Key Facts Verified

Chip codename: Frozen v2
Target: 6-10x tokens/watt vs latest TPUs
Deployment target: 2028
Origin: Jeff Dean concept
Status: In development, not announced

Sources Used

The Information (primary)
Google Cloud Blog (Ironwood specs)
CNBC, Bloomberg, Yahoo Finance
SemiAnalysis (TPU analysis)
The Decoder, Tom’s Hardware

⚠️ Editorial Notes

All claims sourced; no speculation presented as fact
Google has not formally announced Frozen v2
2028 deployment is a target, not a guarantee
Efficiency claims are internal projections, not measured
Distinguishes “architecture frozen” vs “weights frozen”

What Is Google’s Frozen v2 Chip?

AI & Cloud Infrastructure: Google is quietly building a server processor that would hardwire parts of its Gemini AI model directly into the silicon itself. Internally codenamed Frozen v2, the chip is designed to serve AI responses with dramatically less power than today’s general-purpose accelerators. Engineers estimate it could deliver six to ten times more AI tokens per watt than Google’s latest Tensor Processing Units (TPUs), according to sources cited by The Information. Google is targeting deployment as early as 2028, though the design remains unfinished and the exact scope of what gets etched into hardware is still under debate.

The concept originated with Jeff Dean, Google DeepMind’s chief scientist. His first “Frozen” design proposed embedding Gemini’s model weights – the tuned parameters that determine how the model responds – directly into the chip. That approach was scrapped because the hardware would have locked Google to a single Gemini version, rendering it obsolete within months. Frozen v2 takes a more flexible path: it freezes the architecture (the underlying blueprint of the model) while keeping the weights updatable. New model versions can still be loaded, but the chip’s circuitry is optimized for Gemini’s specific data movement patterns, matrix operations, and attention mechanisms.

Google has not formally announced Frozen v2. In a statement to multiple outlets, the company said its teams “are constantly researching and experimenting with new innovations” and cautioned that “not every project moves into production.” It also emphasized that co-designing hardware and software “from the ground up” remains central to its infrastructure strategy.

Why Google Needs a More Efficient Inference Engine

The project is driven by a severe and well-documented capacity crunch. Google is reportedly struggling to allocate enough AI compute across its own research teams and paying cloud customers. The shortage has fueled internal tensions and, in some cases, forced Google Cloud to turn away external deals because it could not guarantee sufficient capacity. Bloomberg reported in May 2026 that DeepMind researchers are queuing for TPU access while large blocks of chips are committed to customers like Anthropic – which signed a deal worth up to $40 billion for TPU capacity – and Meta.

At the same time, Google faces competitive pressure on the model front. The launch of Gemini 3.5 Pro has been delayed beyond its original June 2026 target after the model reportedly failed to meet internal performance benchmarks, particularly in coding tasks. Rivals have not stood still: Anthropic released Claude Mythos (later Fable 5), OpenAI shipped GPT-5.6 Sol, and Chinese lab Moonshot open-sourced Kimi K3 with 2.8 trillion parameters. Without a frontier model firing on all cylinders, Google needs its infrastructure to run whatever it has as efficiently as possible.

How Frozen v2 Differs from Google’s TPUs

Google’s current TPU lineup – led by the seventh-generation Ironwood chip announced in April 2025 – is a general-purpose AI accelerator. Ironwood delivers 4.6 petaFLOPS of FP8 compute per chip, 192 GB of HBM3e memory, and scales to 9,216-chip superpods capable of 42.5 exaFLOPS. It is designed to train and serve a wide variety of models, including Gemini, Claude, and customer workloads. Google has already announced an eighth generation split into two chips: TPU 8t (training) and TPU 8i (inference), both targeting late 2027.

The Technical Bet: Freezing Architecture, Not Weights

The name “Frozen” borrows from a machine-learning concept: freezing model parameters so they stop updating during training. In hardware terms, it means permanently locking some portion of the model’s behavior into the chip’s logic gates. The critical engineering decision is what to freeze.

