Google's 'Frozen v2' Chip: How This Custom AI Silicon Aims to Supercharge Gemini by 2028 and Challenge Nvidia

Google is developing a new custom AI chip, internally codenamed "Frozen v2," to significantly enhance the efficiency of its Gemini AI models by 2028. This strategic initiative embeds parts of the Gemini model architecture directly into silicon, with engineers projecting a 6 to 10 times improvement in tokens per unit of power compared to current Tensor Processing Units (TPUs), aiming to reduce Google's dependence on Nvidia and manage infrastructure costs.

The Strategic Imperative Behind Frozen v2

The development of "Frozen v2" underscores Google's long-term strategy to optimize its AI infrastructure. By integrating specific elements of the Gemini model directly into the chip's design, Google seeks to dramatically reduce the computational load and data movement required for Gemini queries. This specialization is a direct response to the escalating demands and costs associated with large-scale AI model deployment and inference.

This custom chip development is also a clear signal of Google's intent to lessen its reliance on external hardware providers like Nvidia. While Nvidia currently dominates the AI chip market, Google has been investing in its own custom AI chip solutions for years, with its Tensor Processing Units (TPUs) evolving through multiple generations. The "Frozen v2" project represents a deeper integration of software and hardware, tailored specifically for future Gemini models.

Projected Efficiency Gains and Trade-offs

Google engineers anticipate that the "Frozen v2" chip will deliver substantial efficiency improvements, projecting 6 to 10 times more tokens per unit of power than existing TPUs. This efficiency gain is critical for managing the operational costs of running advanced AI models like Gemini at scale within Google Cloud. Such improvements could translate into more cost-effective AI services and potentially faster response times for users.

However, this specialization comes with a trade-off: the "Frozen v2" chip would only be compatible with future Gemini models that share the same embedded architecture. This design choice sacrifices flexibility for optimized performance, meaning the chip would not be suitable for other AI models or general-purpose computing tasks. This highlights a strategic decision to prioritize deep integration for its flagship AI product.

Timeline and Market Reaction

The projected 2028 timeline for the "Frozen v2" chip's deployment indicates the significant development and integration work still required. This timeframe also suggests the competitive landscape Google navigates, particularly with ongoing advancements from competitors and the reported delay of the next Gemini Pro release. The long lead time underscores the complexity of designing and deploying highly specialized silicon.

Following the news of the chip's development, Alphabet shares reportedly rose by approximately 3% on Monday. This market reaction suggests investor confidence in Google's long-term strategy to control its AI hardware destiny and potentially improve its cost structure for AI operations.

Google's History with Custom AI Silicon

Google has a history of developing custom AI chips, starting with its original Tensor Processing Unit (TPU) designed to accelerate its own AI services, including image recognition and machine translation. The company later introduced Cloud TPUs, making these specialized processors available through Google Cloud for both training and inference of neural networks. More recently, Google Cloud announced its eighth generation of TPUs, splitting them into the TPU 8t for model training and the TPU 8i for inference, further demonstrating its commitment to specialized AI hardware.

Conclusion

Google's development of the "Frozen v2" chip represents a significant investment in its proprietary AI infrastructure, aiming for substantial efficiency gains for its Gemini models by 2028. By embedding Gemini's architecture directly into silicon, Google seeks to reduce its dependence on external hardware and manage the escalating costs of AI operations. While this specialization offers considerable performance benefits, it also introduces a degree of inflexibility, limiting the chip's use to compatible Gemini models. The market's positive response to this announcement reflects the strategic importance of controlling key AI hardware components in the competitive artificial intelligence landscape.

Sources

• https://techcrunch.com/2026/07/20/google-is-working-on-a-new-ai-chip-designed-to-make-gemini-more-efficient/
• Google Rattles the Tech World With a New AI Chip for All | WIRED
• Google Cloud launches two new AI chips to compete with Nvidia | TechCrunch
• https://www.cnbc.com/2026/07/20/alphabet-googl-stock-ai-chip-report.html