Google is building a chip with Gemini baked into the silicon

What happened

Google is developing a new type of AI chip that has its Gemini language model architecture built directly into the hardware. Unlike typical AI chips that run software models loaded onto them, this project, nicknamed “Frozen v2,” physically embeds Gemini’s design into the silicon. This approach was reported by The Information and confirmed by Reuters and Bloomberg Law.

Why it matters

Embedding the AI model directly into the chip shifts the traditional model-hardware relationship. It could make running Gemini faster and more energy efficient since the chip doesn’t just execute instructions for a model—it becomes the model. This could reduce latency and hardware overhead for complex AI workloads. For builders and businesses, this innovation translates to more tightly integrated AI performance optimized for Gemini’s specific architecture, rather than trading flexibility for speed on general-purpose chips.

However, it raises practical tradeoffs. The chip would be less adaptable to new or different AI models from other vendors, locking Google’s AI stack deeper into its custom silicon. This pressures competitors relying on generic chips and cloud providers that offer more flexible AI compute options. Investors and adopters need to watch if Google’s strategy forces a chip-model fusion trend or stays a niche approach due to cost or limited flexibility.

What to watch next

The key will be whether Google scales this chip beyond prototypes and integrates it into its cloud and device ecosystem where Gemini powers real AI services. Watch for performance benchmarks compared to current TPU or GPU solutions, adoption signals from Google Cloud customers, and any ripple effects on the AI chip market. If “Frozen v2” proves effective, expect other companies to explore embedding AI models in hardware for critical workloads, reconfiguring how AI infrastructure is built and priced.

AI Quick Briefs Editorial Desk