Google is designing a new server chip, internally dubbed “Frozen v2,” that hardwires parts of its Gemini AI models directly into the silicon to make serving AI responses far more efficient. The chip, reported by The Information and not officially confirmed by Google, is not expected until 2028, but engineers cited in the report estimate it could be six to ten times more efficient than Google’s existing Tensor Processing Units (TPUs) when measured by tokens generated per unit of power.<\/p>\n
Frozen v2 is a custom AI accelerator in development at Alphabet, Google’s parent company. Rather than running Gemini as software on general-purpose AI hardware, the chip is designed to bake selected portions of the model into the silicon itself. The approach mirrors a broader industry trend in which AI labs co-design hardware and software together to squeeze more performance out of every watt.<\/p>\n
Google declined to confirm or deny the report. A spokesperson told TechCrunch that “our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers,” adding that “not every project moves into production, this rigorous exploration is central to our full stack approach. By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads.”<\/p>\n
Engineers quoted by The Information expect Frozen v2 to deliver a six to ten times improvement in efficiency over Google’s current TPUs, using tokens generated per unit of power as the benchmark. A higher tokens-per-watt figure means the same hardware can serve more model responses for the same energy cost, a meaningful lever given how power-hungry large model inference has become.<\/p>\n
Two pressures are driving the effort. First, Google faces an internal compute crunch: serving Gemini at scale consumes enormous infrastructure, and squeezing more answers from each chip directly lowers operating cost. Second, AI companies across the industry are racing to reduce reliance on Nvidia, whose GPUs have historically dominated the AI accelerator market. Custom silicon lets labs tune their hardware tightly to the workloads that matter most to them.<\/p>\n
Google is not alone. In June, OpenAI announced its first custom chip, an inference processor called Jalape\u00f1o, designed to handle OpenAI model workloads on its own hardware. Earlier in July, it was reported that Anthropic was discussing a new chipmaking partnership with Samsung. The pattern is consistent: each major AI lab is working to own more of the stack underneath its flagship model.<\/p>\n
Frozen v2 is not expected to ship until 2028, according to The Information’s report. That timeline puts it firmly in the “future weapon” category rather than an immediate performance upgrade for current Gemini users. In the meantime, Google will continue serving Gemini on its existing TPUs and on Nvidia hardware.<\/p>\n
The disclosure had an immediate effect on Alphabet’s stock. Shares climbed roughly 3% on the Monday following the report, ahead of Alphabet’s earnings release later in the week. Investors have been watching closely as Google lays out the cost of its AI build-out: earlier in the year the company said it plans to spend between $180 billion and $190 billion on capital expenditures to support its AI strategy. A credible path to six to ten times efficiency on a future chip helps justify that spending by promising lower inference cost per query over time.<\/p>\n
More broadly, the AI chip race is shifting from a contest over raw FLOPs to a contest over specialization. The next competitive advantage may come less from training ever-larger general models and more from designing silicon around a single model family so that every answer costs less power, less time, and less money. Frozen v2, OpenAI’s Jalape\u00f1o, and Anthropic’s reported Samsung talks all point in the same direction: the frontier of AI performance is moving into the chip.<\/p>\n
Frozen v2 is the internal name for a new AI server chip Alphabet is developing to make its Gemini models more efficient. It is designed to hardwire parts of Gemini directly into the silicon rather than running the model purely as software on general AI hardware.<\/p>\n
According to The Information, citing anonymous sources, Frozen v2 is slated for release sometime in 2028.<\/p>\n
Engineers cited in the report expect Frozen v2 to be six to ten times more efficient than Google’s existing AI chips, measured by the number of tokens generated per unit of power. Google declined to directly confirm the figures.<\/p>\n