What is Google's Frozen v2 chip?

Frozen v2 is Google's in-development custom silicon that hardcodes parts of the Gemini model architecture directly into the chip, rather than running Gemini as software on general-purpose AI accelerators. Bloomberg reported the project on July 20, 2026. Engineers estimate 6-10x more tokens generated per watt vs Google's current TPU line. It's designed to run alongside Google's TPU 8t (training) and TPU 8i (inference) chips announced in 2026 — not to replace them. Target production window is 2028. The name 'Frozen' refers to the model architecture being frozen into the hardware — impossible to change once fabricated.

How is Frozen v2 different from a TPU?

TPUs are general-purpose matrix-multiply accelerators — they can run any transformer model (Gemini, Claude, GPT, Llama variants). Frozen v2 is model-specific silicon — it only runs Gemini (or Gemini derivatives with compatible architecture). Because the chip 'knows' Gemini's structure, it can hardwire operations that a TPU has to compute dynamically. The tradeoff: 6-10x better tokens-per-watt for Gemini specifically, versus zero flexibility to run any other model. TPUs remain the workhorse for training and general inference; Frozen v2 becomes the specialist for Gemini production inference at extreme scale.

When will Frozen v2 be available?

Bloomberg's July 20, 2026 report indicates Frozen v2 is targeting 2028 for production availability. Engineers are still finalizing the design and deciding how much of the model to permanently hardwire. This is an internal chip for Google's own datacenters — it will not be sold to third parties (unlike TPUs which are rentable via Google Cloud). External customers will benefit indirectly through cheaper Gemini API pricing and larger free-tier quotas once Frozen v2 goes into production.

What does Frozen v2 mean for competitors like OpenAI and Anthropic?

It widens Google's structural cost advantage on Gemini inference by 6-10x — a gap that competitors can't close without similar model-specific silicon, which requires years of chip design and billions in fab capacity. OpenAI relies on Nvidia GPUs (rented from Microsoft Azure and its own datacenters); Anthropic relies on AWS Trainium and Nvidia. Neither has model-specific silicon in development that's publicly disclosed. Frozen v2 is Google's answer to the 'Google doesn't have a real moat' critique: a moat built in silicon. Long-term implication: Gemini inference cost falls dramatically starting 2028, potentially forcing Sol and Fable 5 pricing down or accelerating the shift to open-weight self-hosting for cost-sensitive workloads.

Quick Answer

What Is Google Frozen v2? Gemini Hardcoded in Silicon (Jul 2026)

Published: July 21, 2026

What Is Google Frozen v2? Gemini Hardcoded in Silicon (Jul 2026)

Bloomberg reported on July 20, 2026 that Google is developing an internal chip codenamed “Frozen v2” that hardcodes the Gemini architecture directly into silicon — delivering an estimated 6-10x improvement in tokens-per-watt over Google’s current TPU line.

This is a fundamentally different design philosophy than TPUs. Where TPUs are general-purpose matrix-multiply accelerators that can run any transformer, Frozen v2 is model-specific silicon that only runs Gemini. It’s a structural bet: give up flexibility, gain massive efficiency, and defend Gemini’s economics for the long term.

Last verified: July 21, 2026

The Core Idea: Hardware Specialization for a Single Model

Modern AI accelerators (Nvidia H200, AWS Trainium2, Google TPU 8t/8i) are general-purpose — they run any transformer model by executing generic matrix operations. This flexibility costs energy: every operation goes through the same pipeline whether it’s needed for the current model or not.

Frozen v2 flips this tradeoff:

The chip knows Gemini’s exact architecture.
Layer counts, attention head configurations, MoE routing patterns are baked into the transistor layout.
Operations that a TPU computes dynamically are hardwired connections.
Data movement paths are optimized for Gemini’s specific memory access patterns.

Result: 6-10x more tokens generated per watt. But the chip can only run Gemini (or Gemini derivatives with compatible architecture) — it can’t run Claude, GPT, Llama, or any other transformer.

Why “Frozen”?

The name is literal. Once the chip is fabricated, the model architecture is frozen into the silicon. You can’t change layer counts. You can’t add attention heads. You can’t swap in a new architecture.

Weights are still updatable — the model parameters flow through the chip during inference. But the structure (layer topology, attention pattern, MoE routing) is permanent.

This creates a dependency: Frozen v2 tied to a specific Gemini architecture family means Google can’t dramatically change Gemini’s structure without invalidating the chip investment. Bloomberg’s report notes engineers are still deciding how much to hardwire — the tradeoff is efficiency (more hardwired = more efficient) vs future flexibility (less hardwired = easier to iterate on the model).

Timeline and Availability

2026 (now): Design in progress. Bloomberg’s July 20 report is the first public confirmation.
2027: Expected tapeout and testing on early silicon.
2028: Target production deployment in Google datacenters.
Never: Public sale or rental. Frozen v2 is an internal Google chip — not offered on Google Cloud (unlike TPUs).

External impact: Third parties benefit indirectly via:

Cheaper Gemini API pricing starting 2028.
Larger free-tier quotas (Google can afford to give away more inference).
Faster Gemini responses (higher throughput per rack).

