Google's Frozen v2 Chip Bakes Gemini Into Silicon

Google is reportedly building a Frozen v2 chip that does something most AI accelerators avoid. It hardwires the Gemini model architecture directly into the silicon. According to a report from The Information, relayed by The Decoder, the server chip could serve AI responses 6 to 10 times more efficiently than Google's current Tensor Processing Units. The target date is 2028, and the goal is blunt. Google wants to cut the cost of running its own models.

The idea sounds narrow, yet it points at the central problem of the AI business. Training a model is a one-time expense. Serving it to millions of users runs forever. Because inference is the bill that never stops, a chip that slashes it changes the economics of an entire company.

What the Frozen v2 chip actually does

Start with the word frozen, because it explains the design. A normal TPU stays flexible. It can run many different models, loading whatever weights a customer sends it. That flexibility costs performance, since general hardware must handle cases it may never see. The Google Frozen v2 chip gives up some of that generality on purpose.

The chip embeds Gemini's structural blueprint into the hardware itself. The Information described the split precisely, saying Frozen v2 "embeds model architecture instead of weights." In plain terms, the shape of the network becomes physical. The wiring assumes a specific pattern of attention layers and feed-forward blocks. New weights can still load dynamically, so Google can keep training fresh versions. What stays fixed is the model architecture underneath.

That fixed structure removes compute steps. When the hardware already knows the model's shape, it skips the overhead of interpreting a general graph. As a result, each token takes less energy to produce. The reported gain, 6 to 10 times better efficiency measured in tokens per watt, comes from that specialization.

An analogy helps here. A general TPU works like a versatile factory that can retool to build many products, which is powerful but slow to switch. A frozen chip works like a purpose-built assembly line that makes one product beautifully and nothing else. The assembly line wins on cost and speed for that single product. It simply cannot make anything else without being rebuilt, and that is the deal Google is choosing.

The Jeff Dean idea, refined

The concept traces back to Jeff Dean, chief scientist at Google DeepMind. His first proposal went further. He wanted to embed the model weights directly into the chip, which would have made it even faster. Google rejected that version, because a chip tied to specific weights would become obsolete the moment a new Gemini update shipped. Freezing the architecture instead of the weights is the compromise. It keeps most of the speed while leaving room to retrain.

Understanding AI inference cost

To see why an AI inference chip matters, separate the two big AI expenses. Training builds the model once. Inference runs it for every user, every day, without end. Because inference cost accumulates with usage, it eventually dwarfs the training bill for any popular product. A frontier assistant that answers billions of prompts pays that tax without pause.

This is why TPU efficiency has become the metric engineers obsess over. They now measure chips in tokens per watt, meaning how much text a chip can generate for each unit of electricity. A higher tokens per watt figure means lower inference cost and less strain on the grid. The Frozen v2 chip attacks exactly this number by stripping out general-purpose overhead.

Data center power is the real constraint behind the whole race. Grid capacity grows slowly, and new plants take years to build. So the fastest way to serve more users is to squeeze more output from each watt. An efficient AI inference chip does that in hardware, which is far more durable than a software trick a rival can copy overnight.

Why 6 to 10 times efficiency matters

Numbers like these are easy to skim past, so it helps to translate them. If a chip serves the same model with one-sixth to one-tenth of the power, then Google can serve far more users on the same electricity budget. A more efficient chip effectively raises the ceiling on growth without building new plants.

The cost angle is even sharper. Google, OpenAI, and Anthropic all pay to answer every query. If Google's per-token inference cost drops by a large factor, it gains room to lower prices or protect margins. The Decoder framed this directly, noting the chip could give Google "a price advantage over OpenAI and Anthropic." In a market where buyers increasingly shop on price, that advantage compounds.

There is a defensive motive too. TechCrunch reported that the effort is part of Google's broader push to reduce its Nvidia dependence. Every accelerator Google designs in-house is one it does not have to buy at a premium. For a company spending as heavily as Alphabet, that shift matters over time.

