Google is developing a specialized server chip called Frozen v2 that embeds parts of its Gemini AI model's architecture directly into silicon — a radical departure from general-purpose AI accelerators that could deliver 6 to 10 times more inference efficiency than Google's latest TPUs. Reported by The Information on July 20, 2026 and confirmed by CNBC and TechCrunch, the project targets deployment as early as 2028 and is Google's most aggressive bet yet on solving the AI compute bottleneck through hardware specialization rather than raw scale.
The chip won't replace Google's TPU line — it's an experimental parallel track. But if it works, it could fundamentally change how AI inference is priced, giving Google a structural cost advantage over OpenAI and Anthropic in the race to serve billions of daily AI queries.
Last verified: July 21, 2026
- Frozen v2 hardwires Gemini's architecture (not weights) into silicon
- Projected 6-10x more tokens per watt vs. Google's latest TPUs (vendor estimate)
- Deployment targeted for 2028; treated as an experiment, not a TPU replacement
- Driven by a severe internal compute shortage — Google Cloud's backlog hit $462B in Q1 2026
- Alphabet stock rose ~3.3% intraday on the report
- Won't be sold to outside customers (unlike TPUs)
What Is Google's Frozen v2 Chip?
Frozen v2 is a purpose-built AI inference chip that permanently embeds the structural blueprint of Google's Gemini model directly into the silicon circuitry. Unlike general-purpose AI accelerators — including Google's own Tensor Processing Units (TPUs), NVIDIA GPUs, and Amazon's Trainium chips — which are designed to run any AI model you load onto them, Frozen v2 is built exclusively for Gemini.
The name comes from the machine learning concept of "freezing" parameters, where you lock certain values in a neural network so they stop changing during training. Frozen v2 takes this literally: it freezes part of Gemini's architecture into the physical chip design, so the hardware doesn't need to compute those decisions at runtime. This eliminates redundant computation steps and reduces the amount of data that must travel between the chip's memory and processor — the single biggest energy cost in AI inference today.
Here's why that matters: Google DeepMind Chief Scientist Jeff Dean has noted that running a single matrix multiplication on an AI accelerator costs roughly 1 picojoule of energy, but moving a model weight from on-chip memory to the compute unit costs about 1,000 picojoules — a 1,000x difference. Most of the power consumed by today's AI chips goes into moving data, not computing it. Frozen v2 attacks this directly by building the data pathways into the silicon itself.
How Does Frozen v2 Differ From Google's TPUs?
Google's TPU line — now on its seventh generation with Ironwood (TPU7x) — is a general-purpose AI accelerator. You can load any model onto a TPU, from Gemini to open-source models like Llama, and it will run. That flexibility is valuable, but it comes with an overhead cost: the chip must make real-time decisions about how to route data and execute computations for whatever model it's running, every single time.
| Feature | TPU (Ironwood) | Frozen v2 |
|---|---|---|
| Model support | Any AI model | Gemini only |
| What's hardcoded | Nothing — general-purpose | Gemini's architecture (blueprint, not weights) |
| Flexibility | High — load any model | Low — tied to Gemini's architecture |
| Efficiency target | 2x perf/watt over prior gen (Google Cloud Blog) | 6-10x tokens/watt vs. latest TPU (vendor estimate) |
| Sold externally? | Yes — available via Google Cloud | No — internal only |
| Deployment | In production now | Targeted for 2028 |
| Risk | Low — proven, shipping | High — obsolete if Gemini architecture changes |
Frozen v2 trades universality for efficiency. By building Gemini's decision-making pathways directly into the circuit, the chip skips the runtime computation that a TPU must perform for every inference call. Fewer steps, less data movement, fewer decisions — and that translates directly into more tokens generated per watt of power.
Why Is Google Building a Model-Specific Chip Now?
The short answer: Google is running out of compute, and the problem is getting worse, not better.
In Q1 2026, Google Cloud reported over $20 billion in revenue — up 63% year-over-year — with its backlog nearly doubling in a single quarter to $462 billion. But Alphabet CEO Sundar Pichai told analysts on the earnings call: "We are compute constrained in the near term. Our cloud revenue would have been higher if we were able to meet the demand." Google Cloud was literally turning away enterprise customers because it didn't have enough infrastructure to serve them.
The compute crunch is severe enough that Google signed a deal with SpaceX in June 2026 to lease access to 110,000 NVIDIA GPUs for $920 million per month — a contract extending through 2029 that could total roughly $30 billion. As The New York Times reported, one of the world's largest cloud computing companies is paying a rocket company for compute capacity. That's how tight the market is.
Frozen v2 is Google's long-term answer to this problem. Rather than perpetually buying or leasing more general-purpose hardware, the bet is to get dramatically more useful AI output out of the hardware Google already controls — by specializing it for the one model that matters most to Google's business.
What Was Jeff Dean's Original "Frozen" Concept?
