Google appears to be developing a custom semiconductor optimized for its artificial intelligence (AI) model "Gemini."
U.S. information technology (IT) outlet The Information reported on the 20th (local time), citing multiple sources, that Google is developing an AI chip with the internal code name "Frozen v2," targeting server deployment in 2028.
Frozen v2 is being developed separately from the tensor processing unit (TPU) that Google has developed and released. While the existing TPU, like Nvidia's graphics processing unit (GPU), has the versatility to handle various AI models, Frozen v2 pre-embeds Gemini's data processing method and some weights into the chip.
This is known to be an idea proposed by Jeff Dean, Google's chief scientist. The name "Frozen" also comes from the concept of fixing the chip's architecture to fit the Gemini model.
Google expects that running Gemini on Frozen v2 will be 6–10 times more efficient in processed tokens per unit of power, with faster response times.
Sources said Google has not yet decided how much information to pre-embed in "Frozen v2" to balance flexibility and efficiency.
Analysts say a shortage of AI compute resources is behind Google's push to develop a Gemini-dedicated chip. As demand for computing power needed for AI model development and service operations has surged, internal conflicts over resource allocation have arisen at Google, and constraints have reportedly emerged in some contracts with external customers.
In fact, Google recently signed a contract to lease data centers to Elon Musk's xAI, paying $920 million per month (about 1.36 trillion won), and it also limited usage by Meta, which uses Google's AI models.
However, Google is not expected to mass-produce "Frozen v2" at scale. Even within Gemini, if the model architecture and weights change by version, the utility of the chip could drop significantly.
A Google Spokesperson said in a statement, "We continue to research and experiment with new innovations to deliver maximum performance and efficiency to users and customers."