Frozen v2 vs Current AI Chips

Chip	Type	Runs	Efficiency signal
Nvidia H200	GPU (general-purpose)	Any model	Industry baseline
Nvidia B200	GPU (general-purpose)	Any model	~2x vs H200
AWS Trainium2	ASIC (transformer-optimized)	Any transformer	~1.5-2x vs H100 for supported models
Google TPU 8t	ASIC (transformer training)	Any transformer	Comparable to B200 for training
Google TPU 8i	ASIC (transformer inference)	Any transformer	Comparable to B200 for inference
Google Frozen v2 (2028)	Model-specific ASIC	Gemini only	6-10x vs TPU 8i on Gemini inference
Cerebras WSE-3	Wafer-scale (transformer-optimized)	Any transformer	Different economics, specialty workloads
Groq LPU	Language inference (deterministic)	Any transformer	Latency-optimized, not throughput-optimized

Key insight: Frozen v2 is the first widely-reported major example of model-specific silicon at frontier scale. If it works, it establishes a new competitive dimension in AI infrastructure.

The Strategic Bet

Google is making three bets with Frozen v2:

1. Gemini’s Architecture Will Remain Stable Long Enough to Justify Fabrication

Chip design cycles are 2-3 years. If Gemini’s architecture undergoes major changes (new attention mechanism, radically different MoE routing, non-transformer components), Frozen v2 becomes obsolete before it ships. Google is betting Gemini 3.5’s architecture is close enough to the multi-year winner that hardcoding it now is worth the risk.

2. Inference-Optimized Efficiency Beats Training Flexibility for Long-Term Economics

By 2028, most AI compute will go to inference (not training). Winning inference economics matters more than winning training flexibility. Frozen v2 optimizes for the winning workload of the future.

3. Vertical Integration Wins in AI

Google is the only major AI lab with a full stack: models (Gemini), datacenter chips (TPU + Frozen v2), datacenters (Google Cloud), and consumer distribution (Search, YouTube, Android). Frozen v2 deepens the moat. Competitors renting Nvidia GPUs from someone else’s cloud can’t replicate this cost structure.

What This Means for Non-Google Users

API pricing pressure starts 2028

Once Frozen v2 ships, Google can profitably run Gemini inference at prices competitors can’t match. Expect:

Gemini 4.x pricing to drop 30-50% year-over-year starting 2028.
Google to offer massive free tiers for developers (loss-leader to grow ecosystem).
GPT and Claude to face pricing pressure — either match Gemini or lose price-sensitive market share.

Open-weight self-hosting becomes more attractive by contrast

If Google is the only lab with model-specific silicon, then closed-model competitors (OpenAI, Anthropic) can’t match Gemini’s inference cost structure. Their alternative is charging premium pricing for premium capability — which pushes cost-conscious workloads to open-weight self-hosted models (DeepSeek V4, Kimi K3, Inkling).

Ironic outcome: Frozen v2 may accelerate open-weight adoption by making it the only cost-competitive alternative to Google’s Gemini moat.

The two-model-family future

By 2028-2029, the market likely bifurcates:

Google Gemini (proprietary, hardware-optimized, cheap at scale) for cost-sensitive workloads.
Open-weight (V4, K3, successors) for compliance/self-hosted workloads.
OpenAI Sol and Anthropic Fable as premium-tier flagships for hardest tasks.

Middle-tier proprietary models struggle. Meta abandoned Llama specifically because this dynamic was already visible in 2026.

Frozen v2 vs Frozen v1

Bloomberg’s report references “Frozen v2” — implying a v1 existed. Publicly disclosed details on v1 are limited. The most likely scenario: v1 was an earlier prototype or internal-only chip that demonstrated the model-specific silicon concept at smaller scale, and v2 is the production version.

Google’s chip disclosure culture is conservative — public details are limited until commercial deployment. Expect more concrete Frozen v2 details at Google Cloud Next 2027 or Google I/O 2028.

Risks and Skepticism

Frozen v2 could fail if:

Gemini architecture requires major redesign before 2028, invalidating the hardwired assumptions.
The 6-10x efficiency estimate proves optimistic in real deployment.
Manufacturing yield on model-specific chips is worse than expected.
Competitor open-weight models (V4, K3) improve fast enough to make Gemini’s cost advantage irrelevant for most workloads.

Broader skepticism: The “hardcode a model into silicon” idea has been tried before (Nervana, Wave, Cerebras variants). None have delivered the promised 10x efficiency at frontier scale. Frozen v2 requires Google to succeed where prior specialized-chip efforts stalled.

Bottom Line

Frozen v2 is Google’s most strategically significant chip bet in a decade. If it ships and delivers on the 6-10x efficiency promise, Google gains a structural inference cost advantage that competitors can’t match without similar model-specific silicon (and OpenAI, Anthropic, xAI don’t have such projects publicly disclosed).

For AI teams evaluating today: Frozen v2 doesn’t affect near-term decisions — it’s a 2028 chip. But it does inform long-term architecture strategy. If you’re betting the future of your product on proprietary Anthropic or OpenAI models, understand that Google is building the cost-structure moat that will make Gemini the cheap-and-good default by 2028-2029.

For infrastructure teams: watch competitor responses. If OpenAI or Anthropic announce model-specific silicon projects in the next 12 months, the chip specialization race is on. If they don’t, Google’s 2028-2029 competitive position becomes structurally advantaged in a way that’s hard to reverse.

Sources

Bloomberg Frozen v2 original report (via Global Banking and Finance): globalbankingandfinance.com
Quartz Frozen v2 analysis: qz.com/google-gemini-chip-frozen-tpu-efficiency-072026
Crypto Briefing Frozen v2 technical breakdown: cryptobriefing.com/google-frozen-v2-ai-chip-gemini
The News (Pakistan) Frozen v2 coverage: thenews.com.pk