The Frozen v2 project has a predecessor. The original "Frozen" design was proposed by Jeff Dean, Chief Scientist at Google DeepMind, who is one of the most influential figures in Google's AI history. Dean's initial concept was even more radical: he proposed embedding the model's weights — the specific numerical parameters that determine how a trained model responds to queries — directly into the chip's circuitry.
That approach would have maximized efficiency, but Google scrapped it. The problem was lifecycle: if you hardwire the weights for a specific version of Gemini into the chip, the chip becomes obsolete the moment that version is updated. Given that Google updates Gemini on a monthly cadence, a weights-locked chip would have a shelf life measured in weeks — economically nonsensical for hardware that takes years to design and manufacture.
Frozen v2 is the more pragmatic do-over. Instead of locking in the weights, it locks in the architecture — the underlying blueprint of how Gemini processes information. New weights can still be loaded onto the chip as Gemini is updated, so the hardware stays useful across model versions. The trade-off is lower efficiency than full weight-embedding, but a vastly longer useful life.
According to The Information's report, Google has not yet finalized how much of the architecture to hardcode. More hardcoded architecture means higher efficiency but lower flexibility — a balance that's still being worked out by the engineering team.
How Does Frozen v2 Compare to Other Custom AI Chips?
Google isn't alone in building custom silicon for AI inference. Every major hyperscaler is investing heavily in purpose-built chips to reduce dependence on NVIDIA and improve unit economics. But Frozen v2 takes a fundamentally different approach than any of them.
| Company | Chip | Approach | Model-specific? | Available externally? |
|---|---|---|---|---|
| TPU Ironwood (7th gen) | General-purpose AI accelerator | No — runs any model | Yes, via Google Cloud | |
| Frozen v2 | Hardwired model architecture | Yes — Gemini only | No | |
| Microsoft | Maia 200 | Inference ASIC (TSMC 3nm, 216GB HBM3e) | No — serves multiple models | No — Azure internal only |
| Amazon | Trainium 3 | Training + inference ASIC | No — runs multiple models | Yes, via AWS |
| Meta | MTIA (300/450/500) | Recommendation + ranking ASIC | Partially — optimized for Meta workloads | No — Meta internal only |
| OpenAI | Custom inference chip (reported) | Purpose-built for LLM inference | Partially — optimized for GPT models | No — internal only |
| Taalas (startup) | HC1 | Hardwired model weights via mask-ROM | Yes — one model per chip | Not yet shipping |
Most hyperscaler chips are still general-purpose or workload-optimized — they're built to run a range of models more efficiently than NVIDIA GPUs, not to serve a single model. Frozen v2 goes further than any major hyperscaler chip by hardwiring a specific model's architecture into the silicon itself.
The closest parallel is Taalas, a Toronto-based startup that raised over $169 million to build chips with entire models hardwired into silicon using mask-ROM technology. Taalas's HC1 chip embeds a full Llama 3.1 8B model directly into the hardware — an even more extreme version of the approach Google's original Frozen concept pursued before pivoting to Frozen v2's architecture-only design.
What Are the Risks of Hardwiring Gemini Into Silicon?
The fundamental risk is architectural lock-in. Frozen v2 only makes economic sense if Google keeps Gemini's underlying architecture stable for the chip's useful life — likely 3-5 years from deployment in 2028. If Google's AI research team discovers a fundamentally better model architecture in 2029, the Frozen v2 chips become expensive paperweights.
This isn't a hypothetical concern. The AI field moves fast. Transformer architectures have dominated since 2017, but alternative approaches — state-space models, mixture-of-experts variants, and entirely new paradigms — are actively being researched. Google itself has been a source of architectural innovation, and the pressure from competitors like OpenAI, Anthropic, and Chinese labs means the incentive to overhaul architecture is high.
Google is reportedly treating Frozen v2 as an experimental side project rather than a core infrastructure commitment, which is the right call. The chip will be produced in smaller volumes than the TPU line, and the company hasn't committed to making it a permanent part of its hardware stack. If the Gemini architecture proves stable enough over the next few years, the concept could scale. If not, the experiment still yields valuable learnings about the efficiency ceiling of model-specific silicon.
Additional risks include:
- Fabrication timeline risk: Chip design cycles are 18-24 months. If Frozen v2 slips, the 2028 target moves, and the efficiency advantage may shrink as TPUs continue improving.
- Architecture drift: Even small changes to Gemini's architecture could require chip redesigns, eroding the efficiency advantage.
- Competitive response: If Google's approach works, competitors will pursue similar strategies, potentially commoditizing the advantage.
What Does Frozen v2 Mean for AI Pricing and Competition?
If Frozen v2 delivers even half of the projected 6-10x efficiency gain, it could reshape AI inference economics. Today, the cost of serving AI responses is one of the central tensions in the industry — the more capable the model, the more expensive every query. Google's Gemini API was processing 16 billion tokens per minute by Q1 2026, up from 10 billion in Q4 2025, and that volume is only growing.
Cheaper inference means Google could:
- Undercut competitors on API pricing — making Gemini the cheapest production-grade AI API per token, which is a direct lever for market share against OpenAI and Anthropic.