Embedding weights directly would have been the most aggressive approach – and the most fragile. Model weights change with every training run. A chip built around Gemini 2.5’s weights would be useless for Gemini 3.0. By freezing the architecture instead – the layer structure, attention patterns, dataflow topology, and quantization schemes – Google gains most of the efficiency benefit while preserving the ability to load new weights. The trade-off is that Frozen v2 only works as long as Gemini’s foundational structure stays recognizable. A radical architectural pivot would strand the hardware.

How much of the architecture actually gets hardcoded is reportedly still undecided. Google could freeze common mathematical paths, memory access patterns, or sparsity structures that persist across Gemini versions. The rest – specific weight values, fine-tuned behavior, new capabilities – would remain in software or external memory.

Why Tokens per Watt Is the New Battleground

AI economics have shifted. Training a frontier model is a one-time capital expense – enormous, but finite. Inference – the act of answering a prompt – is a recurring operational cost that scales with every user, every query, every agentic workflow. A single long conversation or coding session can consume millions of tokens. Multiply that by billions of users and the power bill becomes a strategic constraint.

A sixfold to tenfold improvement in tokens per watt would not just save electricity. It would effectively create capacity without building new data centers. Power, cooling, and grid connections are now the limiting factors for AI expansion. Google operates data centers where available megawatts are already spoken for. A chip that squeezes more useful work from each watt expands the effective capacity of existing infrastructure.

The savings could flow in several directions: lower API prices for developers, higher usage limits for consumers, more aggressive reasoning budgets per query, or simply better margins in a market where inference costs increasingly determine profitability.

The Competitive Landscape: Everyone Wants Custom Silicon

Google is not alone in trying to escape the Nvidia tax. The major AI operators are all building or buying custom chips:

Amazon Web Services deploys Trainium for training and Inferentia for inference, alongside Graviton CPUs.
Microsoft has developed Maia accelerators for Azure AI workloads.
Meta is iterating on its Meta Training and Inference Accelerator (MTIA) for recommendation and generative AI.
OpenAI introduced its first in-house inference chip, developed with Broadcom.
Anthropic has reportedly begun preliminary discussions with Samsung about custom silicon.

None of these efforts signal the death of Nvidia GPUs. General-purpose accelerators remain essential for research, rapidly evolving models, and third-party workloads that demand portability. Custom chips absorb the high-volume, predictable inference traffic – the “boring” work that happens billions of times per day. That is exactly where Frozen v2 would slot in.

Risks: What Could Go Wrong

For all its promise, Frozen v2 carries substantial technical and strategic risks:

Architectural obsolescence. If Gemini undergoes a fundamental redesign before 2028, portions of the frozen silicon could become inefficient or irrelevant.
Manufacturing bottlenecks. Advanced AI chips depend on TSMC’s CoWoS packaging and high-bandwidth memory from a handful of suppliers. Google has already reportedly cut its 2026 TPU production target from 4 million to 3 million units due to packaging constraints. Nvidia alone has booked roughly 60% of available CoWoS capacity.
System-level gaps. A 10× chip-level improvement does not automatically translate to 10× service-level savings. CPUs, networking, storage, cooling, and idle capacity still consume power and money.
Flexibility loss. Hardware tuned for Gemini may perform poorly with third-party models, limiting its utility if Google wants to offer it through Cloud.
Project cancellation. Google has explicitly warned that not every research project ships. Frozen v2 could be folded into a future TPU generation, scaled back, or abandoned entirely.

What This Means for Cloud AI

If Frozen v2 reaches production and hits its targets, it would mark a meaningful shift in how AI infrastructure is built. The industry has long talked about hardware-software co-design. Google would be practicing it at an unprecedented scale – not just optimizing a compiler for a chip, but designing the chip around the model and the model around the chip simultaneously.

