capacity can also affect cloud businesses. The report suggests that capacity constraints have previously caused Google Cloud to avoid certain agreements with external customers. More efficient processors could help Google support its own AI products while making additional infrastructure available to businesses, developers, schools, and research organisations.
Efficiency is therefore not only a technical advantage. It can influence the availability, affordability, and accessibility of AI.
The Trade-Off Between Efficiency and Flexibility
Embedding elements of an AI model into hardware could deliver substantial performance improvements, but it may also introduce limitations.
Software can usually be updated quickly. Hardware is more difficult and expensive to change. If too much of a model is permanently built into a chip, the processor could become less useful when the model evolves.
Google’s engineers will therefore need to decide which components of Gemini are stable enough to be hardwired and which should remain programmable.
This is a classic computational thinking problem involving decomposition, abstraction, optimisation, and trade-offs.
Engineers must break the AI system into smaller components, identify the elements that change frequently, and determine which operations can be accelerated without reducing adaptability. The best design will not simply be the fastest chip. It will be the design that creates the most effective balance between efficiency and flexibility.
A New Phase of AI Innovation
The reported Frozen v2 project represents a possible new phase in the AI race.
Until recently, much of the competition centred on model size, training data, and capabilities. The next stage may be shaped just as strongly by infrastructure efficiency. Companies capable of building highly optimised combinations of chips, software, data centres, and AI models may gain a significant advantage.
However, Frozen v2 is still under development. Its final design, performance, deployment schedule, and commercial availability could change before the reported 2028 launch target.
Even so, the idea behind the project is significant. The future of artificial intelligence may not depend solely on smarter software. It may also depend on specialised hardware designed around the way individual AI models think and operate.
For educators, this is an important reminder: AI education must go beyond tools and prompts. Students need computational thinking skills that help them understand how complex systems are designed, optimised, evaluated, and improved.
Frozen v2 may eventually become part of Google’s AI infrastructure. The larger lesson, however, is already clear—the future of AI will be built through the close integration of human ideas, intelligent software, and increasingly specialised machines.