Google Is Building an AI Chip Just for Gemini—And Investors Already Moved On It

Google is constructing a chip designed for one job: working Gemini quicker and cheaper.

The chip, codenamed Frozen v2, was reported by The Information on Monday and provides Google a possible reply to an issue it may well’t spend its method out of quick sufficient: It’s working out of capability to serve the AI demand it has already generated.

In March, Google instructed Meta it could not fill the quantity of Gemini compute Meta needed to buy. Meta needed to instruct workers to ration their AI utilization. Google—spending as much as $190 billion on AI infrastructure this 12 months—was turning away clients as a result of it did not have sufficient servers to serve them.

So now it is constructing a chip designed just for its AI fashions.

There’s not a lot details about this new chip, however by naming conference it’s not one other improve to Google’s Tensor Processing Models (TPUs)—the custom chips Google has been building since 2015 that energy Gemini and its Cloud companies for out of doors builders.

These Tensor chips run any AI mannequin loaded onto them. Frozen v2 does one thing completely different. Per the experiences, it bakes a part of Gemini’s structure—the structural blueprint that determines how the mannequin routes and processes info—straight into the {hardware}.

In machine studying, “freezing” means locking one thing completely in place. Right here, what will get frozen is the structure, not the mannequin’s weights (the precise information Gemini picks up by way of coaching, which stays updatable). By hardwiring this blueprint into the chip’s circuits, the chip skips redundant calculations and stops shuttling information throughout reminiscence on each question. Engineers challenge a six to 10 instances enchancment in tokens—the small textual content chunks that make up every AI response—generated per watt of electrical energy consumed.

That is the distinction between Google serving ten queries for the facility value of 1.

For those who use Gemini, Frozen v2 will not change the way it feels to you. Nevertheless it adjustments what it prices to run—and a cheaper-to-run Gemini competes tougher towards OpenAI, Anthropic, and Chinese language labs that already account for as much as 45% of U.S. firm AI token utilization, largely as a result of they run 60–90% cheaper. It’s possible you’ll not have cheaper AI, however Google will possible be extra worthwhile.

Alphabet shares climbed roughly 3% throughout Monday’s session on the information, touching $356 intraday. The corporate experiences Q2 2026 earnings on Wednesday, July 22, and the pump receded in in the present day’s session as traders await Google’s most up-to-date outcomes..

That is yet one more effort by a significant AI firm to kill its over-reliance on Nvidia {hardware} to develop its merchandise. Nvidia controls roughly 85% of the GPU marketplace for AI, and each main tech firm desires out.

Nvidia’s {hardware} was initially constructed for video video games, not language fashions—it really works, simply with overhead that purpose-built chips do not carry. At Google’s scale, a 6–10x effectivity hole is not summary. It is billions of {dollars}. Meta, Amazon, Microsoft, and OpenAI all have customized silicon packages for precisely that cause.

As Decrypt reported in March, even AWS—which dedicated to deploying 1 million Nvidia GPUs by way of 2027—is constructing its personal chips concurrently to chop that long-term publicity.

Frozen v2 remains to be exploratory. Key design choices aren’t finalized, Google hasn’t confirmed the challenge exists, and the chip will not be provided to exterior Cloud clients—{hardware} hardwired for one mannequin cannot run anybody else’s. Deployment is focused for 2028 on the earliest, in line with experiences.

Within the meantime, Google is paying SpaceX $920 million a month to hire 110,000 Nvidia GPUs from xAI’s information facilities as a bridge.