Google is reportedly working on a new server chip, internally called Frozen v2, that is designed specifically to run Gemini models more efficiently.

The chip has not been announced by Google. The Information reported that the project could launch as soon as 2028 and may deliver six to ten times more Gemini output per unit of power than Google's latest AI chips. TechCrunch and Bloomberg both summarized the report on July 20, noting that the chip is being framed as a way to reduce the cost of serving Google's own models at scale.

If the report holds, Frozen v2 would be more than another TPU generation. It would push Google's hardware strategy closer to model-specific silicon: instead of making a general AI accelerator that can serve many model families, Google would tune more of the hardware around Gemini's own architecture.

What is different about Frozen v2

Google already has a long-running custom silicon program. Its Tensor Processing Units power internal services and Google Cloud customers, and its current Ironwood generation is built for large-scale training and inference. Google says Ironwood supports dense and mixture-of-experts models, decode-heavy inference, and pod-scale deployments with up to 9,216 chips.

Frozen v2 appears to be a different bet. The Information's report says the chip would embed parts of Gemini's architecture more directly into silicon. The practical goal is straightforward: reduce the energy and hardware cost of generating tokens.

That matters because the AI infrastructure race is shifting from "who can train the biggest model" to "who can serve the most useful model at the lowest cost." Training remains expensive, but consumer chatbots, coding agents, AI search, Workspace features, and multimodal assistants all create a recurring inference bill. If a company can serve each response with less power, it can either improve margins, lower prices, or spend the savings on larger context windows and more agentic workflows.

Why Google would do this

Google has one advantage most AI labs do not: it controls the model, the serving stack, the cloud platform, and the hardware roadmap. That makes vertical integration more realistic.

Nvidia GPUs are flexible and dominant because they can run many workloads across many customers. Google does not need Frozen v2 to be equally general if the highest-value workload is Gemini. A narrower chip can be a good trade if it serves billions of Gemini requests more efficiently than a broader accelerator.

This is also why the reported timing matters. A 2028 target would give Google time to align the chip with future Gemini architectures, not just today's model. But that creates the central risk: if too much of the model shape is baked into hardware, Google has to predict where its own model family is going years in advance.

The caveats

The biggest caveat is that Frozen v2 is still a reported project, not a public product. Google has not published a chip spec, benchmark, launch date, model support list, or cloud availability plan.

The second caveat is architectural lock-in. A model-specific chip can be extremely efficient when the model shape stays stable. It can become awkward if the model architecture changes faster than the hardware cycle. AI labs are still experimenting with routing, memory, multimodal encoders, test-time compute, and agent tool use, so a design that looks perfect in 2026 could be less ideal by 2028.

The third caveat is comparability. "Six to ten times more efficient" sounds huge, but it needs a denominator. Efficiency can mean tokens per watt, cost per query, latency at a given quality level, or throughput under a specific serving setup. Until Google publishes numbers, the claim is best treated as a direction of travel rather than a benchmark.

Why builders should care

For developers, the important story is not the chip name. It is the direction: frontier AI is becoming a full-stack systems problem.

The winning model providers will not only compete on benchmark scores. They will compete on how cheaply they can serve long-context reasoning, code execution, search grounding, image and video understanding, and background agents. Hardware tuned for a specific model family could make those workloads cheaper and more available.

That also raises a platform question. If Gemini becomes materially cheaper to run on Google-owned hardware than rival models are on generic accelerators, Google can use that cost structure across Search, Android, Workspace, Cloud, and developer APIs. The advantage would not show up only in model leaderboards. It would show up in product margins and how aggressively Google can bundle AI into everyday tools.

Frozen v2 is still early. But the signal is clear: the next AI chip war may be less about generic acceleration and more about how tightly a company can bind its model architecture, compiler stack, data centers, and products into one system.