Google is developing a new server chip intended to run its Gemini AI models with less power and greater efficiency. The processor, internally called Frozen v2, could enter service as early as 2028, although its design remains under development.
Frozen v2 could generate six to 10 times more tokens for each unit of power than Google’s existing AI chips, according to The Information. Alphabet shares rose about 3% after the report was published, ahead of the company’s next earnings release.
Chip Could Integrate Gemini Into Hardware
The planned processor would incorporate parts of Gemini’s architecture directly into the silicon. This could reduce the amount of data that must move between memory and processing components while the model generates responses.
Google did not confirm that Frozen v2 would enter production. The company said its teams regularly test new designs, but not every research project becomes a commercial product.
Google added that designing hardware and software together allows its systems to be optimised for specific workloads. The company already applies this approach to its Tensor Processing Units, or TPUs.
Its current Ironwood TPU was designed specifically for AI inference, the process used when trained models respond to user requests. Google has described Ironwood as its most powerful and energy-efficient TPU so far.
Technology Companies Develop Custom Chips
AI developers are producing their own processors to reduce computing costs and reliance on Nvidia. Custom hardware can be designed around a company’s models rather than supporting a wider range of software.
OpenAI announced an inference processor called Jalapeño in June, while Anthropic has reportedly discussed a possible chip partnership with Samsung. Google has developed several generations of TPUs for its internal services and Google Cloud.
Demand for AI computing has also placed pressure on available data-centre capacity. The reported Frozen v2 design is intended to improve the number of responses Google can generate without requiring a similar increase in electricity use.
Alphabet Faces Pressure Over AI Spending
Alphabet plans to spend between $180 billion and $190 billion on capital projects in 2026, with much of the money directed towards servers, data centres and networking infrastructure. The company raised the range from an earlier forecast of $175 billion to $185 billion.
Investors are watching whether revenue from Gemini and Google Cloud can justify those expenses. A processor that lowers inference costs could help Google serve more AI requests from the same amount of computing and electrical capacity.
Alphabet provides its financial reports through its investor relations website. Google has not announced a release date, technical specifications or production commitment for Frozen v2.
Featured image credits: Carlos Luna via Flickr
For more stories like it, click the +Follow button at the top of this page to follow us.