Google embeds Gemini architecture directly into silicon: details about the Frozen v2 chip
Google is preparing to release a fundamentally new specialized processor under the codename Frozen v2. This is not just another tensor accelerator — it is a chip that physically integrates key elements of the Gemini neural network architecture directly into the silicon substrate. This approach radically reduces the volume of data transferred and the number of computational operations required to process user requests.
The project is not intended as a replacement for the existing line of tensor processing units (TPUs), but rather as a highly specialized complement to them. According to Google engineers' estimates, the energy efficiency of Frozen v2 will be 6–10 times higher compared to the company's latest generations of AI accelerators — meaning the chip will be able to process significantly more tokens per unit of energy consumed. This is particularly relevant amid an acute shortage of computing power, which, according to available data, has forced Google Cloud to turn down some contracts with external clients.
Experimental Status and Architecture Dependency
It is worth emphasizing that Frozen v2 is currently considered an experimental project. The company does not plan to produce it in the same volumes as universal TPUs. Moreover, the chip will only work effectively with future versions of Gemini models, provided the basic architecture is maintained. If Google decides to radically change its approach to building neural networks, the investment in silicon could be called into question. The chip is scheduled to be deployed in 2028.
Interestingly, after this information emerged, Alphabet shares (Google's parent company) on the Nasdaq rose by 1.5% — the market clearly views the long-term plans for hardware AI optimization positively.
Context: Challenges for Google's AI Division
Recent months have been difficult for Google. The launch of the next release, Gemini 3.5 Pro, has been delayed, and the company has lost four leading researchers who moved to competitors — Anthropic and OpenAI. At the same time, Chinese models continue to strengthen their positions, now accounting for up to 46% of tokens processed by American companies. Against this backdrop, the initiative to create a super-efficient specialized chip looks not only like a technological move but also a strategic one — an attempt to regain leadership in the performance race.
Expert opinion. Frozen v2 is a signal that Google is betting on deep integration of software and hardware, essentially turning the neural network into a physical object. If the project proves successful, it could radically change the economics of AI computing. However, 2028 is a long way off, and competitors are not standing still. The question is whether Google can maintain its technological advantage until the chip actually reaches the market.