- Improve Gemini product margins — getting more revenue per compute dollar, which matters as AI spending faces increasing investor scrutiny.
- Expand free-tier offerings — lower serving costs make it economically viable to offer more generous free AI features, driving adoption and ecosystem lock-in.
- Sustain aggressive API token throughput growth — the 16B tokens/minute trajectory needs exponentially more hardware, and efficiency gains are cheaper than buying more data centers.
This is why investors reacted positively. Alphabet shares rose as much as 3.77% intraday on July 20, 2026, closing approximately 1.5% higher, according to CNBC. The market read Frozen v2 as Google's answer to the compute crunch — a credible path to lowering the unit cost of AI serving at a time when the entire industry is grappling with infrastructure economics.
What This Means for You
If you're building AI products on Google's ecosystem (Gemini API, Vertex AI, Google Cloud): Frozen v2 is a long-term signal that Google is investing in making Gemini inference cheaper and more scalable. If the chip ships in 2028 and delivers on efficiency projections, expect API pricing for Gemini to become more competitive over time — potentially meaningfully cheaper than alternatives for production workloads. This isn't actionable today, but it's a factor in your 2027-2028 platform decisions.
If you're an AI startup or small business evaluating cloud providers: The hyperscaler custom silicon race means inference costs will keep falling across all major clouds — not just Google. Microsoft's Maia 200 already claims 30% better performance per dollar over prior hardware, and Amazon's Trainium line is driving down costs on AWS. Your strategy should be to architect for portability (use open-source models where possible, avoid deep proprietary lock-in) while benefiting from the pricing pressure this competition creates.
If you're investing in or tracking AI infrastructure: Frozen v2 represents the leading edge of a shift from general-purpose to model-specific silicon. The efficiency gains are real in principle, but the execution risk is substantial. The key metric to watch is whether Google commits to production-scale deployment post-2028 or shelves it like the original Frozen concept. The broader trend — hyperscalers building custom inference chips to reduce NVIDIA dependence — is confirmed and accelerating regardless of Frozen v2's individual fate. For more on how this dynamic plays out across the industry, see our analysis of why enterprise AI strategies fail on the infrastructure readiness gap.
If you're concerned about AI compute accessibility: Model-specific chips like Frozen v2 are inherently captive — Google won't sell them to you. This means the efficiency gains accrue to Google's internal cost structure, not to the broader market. The compute war between free and paid AI models is intensifying, and specialized silicon is one more way large platforms are pulling ahead of smaller players. Open-source models running on commodity hardware remain the counterbalance, and the case for buying your own GPU for local AI continues to strengthen as frontier inference costs diverge.
FAQ
Q: What is Google's Frozen v2 chip? A: Frozen v2 is an experimental server chip Google is developing that hardwires parts of its Gemini AI model's architecture directly into silicon. Unlike general-purpose chips like TPUs or GPUs, it's built exclusively for Gemini, trading flexibility for a projected 6-10x improvement in inference efficiency (tokens per watt). Deployment is targeted for 2028.
Q: How is Frozen v2 different from a TPU? A: TPUs are general-purpose AI accelerators that can run any model. Frozen v2 hardcodes Gemini's architecture into the chip itself, eliminating runtime computation steps and data movement. TPUs are sold to external customers via Google Cloud; Frozen v2 will be internal only. Frozen v2 is an experimental parallel track, not a TPU replacement.
Q: Why did Google scrap the original Frozen chip design? A: The original Frozen concept, proposed by Jeff Dean, would have embedded Gemini's model weights directly into the chip. This would make the chip obsolete the moment Gemini was updated to a new version — an unworkable lifecycle for hardware that takes years to design. Frozen v2 embeds the architecture (blueprint) instead of weights, allowing new weights to be loaded while still gaining efficiency from the hardcoded structure.
Q: When will Frozen v2 be deployed? A: The Information's report, confirmed by CNBC and TechCrunch, targets deployment as early as 2028. Google is treating this as an experimental project with smaller production volumes than its TPU line. The exact timeline depends on how much architecture Google decides to hardcode and whether the Gemini architecture remains stable enough to justify the investment.
Q: Will Frozen v2 make Gemini API cheaper? A: If the 6-10x efficiency projections hold, Frozen v2 would significantly lower Google's cost per token for serving Gemini. That could translate into more competitive API pricing, more generous free tiers, or higher margins — or some combination of all three. However, the impact wouldn't be felt until post-2028 deployment, and the projections are vendor estimates, not independently verified benchmarks.
Q: Is Google the only company building model-specific AI chips? A: No, but Google's approach is the most aggressive among hyperscalers. Microsoft's Maia 200, Amazon's Trainium 3, and Meta's MTIA are all custom inference chips, but they're designed to run multiple models — not hardwired for a single one. The closest parallel is Taalas, a startup that embeds entire models into silicon via mask-ROM, but it's not yet shipping at scale. For context on the broader custom silicon race, see our guide to India's indigenous AI inference chip efforts.
Discussion
0 comments