How specialization compounds

Specialization tends to pay off more the longer it runs. A general chip wastes a little energy on every single token, because it must stay ready for any model. A frozen chip pays that generality tax only once, at design time. After that, every token it serves is cheaper than the same token on flexible hardware. Across the lifetime of the chip, those small savings stack into a large structural advantage that is hard for a competitor to erase.

The flip side is real, though. A general chip stays useful even when models change, because it never assumed any particular design. A frozen chip trades that safety for speed. Google is betting the trade is worth it for a model as central as Gemini, where the volume of queries is so high that even a modest per-token saving becomes enormous in aggregate. Smaller or faster-changing models would not justify the same gamble.

A chip that cannot leave home

The Frozen v2 chip carries an obvious catch, and Google seems to accept it. The chip only works as long as the company keeps the same Gemini architecture. The moment Google redesigns the model in a fundamental way, the frozen hardware loses its fit. This inflexibility likely rules out any commercial release to outside customers, according to The Information.

That limitation shapes how Google plans to use it. Unlike TPUs, which Google leases to Meta and other cloud customers, Frozen v2 is built for internal efficiency alone. The reported production volume is smaller than the main TPU line. Sources describe it as a test run for specialized chips rather than a mass product. In other words, Google is experimenting with how far hardware specialization can go before it becomes too brittle.

The bet is subtle. Google is wagering that its model architecture will stay stable enough, for long enough, to justify casting it in silicon. If Gemini's core design holds through 2028 and beyond, the chip pays off. If a breakthrough forces a redesign, the frozen investment ages fast. That tension between speed and flexibility sits at the heart of the whole project.

Google's TPU foundation

Frozen v2 does not appear from nowhere. It sits on top of a decade of custom-chip work. Google's Tensor Processing Units are among the most mature AI accelerators outside Nvidia's lineup. In 2026 the company made its seventh-generation chip, the Ironwood TPU, generally available at Google Cloud Next. Google branded it "the first Google TPU for the age of inference," a phrase that signals where its priorities now sit.

The specifications are large. As Google detailed on its own blog, the Ironwood TPU delivers 4.6 petaFLOPS per chip and 42.5 exaFLOPS across a 9,216-chip superpod. Those are the kinds of figures that let Google run frontier models at global scale. Ironwood is the current workhorse, and the Frozen v2 chip is the experimental step beyond it.

The training and inference split

Google's next move already points toward specialization. The company is previewing an eighth generation that splits into two chips for the first time. According to The Next Web, the TPU 8t, codenamed Sunfish, is a Broadcom-designed training chip. The TPU 8i, codenamed Zebrafish, is a MediaTek-designed inference chip. Both target TSMC's 2nm process and late 2027. Anthropic is reportedly the anchor customer, with its deal expanding to 3.5 gigawatts of compute in 2027.

Broadcom sits at the center of this supply chain. It designs the Ironwood TPU and the 8t training chip under an agreement that runs through 2031, and it commands more than 70 percent of the custom AI accelerator market. Its AI revenue reportedly hit 8.4 billion dollars in a recent quarter, up 106 percent year over year. Google has also held talks with Marvell to widen its options beyond Broadcom, with TSMC handling fabrication. The Frozen v2 chip fits this pattern, taking the inference-specialization logic of the TPU 8i even further. It shows how deep Google is willing to go to control its own hardware destiny.

The custom silicon race

Google is not alone in this shift, and the timing is not a coincidence. Every large AI company now faces the same inference bill, so every one of them is chasing custom silicon. The result is a quiet arms race in accelerators built for serving models rather than training them.

OpenAI made its move first. In June 2026 it unveiled Jalapeño, its first custom chip, built with Broadcom. As TechCrunch reported, Jalapeño is an inference-only accelerator designed from scratch in roughly nine months. It anchors a 10-gigawatt collaboration with Broadcom, with deployment starting in late 2026. Microsoft is expected to buy about 40 percent of the initial production for its own data centers.

