In July 2026, multiple reputable outlets reported that Google is secretly developing a new dedicated server chip codenamed Frozen v2. The chip's core breakthrough is hardwiring parts of the Gemini large model's low-level computation logic into the hardware, greatly improving computational efficiency. According to reports, compared with existing TPUs, Frozen v2 could increase token processing capacity per unit of power by 6 to 10 times and is expected to be deployed beginning in 2028. The project reflects Google's strategic response to internal compute shortages and pressure to expand AI infrastructure.

Frozen v2 Core Principle: Architecture Hardening and Compute Optimization

The innovation of Frozen v2 lies in "hardwiring the Gemini architecture directly into silicon":

Architecture hardening mechanism Traditional AI chips (such as TPUs) are general-purpose processors that load model weights via software to perform computation. By contrast, Frozen v2 permanently "freezes" part of Gemini's core architecture (not the specific weights) into hardware circuits. This design greatly reduces inference computation steps and data-movement overhead.
Efficiency improvement logic
- Reduced data movement: Traditional inference requires frequent transfers between memory and compute units, while a hardened architecture can implement common computation paths in hardware, significantly lowering latency.
- Power optimization: Dedicated circuits enable high-frequency operations so more tokens can be processed per unit of power.
- Flexibility balance: Unlike earlier "frozen weights" approaches, Frozen v2 hardens the architectural blueprint, allowing future Gemini versions to update parameters while preserving the core framework, avoiding rapid obsolescence of the chip.

The project is reportedly rooted in an early concept by DeepMind chief scientist Jeff Dean. It remains in the experimental stage and is intended as a specialized supplement rather than a complete replacement for general-purpose TPUs.

Project Background: Compute Shortages Driving Hardware Innovation

The primary driver for Google's development of Frozen v2 is the increasingly pronounced internal imbalance between compute supply and demand:

Explosive AI demand: Wide adoption of models like Gemini has strained compute for Google Cloud and internal services, reportedly leading to refusals of some external customer workloads.
General-purpose chip bottlenecks: Existing TPUs, while powerful, still have room for efficiency gains in specific Gemini inference scenarios.
Full-stack optimization strategy: Google adheres to a "software-hardware integrated" approach, using custom chips to achieve software-hardware co-optimization and reduce total cost of ownership (TCO).

The project is a major component of Google's AI infrastructure investments. Alphabet plans to invest heavily in AI to maintain leadership against competitors such as OpenAI and Anthropic.

Potential Impact: Multi-dimensional Industry Effects

1. Impact on Google itself

Efficiency and cost optimization: Frozen v2 could significantly reduce Gemini inference energy consumption and latency, improving Google Cloud's service capability and profitability.
Strengthened competitive moat: A specialized chip would create a unique technological barrier, reinforcing the Gemini ecosystem advantage.
Risks and challenges: Specialization may reduce flexibility; if Gemini's architecture evolves rapidly, chip compatibility could become an issue. Development failure or delays would also affect the broader AI strategy.

2. Impact on the AI hardware industry