Google developing Gemini-dedicated AI chip alongside existing TPUs

Google is developing a dedicated AI chip optimized for its Gemini AI model as part of the company's strategy to maximize efficiency amid a deepening shortage of AI computing resources.

The Information, a US technology news outlet, reported Monday, citing multiple sources, that Google is developing a new AI chip internally codenamed "Frozen v2," targeting server deployment by 2028.

Frozen v2 is being developed separately from the tensor processing units (TPUs) Google has introduced to date.

While existing TPUs are designed to run a wide range of AI models — much like Nvidia's GPUs — Frozen v2 is purpose-built for Gemini.

The chip incorporates an idea from Google Chief Scientist Jeff Dean to pre-embed Gemini's data-processing methods and model weights directly into the hardware. The name "Frozen" derives from the concept of locking a Gemini-optimized architecture into the chip itself.

Google expects that running Gemini on the chip will increase the number of tokens processed per unit of power by six to 10 times compared with current levels, while also significantly improving response speeds.

However, Google has not yet decided how much Gemini-specific information to pre-embed in the chip as it seeks to balance flexibility against efficiency.

Google's push to develop a Gemini-dedicated chip stems from a worsening shortage of AI computing resources.

The company has reportedly faced internal friction over the scarcity of computing resources and has encountered constraints in contracts with external customers.

Google recently signed a contract to lease data centers from Elon Musk's SpaceX AI for $920 million per month and has also restricted usage by Meta, a competitor that uses Google's AI models.

Google is not planning to mass-produce Frozen v2 at scale, however, given the chip's limited flexibility.

Even within the Gemini family, differences in processing methods or model weights across versions could make the chip difficult to use.

In a statement, a Google spokesperson said the company "continues to research and experiment with new innovations to deliver maximum performance and efficiency to users and customers."

Shares of Alphabet, Google's parent company, climbed nearly 4% during trading after news of the custom AI chip broke, but gave back some of those gains in the afternoon. As of 1:20 p.m. Eastern time, the stock was trading around $353, up about 1.8% from the previous close.