Google's Frozen v2 Chip Hardens Gemini Architecture into Silicon

According to The Information, Google is progressing on a dedicated server chip called Frozen v2, which hardens Gemini's model architecture directly into hardware, unlike TPUs that support various models. Unlike the original Frozen approach that fixed model weights, v2 only fixes the architecture (the model blueprint), allowing weights to be updated for greater flexibility. The chip is reportedly 6 to 10 times more efficient than current TPUs for AI inference, with deployment planned for 2028 but initially in small volumes as an experiment in specialized chips. Frozen v2 will not be sold externally but aims to ease Google's internal AI compute crunch. If performance meets targets, it will become a key competitive advantage: lower inference costs enable cheaper large model services, stealing market share from OpenAI and Anthropic. For developers using Gemini API, this means potentially lower prices or higher performance inference in the future, though no immediate changes to API usage.