Google Is Building an AI Chip for the Gemini Bus—And Retailers Are Already Getting Started

Google is developing a chip designed for one task: running Gemini quickly and cheaply.

The chip, called Frozen v2, was by The Information On Monday, Google offered a potential solution to a problem that won’t go away any time soon: It’s running out of power to serve what AI has already created.

In March, Google told Meta that it could not fulfill the amount of Gemini compute Meta wanted to buy. Meta had to instruct employees to share the use of AI. Google — spending up to $190 billion on AI infrastructure this year — was turning customers away because it didn’t have enough servers to serve them.

So now it’s building a chip designed just for its AI models.

There is no information about this new device, but the conference mention is not another change of Google Tensor Processing Units (TPUs)—the chips that Google has been making since 2015 That strength of Gemini is its Cloud services for external developers.

Tensor chips run any type of AI loaded on them. Frozen v2 does the opposite. According to reports, it builds part of the Gemini architecture – the planets that determine the model’s behavior and transfer information – directly into the hardware.

In machine learning, “freezing” means shutting down something forever. Here, what is cool is the construction, not the weight of the model (the real knowledge Gemini takes through training, which is flexible). By incorporating this design into the chip’s circuits, the chip skips multi-threading and stops storing data in memory for each query. Token engineers manage six to ten times the amount of electricity — the tiny particles that make up any AI solution — generated for every watt of electricity used.

That’s the difference between Google sending ten queries for the price of one email.

If you use Gemini, Frozen v2 will not change your experience. But it changes what costs to run – and Gemini’s low-cost processor competes closely with OpenAI, Anthropic, and Chinese labs that already account for up to 45% of the US company’s AI applications, mainly because they run 60-90% cheaper. You may not have cheap AI, but Google can be very profitable.

Shares of the label rose nearly 3% during Monday’s news, touching $356 intraday. The company reported its Q2 2026 earnings on Wednesday, July 22, and the pump dropped on today’s stock as investors await Google’s latest results.

This is yet another attempt by the AI ​​giant to kill the heavy reliance on Nvidia hardware to create its products. Nvidia controls about 85% of the AI ​​GPU market, and every major tech company wants it.

Nvidia hardware is designed for video games, not languages ​​- they work, with themes that purpose-built chips don’t handle. At Google’s level, the 6-10x difference is barely noticeable. It’s billions of dollars. Meta, Amazon, Microsoft, and OpenAI all have silicon software for the same purpose.

As Decrypt reported in Marcheven AWS—which has committed to shipping 1 million Nvidia GPUs through 2027—is manufacturing its chips at the same time to reduce exposure in the long run.

Frozen v2 is still a research project. The main options are not finalized, Google has not confirmed that the service is available, and the chip will not be offered to external customers of the Cloud-hardware of one type cannot run anyone else’s. Delivery is expected in 2028 at the earliest, according to reports.

Meanwhile, Google is paying SpaceX $920 million a month to rent 110,000 Nvidia GPUs from the xAI data center as a bridge.