Anthropic is moving too. TechCrunch noted the Anthropic Samsung talks over a custom chip partnership in July 2026, even as the company leans on Google TPUs for training. The picture is clear. The three leading model builders are all trying to own more of their hardware stack, because owning the chip means controlling the cost.

Notice the common thread running through all of it. Broadcom builds Google's Ironwood TPU and its 8t training chip, and Broadcom also built OpenAI's Jalapeño. One supplier now sits behind much of the industry's custom hardware, which explains its soaring revenue. For the model builders, the lesson is that designing a chip and manufacturing one are separate skills. They can own the design while leaning on a partner for the hard physics of fabrication.

Frozen v2 embeds model architecture instead of weights.The Information

The economics behind the bet

To understand why Google would freeze a model into silicon, follow the money. The Alphabet capex plan reportedly runs in the range of 180 to 190 billion dollars. Much of that goes to chips and data centers. When you spend at that scale, a chip that cuts serving costs by a large factor is not a curiosity. It is a lever on the entire budget.

Consider the compounding effect. Suppose the Frozen v2 chip hits even the low end of its target and serves Gemini six times more efficiently. Every query then draws a fraction of the power it does today. Across billions of daily requests, the saved electricity and freed capacity add up to real money and real headroom. That is the prize Google is chasing, and it explains the willingness to accept a rigid chip.

The competitive stakes sharpen the math. If Google can serve its models cheaper than OpenAI or Anthropic, it can compete harder on price without bleeding margin. Buyers of AI increasingly compare cost per token across providers. A structural cost edge, baked into hardware, is difficult for a rival to copy quickly. That durability is exactly what makes a frozen chip attractive despite its inflexibility.

The comparison with buying Nvidia hardware makes the logic concrete. A top Nvidia accelerator carries a large margin that flows to the chipmaker, not to Google. When Google designs its own silicon, it keeps that margin and gains supply it controls. Over a multi-year build-out, avoiding the Nvidia premium on millions of chips is itself a strategy, separate from any efficiency gain the Frozen v2 chip might add on top.

The risk of freezing too early

Still, the strategy carries genuine uncertainty. AI architectures have changed fast over the past few years. Attention variants, mixture-of-experts routing, and new memory schemes keep arriving. If one of those shifts proves essential, a chip locked to today's Gemini design could look dated by the time it ships in 2028. Google is betting that the pace of architectural change will slow, or at least that its core design will endure. That is a real wager, not a sure thing.

Where inference goes next

The Frozen v2 chip is a signal, and the signal is louder than the chip itself. It says the industry has moved past the training obsession and into the serving era. For years the story was about who could build the biggest model. Now the story is about who can run that model cheaply enough to make a profit. Hardware specialization is the natural answer, and freezing an architecture is the most aggressive form of it yet reported.

Whether Frozen v2 ships on schedule or slips, the direction is set. Google's TPU roadmap already splits training from inference with the 8t and the TPU 8i. OpenAI has Jalapeño. Anthropic is courting Samsung. The general-purpose accelerator is giving way to chips tuned for one job. Frozen v2 simply takes that trend to its logical edge, casting a single model's shape into permanent hardware.

For anyone building on these platforms, the practical effect will show up in pricing and availability rather than in spec sheets. Cheaper inference tends to reach users as lower prices and higher rate limits. If Google's bet works, the payoff will not be a headline about a chip. It will be a quiet, steady drop in what it costs to run a frontier model, and that is the number that ultimately shapes the market.

There is also a broader lesson for the field. The Frozen v2 chip suggests that model design and hardware design are converging into a single discipline. When a company can cast its architecture into silicon, the line between software and hardware blurs. Future model teams may plan their designs around what silicon can freeze cheaply, rather than treating the chip as an afterthought. If that happens, Google's experiment will look less like a one-off and more like the start of a new way to build AI systems from the metal up.