Google's "Frozen v2" chip reportedly bakes Gemini's architecture directly into silicon for efficiency gains

2026-07-20 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Cloud Computing & IT Infrastructure · Depth: Intermediate, quick

Summary

Google is reportedly developing an internal server chip named "Frozen v2," designed to integrate the Gemini AI model's architecture directly into its silicon. This specialized chip is projected to be 6 to 10 times more efficient at serving AI responses compared to Google's existing TPU chips. Scheduled for deployment starting in 2028, Frozen v2 represents a strategic test run for highly specialized hardware, with a smaller production scale than the general-purpose TPU line. Unlike TPUs, which support various models, Frozen v2 embeds specific structural components of the Gemini model, allowing new weights to be loaded while the core architecture remains fixed. This approach, evolving from an earlier concept to embed model weights, aims to reduce compute steps and accelerate response times, ultimately enhancing Google's internal AI compute capacity and potentially improving inference cost margins against competitors like OpenAI and Anthropic.

Key takeaway

For AI Architects evaluating future infrastructure investments, Google's Frozen v2 initiative signals a shift towards highly specialized, model-architecture-embedded silicon. You should anticipate increased pressure to optimize inference costs and explore custom hardware solutions for your core AI models. This trend suggests that general-purpose accelerators may become less competitive for high-volume, specific model deployments, urging you to consider long-term architectural stability in your model development.

Key insights

Google's "Frozen v2" chip hardcodes Gemini's architecture for 6-10x AI inference efficiency, deploying 2028.

Principles

• Embedding architecture boosts AI inference efficiency.
• Specialized hardware can optimize specific model structures.
• Flexibility in weights is crucial for chip longevity.

In practice

• Optimize inference costs for competitive advantage.
• Consider specialized hardware for core AI models.
• Prioritize architectural flexibility over fixed weights.

Topics

• AI Hardware
• Google Gemini
• Custom Silicon
• AI Inference
• Chip Architecture
• TPU

Best for: Investor, CTO, VP of Engineering/Data, AI Hardware Engineer, AI Architect, Director of AI/ML

Related on AIssential

• Google Plans New ‘Frozen’ Chip to Run Its AI Models Much More Efficiently — Google's "Frozen v2" chip directly integrates the Gemini AI model blueprint for 6-10x greater efficiency.
• Google reportedly develops Frozen v2 chip for Gemini AI — Google is developing a highly efficient custom AI chip, Frozen v2, reflecting a broader industry shift towards proprietary silicon for AI.
• Google is working on a new AI chip designed to make Gemini more efficient — AI companies are developing custom chips to boost model efficiency and reduce reliance on external hardware providers.
• Rebellions Bets on Memory-Centric Architecture as It Weighs IPO Options — Memory-centric AI architectures and supply chain security are critical for large-scale LLM inference and market traction.
• Alphabet Stock Jumps After Information Report of New Google Chip — Google is developing "Frozen v2," a new AI chip 6-10x more efficient, integrating Gemini's blueprint.
• Partnering with Etched: Building the Inference Machine — Etched is building production-ready, cluster-scale custom silicon for AI inference, achieving superior throughput and interactivity.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.