Google Reportedly Developing Frozen v2 Chip to Boost Gemini Model Efficiency

Google is reportedly developing a specialized server chip code-named "Frozen v2" that hardcodes the underlying computational architecture of its Gemini models directly into silicon. The chip, targeted for deployment by 2028, aims to deliver 6 to 10 times the energy efficiency of Google's current AI chips during inference. This represents a major shift toward hardware-software co-design, addressing Google's severe AI compute shortages by optimizing hardware specifically for its flagship models. If successful, it could drastically reduce operational costs and latency for Gemini-powered applications, though it binds Google to a specific model architecture. Unlike its predecessor which failed because it etched entire model weights into silicon, Frozen v2 only hardcodes the architectural logic, allowing model weights to be updated. However, the chip will not replace Google's general-purpose TPUs and is currently treated as a low-volume technology testbed.