Google embeds Gemini architecture into silicon: Frozen v2 chip announced
Google is betting on hardware optimization for artificial intelligence. The company has announced the development of a specialized chip, Frozen v2, which will be a significant step in the evolution of its AI infrastructure. The main innovation is the integration of key elements of the Gemini neural network architecture directly into the processor's silicon substrate.
This approach allows for a drastic reduction in computation and data movement when processing requests. Instead of shuttling information between memory and computing units, the chip will perform some operations "in place," significantly speeding up inference and reducing power consumption. According to Google engineers' estimates, Frozen v2 will be able to process six to ten times more tokens per unit of energy compared to the company's current AI accelerators.
It is important to understand: Frozen v2 does not replace the line of tensor processing units (TPUs) but complements it. The project is positioned as a solution to the acute shortage of computing power, which, according to available information, has forced Google Cloud to turn down some contracts with external clients. However, the chip will only be able to work with future versions of Gemini if the basic architecture of the models remains unchanged. For now, Google views Frozen v2 as an experimental project and does not plan mass production comparable to TPUs. Commissioning is scheduled for 2028.
Against the backdrop of this news, Alphabet (GOOG) shares rose by 1.5% on July 20, reflecting the market's positive reaction to the company's long-term strategy.
However, it is worth noting that Google's AI division is going through a challenging period. The launch of the next version, Gemini 3.5 Pro, is delayed, and the company has lost four leading researchers who moved to competitors — Anthropic and OpenAI. Moreover, Chinese models have already captured up to 46% of tokens among American companies, creating additional pressure.
Expert opinion: Frozen v2 is not just another chip, but an attempt by Google to close the "software-hardware" loop, much like Apple does with the Neural Engine or Tesla with Dojo. If the project proves successful, Google will gain a unique competitive advantage in inference efficiency. But 2028 is too far off for the rapidly changing AI market. By then, the Gemini architecture may undergo radical changes, and the chip tailored to it risks becoming obsolete even before its release.