Picture a restaurant kitchen where the chef spends 80% of his time flipping through recipes, hunting for pans, and scrubbing utensils — and 20% actually cooking. That is what AI chips do today. GPUs and TPUs massively outperform CPUs, yet they still pay an enormous efficiency tax for “generality.”
Google is trying to end that waste with a chip codenamed Frozen v2. Per The Information, citing people directly involved, Google is secretly developing a server chip designed specifically for the Gemini models — and project employees say its energy efficiency, measured in tokens generated per watt, could reach six to ten times Google’s latest TPU.
This is not just another chip. It may mark AI compute formally leaving the general-accelerator era for the model-specific silicon era.
The Generality Tax AI Has Paid Too Long
To understand Frozen v2, return to a root question: why aren’t existing GPUs and TPUs good enough? In 1945, von Neumann proposed the stored-program computer — instructions and data sharing one memory, the processor fetching and executing in sequence. That architecture has ruled computing for eighty years, and its core virtue is generality: the same chip runs spreadsheets, then video editing, on a program swap.
But generality has a price. Every token a large model generates is, at bottom, the same handful of operations — matrix multiplies, attention — repeated. A general processor must still carry the control logic and data paths that serve countless other programs, and much of that circuitry sits idle during inference. Worse is the memory wall: processor and memory are physically separate, data shuttles between them, and when compute outruns data movement the chip idles waiting. For inference workloads, the idle fraction is startlingly high. The industry has answered in steps — CPU to GPU to NPU/TPU, each more specialized — but even a TPU keeps substantial generality to serve many model families. Google now wants to take the decisive step.
Frozen v2: Etching the Model Architecture Into Silicon
The core idea is brutally direct: if you are going to run Gemini forever, etch Gemini’s architecture directly into the chip’s physical layout. A model decomposes into two parts — architecture (the rules of information flow) and weights (the trained parameters). Frozen v2’s strategy: freeze the architecture, keep the weights updatable.
A traditional chip runs fetch → decode → execute, never knowing what comes next. Frozen v2, knowing Gemini’s computation graph in advance, arranges compute units, storage placement, and data flow around that fixed structure — skipping addressing and decode overhead because the hardware already knows what comes next. That is the name: Google freezes part of Gemini in exchange for chip-grade efficiency.
The project is driven personally by Google chief scientist Jeff Dean — co-founder of Google Brain, original lead designer of TensorFlow, and long the hinge between Google’s distributed computing and AI infrastructure. Having the AI architect in the chip-design loop is itself the signal. And this is not Google’s first Frozen attempt: the original conceived freezing weights too, which would have bound one chip to one Gemini version — by the time design, tape-out, and deployment finished, that version would be obsolete. Frozen v2 absorbs the lesson: as long as the next Gemini keeps a similar underlying architecture, the chip stays in service.
Why Every Model Company Is Building Chips
Frozen v2 is not an isolated event. In 2026, essentially every frontier lab is betting on custom silicon: OpenAI unveiled its Jalapeño inference chip with Broadcom in June, designed around real ChatGPT and Codex workloads with the models themselves accelerating the design — nine months from design to tape-out, deployment targeted for late 2026. Anthropic has a chip program reportedly underway. DeepSeek has secretly run an inference-chip project for about a year, in talks with design houses, foundries, and memory vendors. Zhipu has similar plans.
One driver behind them all: compute cost has become a life-or-death line for model companies. Take Gemini 3.1 Pro Preview at $12 per million output tokens. One heavy request emitting 1,000 tokens costs 2.2 cents. At 100 million such calls a day, that is $2.2 million daily — roughly $800 million a year. Gemini’s monthly active users exceed 900 million; any small per-call spend, multiplied by ten-billion-scale volume, is astronomical. Compute scarcity is already degrading products: in March–April, developers complained Claude Code had gotten dumber, and investigation found median reasoning depth had collapsed from 2,200 to 720 characters — down 67%. Google faces its own shortage, severe enough that Google Cloud has declined some external orders.
The Paradigm Turn: From Models Running on Chips to Chips Shaped Like Models
Frozen v2’s meaning outruns any single product’s fate. It represents a larger turn: yesterday, chips were general and models were software, with code deciding everything; now, model architecture is beginning to define chip design; next, chip and model will be designed as one, with no clean hardware-software boundary. The historical path is legible — CPU (general) → GPU (graphics-specialized, accidentally good at AI) → NPU/TPU (AI-acceleration-specialized) → model-level ASIC (Architecture-Specific Integrated Circuit) — and Frozen v2 is that path’s end state.
The analogy: evolving from “one toolbox that fixes everything” to “forging a dedicated key for every lock.” The toolbox is flexible; the key is efficient. Once the lock — the model — is fixed, forging the dedicated key becomes inevitable. And Google holds cards rivals cannot copy: from Gemini models to TPU silicon to Google Cloud datacenters and the product ecosystem, one company owns the entire chain. Researchers can be poached and flagships delayed, but a decade of TPU compilers and datacenter infrastructure does not resign and leave.
Risks and Limits
Frozen v2 is far from a sure thing. First, the timing is awkward: reportedly the chip deploys no earlier than 2028 — two years in which AI models may iterate several generations. If Gemini’s underlying architecture shifts materially, the frozen advantage becomes frozen baggage. Reportedly Google is still deciding how much to freeze: too little, and the efficiency gain is thin; too much, and model iteration loses flexibility. Striking that balance is itself a formidable engineering challenge. Second, talent drain: Gemini co-lead Noam Shazeer departed for OpenAI; John Jumper and multiple Gemini researchers moved to Anthropic; the once-June Gemini 3.5 Pro is reportedly months behind plan with coding capability missing internal targets. When the model team itself keeps churning, does the premise of building chips for that model still hold?
What Developers and Enterprises Should Note
-
Inference costs will fall structurally. As model-specific silicon matures, inference may drop an order of magnitude — long-term good news for inference-heavy applications.
-
Architecture stability becomes a competitive factor. Frequently swapping model architectures forfeits hardware-optimization dividends; when choosing model services, evaluate the architecture’s evolution path, not just current performance.
-
Hardware-software co-design becomes a core capability. Algorithms alone are no longer enough; companies that connect software, hardware, and system layers will hold a structural advantage around 2028.
-
The API price war is far from over. With custom silicon deployed at scale, top-model API prices have more room to fall — not because the model got cheaper, but because running it did.
Is etching large models into chips the optimal answer? Two years until we know. But the direction is set: AI compute is heading for ultimate specialization. As the software-hardware boundary blurs, whoever owns the whole chain holds final pricing power. The name Frozen carries two meanings — freezing the model architecture for efficiency, and, if the road fails, Google’s standing in AI hardware freezing with it. It is a gamble. But AI history says gambles are often the only option.
