Mirror Review
July 21, 2026
Google is developing a specialized server chip, internally codenamed “Frozen v2,” to run its Gemini AI models with vastly improved efficiency.
According to reports from The Information, the hardware hardwires parts of Gemini’s architecture directly into silicon.
Engineers project that Google’s Frozen v2 chip will serve 6 to 10 times more tokens per unit of power than current Tensor Processing Units (TPUs).
Target deployment is set for 2028 as an addition to Google’s hardware lineup. This innovation aims to alleviate severe compute shortages across Google Cloud’s infrastructure amid rising AI demand.
What Is Google’s New Chip Strategy?
Google plans a new ‘frozen’ chip to run its AI models much more efficiently by embedding core software instructions into hardware components. Traditional accelerators remain programmable to run any model architecture. In contrast, Google’s frozen v2 chip locks specific neural network logic directly into the silicon logic.
This direct hardware integration drastically cuts down computational steps. It minimizes data movement between processors and memory modules during inference.
By turning software design into physical circuits, the system achieves unprecedented throughput without drawing excess power.
“Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers… By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads.”
— Google Cloud Official Statement
How Does Frozen v2 Compare to Traditional TPUs?
The new chip project does not aim to replace Google’s established Tensor Processing Units (TPUs). Instead, Google’s new chip serves as a specialized parallel line focused entirely on efficient model serving.
| Feature | Standard Google TPUs | Google’s Frozen v2 Chip |
| Primary Design | General-purpose AI accelerator | Architecture-specific application chip |
| Flexibility | Runs diverse models and software updates | Tailored specifically for Gemini models |
| Power Efficiency | Standard baseline power consumption | 6x to 10x higher tokens served per watt |
| Deployment Role | Large-scale training and broad inference | Dedicated high-efficiency inference serving |
While TPUs handle varying workload demands and general training runs, Google’s frozen v2 chip targets maximum output for fixed production environments.
Why Is Google Facing an AI Capacity Crunch?
Google faces unprecedented demand for computational power across its search products, Gemini subscriptions, and enterprise services.
Data center power limits and hardware supply boundaries have created internal operational bottlenecks.
- Declined Customer Deals: Capacity limitations forced Google Cloud to turn away prospective enterprise clients seeking large compute capacity.
- Third-Party Infrastructure Spending: To bridge immediate gaps, Google agreed to pay SpaceX nearly $1 billion a month for complementary infrastructure support.
- Competitor Gains: Chinese frontier models from teams like Moonshot AI and Alibaba have captured market share, currently accounting for 45% of U.S. company token usage.
- Internal Launch Delays: Reports indicate delays in upcoming Gemini models as teams refine performance and coding capabilities amidst compute constraints.
What Are the Risks of Google’s Frozen v2 Chip?
Developing a dedicated chip with hardcoded software logic carries distinct operational risks alongside its cost benefits.
- Reduced Architecture Flexibility: If researchers discover a vastly superior neural network design, hardwired chips cannot adapt to the new format.
- Obsolescence Risks: Fast updates in artificial intelligence can make fixed silicon designs outdated before large-scale deployment.
- Weight Iteration Support: Despite hardcoded pathways, engineers confirm the chip still supports updating model weights, preserving core tuning abilities.
Because of these trade-offs, Google currently treats the development as a specialized technology test platform rather than a complete replacement for broad hardware investments.
How Does Frozen v2 Fit Into Google’s Long-Term Roadmap?
The Frozen v2 chip release timeline targets deployment around 2028 as engineers finalize design details and choose which neural structures to freeze.
Markets responded favorably to the hardware strategic shift. Alphabet shares rose over 1.5% during regular trading and spiked up to 3.3% in early activity following the report.
At the same time, industry leadership continues addressing broader policy frameworks. Google DeepMind Chief Executive Demis Hassabis recently engaged lawmakers regarding federal safety oversight standards for national security risks in frontier AI development.
Maria Isabel Rodrigues