For developers and enterprises, the immediate impact is indirect. Frozen v2 is expected to remain an internal asset, not a cloud product. The benefits would arrive through faster Gemini APIs, lower pricing, higher rate limits, or more capable reasoning modes in Google’s consumer and enterprise products. For Google Cloud customers renting TPUs, the relief would come indirectly: every Gemini request moved to Frozen silicon frees up a TPU or GPU for other workloads.

The longer-term signal is clearer. The era of one-size-fits-all AI accelerators is ending. The next decade of cloud AI will likely be built on heterogeneous fleets: general-purpose GPUs for research and flexibility, training-optimized chips for model development, inference-optimized chips for serving, and model-specific processors like Frozen v2 for the highest-volume, most predictable workloads. The winners will be the companies that can orchestrate across all of them seamlessly.

What to Watch Next

Alphabet reports Q2 2026 earnings on July 22, 2026. While detailed Frozen v2 disclosures are unlikely this early, comments on AI capital expenditure, capacity constraints, and Cloud demand will provide context. More concrete signals will emerge as Google finalizes the architecture, selects manufacturing partners, and decides whether Frozen v2 graduates from testbed to production fleet.

Until then, Frozen v2 remains a compelling idea with a long road to reality – and a reminder that in the AI race, the most important competition may not be between models, but between the data centers that run them.

Sources

TradingKey, “Gemini Architecture Written Directly Into Silicon: Technical Details of Google’s New AI Inference Chip ‘Frozen v2’ Revealed,” July 20, 2026. https://www.tradingkey.com/analysis/stocks/us-stocks/262041886-google-plans-frozen-v2-ai-inference-chip-gemini-models-tradingkey
TechCrunch, “Google is working on a new AI chip designed to make Gemini more efficient,” July 20, 2026. https://techcrunch.com/2026/07/20/google-is-working-on-a-new-ai-chip-designed-to-make-gemini-more-efficient/
THE DECODER, “Google’s ‘Frozen v2’ chip reportedly bakes Gemini’s architecture directly into silicon for efficiency gains,” July 21, 2026. https://the-decoder.com/googles-frozen-v2-chip-reportedly-bakes-geminis-architecture-directly-into-silicon-for-efficiency-gains/
The New Stack, “Google just bet its inference future on a chip built for one model,” July 21, 2026. https://thenewstack.io/google-frozen-gemini-chip/
Windows Forum, “Google Frozen v2 Targets 6–10x Gemini Tokens per Watt by 2028,” July 21, 2026. https://windowsforum.com/threads/google-frozen-v2-targets-6-10x-gemini-tokens-per-watt-by-2028.439722/
Google Cloud Blog, “Ironwood: The first Google TPU for the age of inference,” April 9, 2025. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/
Data Gravity, “Google’s TPU Supply Chain,” June 2, 2026. https://www.datagravity.dev/p/googles-tpu-supply-chain

Was this article helpful?

Yes0No0

Feature	Frozen v2 (Projected)	TPU v7 Ironwood (Current)
Primary Use	Gemini inference only	Training + inference, general-purpose
Model Integration	Architecture hardwired into silicon	Programmable, loads any compatible model
Efficiency Target	6–10× tokens per watt vs. latest TPUs	2× perf/watt vs. Trillium (prior gen)
Weight Updates	Supported (architecture frozen, weights flexible)	Fully programmable
Target Deployment	2028	Available now (GA 2026)
Production Volume	Smaller run, internal/testbed	Mass production, cloud + internal
External Availability	Likely internal only	Available via Google Cloud

Google Frozen v2 AI Chip: How Gemini-Powered Silicon Could Change Cloud AI

Google is reportedly developing Frozen v2, a chip that embeds Gemini's architecture directly into silicon for 6–10× better inference efficiency. Here's what it means for cloud AI.

Quick Summary

Key Topics

Key Takeaways

Conclusion