Summary
- Google is in the process of creating a chip named Frozen v2, which integrates elements of Gemini's architecture directly into silicon.
- Engineers estimate that this new chip could achieve six to ten times the efficiency of current TPUs, with a projected rollout in 2028.
- Following the announcement, Alphabet's stock rose approximately 3% during Monday's trading session, just ahead of the Q2 2026 earnings report set for Wednesday, July 22.
Google is working on a specialized chip aimed at enhancing the performance and cost-effectiveness of its Gemini AI system.
The chip, referred to as Frozen v2, was reported by The Information on Monday, as Google faces a pressing challenge: a shortage of capacity to meet the surging demand for its AI services.
In March, Google informed Meta that it could not accommodate the volume of computing power Meta sought for Gemini, prompting Meta to advise its employees to limit their AI usage. Despite investing up to $190 billion in AI infrastructure this year, Google has had to turn away customers due to insufficient server capacity.
Consequently, the tech giant is developing a chip specifically for its AI models.
Details on Frozen v2 are limited, but it is not simply an enhancement of Google's existing Tensor Processing Units (TPUs)—the custom chips that have been in use since 2015 to power Gemini and support external developers' cloud services.
Unlike these Tensor chips, which can run any AI model, Frozen v2 integrates part of Gemini's architecture—the foundational design that dictates how the model processes and routes information—directly into the chip.
In the context of machine learning, "freezing" refers to permanently securing a structure. Here, it is the architecture that is frozen, not the model's weights (the knowledge acquired during training, which remains updatable). By embedding this blueprint into the chip's circuitry, it eliminates unnecessary calculations and reduces the need for data transfer across memory for every query. Engineers anticipate a six to tenfold increase in tokens—the small text segments that compose each AI response—generated per watt of energy consumed.
This improvement means Google could handle ten queries using the energy costs typically associated with just one.
For users of Gemini, the experience will remain unchanged; however, the operational costs will decrease, allowing a more competitive edge against OpenAI, Anthropic, and Chinese labs that currently dominate up to 45% of AI token usage in the U.S., primarily due to their significantly lower operational costs—60% to 90% cheaper. While AI services may not become cheaper for consumers, Google is likely to see enhanced profitability.
Alphabet's stock increased by about 3% on Monday, reaching an intraday high of $356. However, as investors await the company's latest earnings report on July 22, this uptick has since leveled off.
This initiative represents another strategic move by a leading AI firm to reduce its dependency on Nvidia hardware for product development. Nvidia currently holds approximately 85% of the AI GPU market, and major tech companies are eager to lessen this reliance.
Nvidia's hardware was initially designed for gaming, not specifically for language models, and while it is functional, it incurs overhead that dedicated chips do not. For a company like Google, a 6-10x efficiency difference translates into billions of dollars. Other tech giants, including Meta, Amazon, Microsoft, and OpenAI, are also pursuing custom silicon for this reason.
As previously reported by Decrypt in March, even AWS, which has committed to deploying 1 million Nvidia GPUs by 2027, is simultaneously developing its own chips to mitigate long-term risks associated with this reliance.
Frozen v2 is still in the exploratory phase. Major design choices have yet to be finalized, and Google has not officially acknowledged the project's existence. Additionally, this chip will not be available for external cloud customers, as it is tailored specifically for a single model. The earliest deployment is anticipated in 2028, according to reports.
In the interim, Google is paying SpaceX $920 million each month to lease 110,000 Nvidia GPUs from xAI's data centers as a temporary solution.