Google developing new chip to run Gemini AI model more efficiently – report

Google is developing a new AI chip called 'Frozen v2' that could run its Gemini AI model six to ten times more efficiently than current Tensor Processing Units (TPUs), according to a report from The Information. The chip, which embeds Gemini decision-making to reduce data movement, is expected to deploy in 2028 and will complement rather than replace Google's TPU line. Google's VP and GM of AI and infrastructure, Amin Vahdat, said in October 2025 that TPU demand is oversubscribed, with seven- and eight-year-old TPUs at 100% utilization.

Google is reportedly working on a new AI chip that would allow it to run its Gemini AI model more efficiently. According to a report from The Information https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently?rc=whojzp , the chip, known internally as ‘Frozen v2,’ would have Google Gemini decisions built in, thus reducing the number of steps necessary to move data around – decision-making is often the most time-consuming process for chips that are able to run multiple AI models. As a result, the report claimed that the new chip could be six to ten times more efficient than Google’s current Tensor Processing Unit TPU offerings and will be faster at responding to queries, potentially enabling the company to explore new AI applications. However, it was noted that Frozen is not intended to replace the TPU, but rather function as a separate branch of Google’s home-grown silicon efforts. Google plans to deploy the chips in 2028, The Information said, with Frozen only set to be usable for later generations of Gemini. The company also doesn’t intend to produce as many Frozen chips as TPUs, with sources cited in the report claiming this would allow Google to bring external design and manufacturing partners in on the project at a later date. In October 2025, Google’s VP and GM of AI and infrastructure, Amin Vahdat, said demand for the company’s TPUs was so oversubscribed https://www.datacenterdynamics.com/en/news/google-says-tpu-demand-is-outstripping-supply-claims-8yr-old-hardware-iterations-have-100-utilization/ that the company is having to turn customers away, adding that the company’s “seven and eight-year-old TPUs have 100% utilization.” Google released its eighth-generation TPUs https://www.datacenterdynamics.com/en/news/google-unveils-eighth-generation-tpus-two-dedicated-training-and-inference-chips/ in April 2026, with the latest generation offering two separate chips: one for inference workloads, and another for training. That same month, Broadcom stated in a regulatory filing https://investors.broadcom.com/node/64166/html that it had entered into a Long-Term Agreement with Google to develop and supply future generations of its TPUs, in addition to signing a Supply Assurance Agreement to supply “networking and other components” to be used in Google’s next-generation AI racks until 2031.