Google reportedly develops Gemini-specific chip to cut AI inference costs

Google is reportedly developing a server chip called 'Frozen v2' that hardwires parts of its Gemini AI model architecture into silicon, aiming to cut inference costs by producing six to 10 times as many tokens per unit of power as its newest TPUs, according to The Information. The project targets deployment as early as 2028 and addresses the 'inference tax' Google Cloud has described from costs including idle hardware and power consumption, though Google has not confirmed the chip and cautioned that not every design reaches production.

Google https://about.google/company-info/our-story/?ref=runtimewire , founded by Larry Page and Sergey Brin as a Stanford search project, is reportedly developing a server chip that would hardwire parts of Gemini's architecture into silicon. The project amounts to a long-dated bet that Google's AI models can become stable enough to justify hardware designed around them. The chip, informally called "Frozen v2," was first reported by The Information https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently?ref=runtimewire . A Reuters markets note https://www.reuters.com/commentary/reuters-open-interest/global-markets-view-usa-2026-07-21/?ref=runtimewire on July 21st said the report helped Alphabet shares, which had trailed other large technology stocks ahead of the parent group's second-quarter results. Google has not announced Frozen v2. The reported 2028 deployment target and projected efficiency gains remain internal estimates, and Google told The Information that its hardware exploration does not guarantee every design will reach production. Page and Brin met at Stanford in 1995 and built an early search engine called Backrub from their dorm rooms. Google says Sun Microsystems co-founder Andy Bechtolsheim supplied its first outside check, worth $100,000, in August 1998. Frozen v2 extends that hardware instinct into the most expensive part of the current AI race. Google already controls the Gemini models, the cloud service that sells access to them, and the data centers and Tensor Processing Units used to run them. Frozen v2 would push that control deeper into the processor itself. The model becomes part of the machine General-purpose AI accelerators load different models and make a series of runtime decisions while moving large volumes of data between memory and compute. Frozen v2 would reportedly encode some Gemini-specific architectural decisions in transistors, reducing the work and data movement required for each query. Engineers working on Frozen v2 have projected that it could produce six to 10 times as many tokens per unit of power as Google's newest TPUs, according to The Information. The project is targeting deployment as early as 2028, and Google reportedly views the first generation as an experiment rather than a processor destined for TPU-scale production. An earlier design led by Google DeepMind chief scientist Jeff Dean reportedly went further by placing Gemini's model weights directly into the chip. Google set that version aside because hardware tied to one set of weights could become obsolete as soon as the model changed. Frozen v2 would preserve updatable weights while fixing part of the underlying architecture in hardware. That compromise gives Google room to improve Gemini after fabrication, provided future Gemini releases retain enough of the same structure. The risk is clear. Chip design and manufacturing work on multi-year schedules, while frontier model architectures can change within months. A significant redesign of Gemini before Frozen v2 enters service could reduce the chip's usefulness or force Google to limit which models run on it. The reported project has not yet settled how much of Gemini's architecture to fix in silicon. Google's immediate constraint is capacity The report said Google Cloud has turned away some outside business because of insufficient AI compute. A chip targeted for 2028 cannot relieve that near-term shortage. Frozen v2 instead addresses the operating cost and power required to serve AI at far greater volume. Google has been making that pressure explicit. In a July AI infrastructure report https://cloud.google.com/blog/products/compute/state-of-ai-infrastructure-report-overview/?ref=runtimewire , Google Cloud described an "inference tax" from costs including idle specialized hardware, storage and data movement, and framed power consumption as a hardware-selection constraint. Those figures come from Google's own survey, but the incentives behind Frozen v2 match the problem it describes. Each additional agent, search query or generated response consumes inference capacity. Lowering the energy and compute required per token would let Google serve more demand from a fixed power envelope, improve response times and protect margins as AI spreads through Search, Workspace and Google Cloud. Alphabet's second-quarter results call https://alphabet2025ir.q4web.com/investor/news/news-details/2026/Alphabet-Announces-Date-of-Second-Quarter-2026-Financial-Results-Conference-Call-2026-2h R0kzZHY/default.aspx?ref=runtimewire is scheduled for July 22nd at 4:30 p.m. Eastern. Reuters cited analyst expectations for $117 billion in quarterly revenue, up 21% from a year earlier, and 64% annual growth at Google Cloud. Those are forecasts rather than reported results. Investors will be looking for evidence that AI demand is translating into revenue quickly enough to cover the infrastructure needed to supply it. Specialized chip founders have an incumbent at their heels Google's reported work also echoes a thesis that AI chip founders have spent years trying to prove: inference can support processors designed around narrower workloads, rather than hardware expected to run every possible model. Ljubisa Bajic's Toronto-based Taalas https://taalas.com/the-path-to-ubiquitous-ai/?ref=runtimewire has already taken the more aggressive route that Google rejected for its first Frozen design. Taalas hardwired the weights of Llama 3.1 8B into its HC1 processor. Taalas claims HC1 can produce 17,000 tokens per second per user, though the benchmark was run by Taalas and has not been independently established in the materials reviewed here. Taalas has raised more than $200 million. Etched https://www.globenewswire.com/news-release/2026/06/30/3319922/0/en/Etched-Emerges-From-Stealth-With-Working-Chip-800M-Raised-and-Over-1B-in-Customer-Contracts.html?ref=runtimewire is pursuing transformer-focused inference systems. On June 30th, Etched said it had emerged from stealth with a working chip, raised $800 million and signed more than $1 billion in customer contracts. Those figures are self-reported. The founders of specialized chip developers now have stronger evidence that hyperscalers see value in their core premise. They also face a harder sales argument. Google can co-design a model, processor, data center and cloud service without waiting for an external customer to adopt the hardware. Independent chip developers must deliver enough speed or cost savings to offset the incumbent's distribution, manufacturing scale and control of the workload. Frozen v2 remains a research project with an unresolved design and a distant target date. Its strategic direction is already legible. Google is preparing for an AI market where the cost of running models matters as much as the cost of training them, and where the winning model may increasingly arrive with its own machine.