Alphabet is developing a new server chip aimed at making its Gemini models run more efficiently, The Information reported Monday, citing anonymous sources. TechCrunch confirmed the existence of the report and sought comment from Google, which neither confirmed nor denied the project.
The chip, internally named Frozen v2, is reportedly targeted for release in 2028. That timeline sits two and a half years out, well beyond the current product cycle, and hardware roadmaps at that distance shift often. According to the sources cited by The Information, Frozen v2 could deliver between six and 10 times the efficiency of Google’s existing AI chips, measured in tokens generated per unit of power.
Google’s public response avoided specifics. “Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers,” the company told TechCrunch, adding that “not every project moves into production.” That is a company statement, not a confirmation of Frozen v2’s existence, specifications, or timeline.
The stakes extend beyond one chip. Google has built its own AI silicon lineage since its first Tensor Processing Unit shipped in 2015, giving it roughly a decade of custom hardware experience that few competitors can match. A tokens-per-watt gain of six to 10 times, if it materializes, would attack the largest recurring cost in running frontier models: the power and data center capacity consumed at inference time, not training time. That is where usage scales with every new chatbot query and API call.
Vertical integration also reduces Google’s exposure to Nvidia, which has controlled the AI chip market and set the price and supply terms most labs live with. OpenAI announced its own inference chip, Jalapeño, in June, and TechCrunch reported earlier this month that Anthropic was discussing a custom chipmaking partnership with Samsung. Three of the largest AI labs are now pursuing chip independence in parallel, a signal that Nvidia’s margins, not just its supply constraints, have become a target.
The market reaction arrived faster than the chip itself will. Alphabet shares climbed roughly 3% Monday morning after The Information’s report published, ahead of the company’s earnings release later this week. Investors have questioned Alphabet’s planned $180 billion to $190 billion in AI spending this year; a credible path to cheaper inference gives that spending a payoff story, even on a report Google has not confirmed.
Treat the 2028 date as a target, not a delivery promise. Chip programs at this stage routinely slip or get redesigned before tape-out, and Google has not disclosed a specification sheet, a customer list, or a manufacturing partner. Cloud customers weighing multi-year Gemini API commitments should watch Alphabet’s earnings call this week for any acknowledgment of Frozen v2, since a confirmed roadmap would be the clearest sign yet that Google intends to price inference against its own hardware rather than Nvidia’s.
Reported by TechCrunch on July 20, 2026, citing an earlier report from The Information.
