Google designs Frozen v2 AI chip for Gemini

Alphabet designs new Gemini chip Frozen v2, The Information says server silicon targets six-to-10-fold token-per-watt gains, investors cheer 2028 roadmap while AI spending bill keeps rising

Alphabet is designing a new server chip intended to make its in-house Gemini AI models run more efficiently, according to TechCrunch’s summary of a report by The Information. The chip, internally dubbed Frozen v2, is slated for release sometime in 2028, and is described as potentially delivering between six and 10 times more efficiency than Google’s current AI chips when measured by tokens generated per unit of power.

Google did not directly confirm the report when asked, telling TechCrunch that its teams are “constantly researching and experimenting” and that not every project reaches production. Still, the company’s emphasis on co-designing hardware and software “from the ground up” points to the same direction the rest of the sector has been taking: the fastest way to cut AI costs is not merely to negotiate better cloud contracts but to change the machine that runs the model.

The timing matters because AI’s economics are shifting from novelty to operating expense. Investors have repeatedly questioned Alphabet’s planned outlays for AI infrastructure; earlier this year, Google said it expects to spend between $180 billion and $190 billion on AI-related efforts, according to TechCrunch. If a future chip can materially increase output per watt, it changes the internal price of serving Gemini—how many answers can be produced before power and data-centre capacity become the bottleneck.

Frozen v2 also fits a broader push by AI companies to reduce dependence on Nvidia, whose chips have dominated the market and whose supply constraints have become a strategic variable in product roadmaps. TechCrunch notes that OpenAI announced its first custom chip in June, an inference processor called Jalapeño, while Anthropic has reportedly discussed a chipmaking partnership with Samsung. The common thread is that model makers are trying to own the part of the stack that sets the marginal cost of every prompt.

There is a second-order effect for customers: if the leading AI providers succeed in driving down inference costs with custom silicon, the price war shifts from marketing to metering. Cheaper tokens can mean broader deployment, but it also means more total usage, more logging, and more dependence on whichever platform controls the hardware pipeline. The Information report appears to have been taken as good news by markets; TechCrunch says Alphabet’s stock climbed about 3% on Monday morning after the story.

For now, Frozen v2 remains a name on an internal roadmap and a date two years away. The company that told Wall Street it plans to spend roughly $180–$190 billion on AI is also asking investors to wait until 2028 for the chip that could make those workloads cheaper to run.