Inside Google's 'Frozen v2' Chip: Why It's Hardwiring Gemini Directly Into Silicon

The efficiency gain is the headline: Google engineers reportedly project Frozen v2 could deliver 6 to 10 times more AI tokens per unit of power consumption than the latest TPU chips .

Deployment Timeline

According to The Information report and corroborating coverage, Frozen v2 is targeted for deployment as early as 2028 . This means the chip is still years from production and will only enter service after the current eighth-generation TPU family (Sunfish and Zebrafish) has been deployed and matured.

The Taalas Precedent — What Hardwiring a Model Actually Looks Like

The startup Taalas shows exactly what this tradeoff means in practice. In February 2026, Taalas unveiled its HC1 chip, which encodes the entire Llama 3.1 8B model — every parameter — directly into the chip's transistor circuits using mask ROM, with no external HBM for weight storage .

The numbers are stunning:

• ~17,000 tokens/second on Llama 3.1 8B
• About 48x the throughput of an Nvidia B200
• Roughly 10x faster than Cerebras
• Cost claimed at 20x cheaper per inference than equivalent GPU setups
• Inference cost as low as 0.75 cents per million tokens

The downside is equally dramatic:

• Model lock-in. The chip runs exactly one model architecture. If the model is updated (new attention mechanism, different layer count), the chip is unusable .
• No weight updates. The parameters are physically encoded; fine-tuning, LoRA adapters, or post-training updates require a new chip mask .
• Zero multi-model serving. A datacenter with HC1 can serve only Llama 3.1 8B — no other LLM .

Taalas's founder described the architecture as pairing a "mask ROM recall fabric" with an SRAM recall fabric to store both the model and compute all KV cache operations on-chip, eliminating external memory movement entirely . This is the same fundamental approach Frozen v2 would use.

The Infrastructure Crisis Driving Frozen v2

Google's AI compute capacity is so strained that it has begun rationing access even to paying giants. Several data points make the motivation clear:

The Meta capacity denial (March 2026): Google told Meta in March 2026 that it could not supply all the Gemini computing capacity Meta wanted to purchase . The shortfall delayed some of Meta's internal AI projects and forced Meta to instruct its engineers to conserve AI tokens .

Backlog figures: Google Cloud's order backlog nearly doubled to $462 billion in Q1 2026, representing demand roughly 23 times its available processing capacity . CEO Sundar Pichai confirmed 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 that demand" .

Broader capacity squeeze: Google is even leasing approximately $920 million per month in Nvidia GPUs from SpaceX to bridge its compute gap . Google's cloud VP Amin Vahdat said in November 2025 that the company would need to double AI capacity every six months to meet demand .

These constraints directly motivate Frozen v2: if Google can get 6–10x more Gemini inference per watt, it effectively multiplies capacity without needing new data centers.

Other AI Chip Initiatives Google Is Pursuing

Frozen v2 is one piece of a much broader hardware strategy:

1. Eighth-generation TPUs split into training and inference chips. At Cloud Next 2026, Google previewed the eighth-generation TPU family, split for the first time into two purpose-built variants :

• TPU 8t "Sunfish" — training-focused, designed by Broadcom, built on TSMC N2 with CoWoS packaging, targeting second half of 2027
• TPU 8i "Zebrafish" — inference-focused, designed with MediaTek leading the I/O die

2. Broadcom long-term agreement through 2031. Broadcom filed an SEC 8-K disclosing a Long Term Agreement to develop and supply custom TPUs for Google's future generations, plus a Supply Assurance Agreement for networking components through 2031 . Separately, Anthropic signed a deal for ~3.5 GW of Google TPU capacity, coming online from 2027 onward .

3. TPU rentals to customers including OpenAI. Reports indicate Google is increasingly offering TPU capacity as a rental product to external AI companies, with OpenAI named as a customer .

4. "Icefish" project with MediaTek. The "Icefish" codename appears linked to MediaTek's involvement on a Google-specific custom chip project beyond TPU 8i — possibly a lower-power edge or inference accelerator .

5. Discussions with Intel. Google has reportedly held discussions with Intel about custom AI chip designs, though no formal agreement has been announced .

6. Ironwood (seventh-generation TPU) generally available. Ironwood was made generally available to cloud customers in April 2026 . Built on a 3nm process with a dual-chiplet architecture, it was optimized for MoE models powering the Gemini ecosystem .

How Frozen v2 Complements, Not Replaces, the TPU Lineup

The Information's report explicitly states Frozen v2 is designed to complement, not replace, Google's TPU roadmap . The logic is straightforward:

• TPUs remain the general-purpose AI compute backbone. They handle training, multi-model serving, fine-tuning, and inference for arbitrary customer models. They are Google Cloud's main product for external AI customers .
• Frozen v2 is a hyper-specialized accelerator for Google's own high-volume Gemini inference. It is essentially a "Gemini appliance" that offloads the most predictable, highest-volume inference traffic from TPUs, freeing TPU capacity for more diverse workloads and external customers .
• No other model, no training, no flexibility. Frozen v2 cannot train models, cannot run a competitor's LLM, and cannot adapt to Gemini architecture changes. It occupies a narrow but massive-volume niche .
• TPU 8i (Zebrafish) and Frozen v2 coexist. The inference-focused TPU 8i handles the broad range of inference workloads. Once Gemini's architecture stabilizes and volume justifies a fixed chip, Frozen v2 serves that specific tail at radically lower power .

This is analogous to how hyperscalers use both general-purpose CPUs for most workloads and fixed-function ASICs for specific high-volume tasks (e.g., video transcoding, networking offload). Frozen v2 is the AI equivalent: a purpose-built inference engine that sits alongside the programmable TPU fleet.