Google embeds Gemini AI directly into new processor chip

Most AI accelerator chips are designed as blank slates. You load a model onto them, and they execute the instructions. Google is reportedly taking a radically different approach: building a processor where the AI model is not loaded onto the chip, but is instead etched directly into the hardware itself. This means the chip would effectively be the model, with Gemini’s neural network architecture permanently embedded at the silicon level.

The project, internally code-named “Frozen v2,” was first reported by The Information before being picked up by Reuters and Bloomberg Law. The name suggests a successor to an earlier, unannounced effort, and it signals Google’s ambition to move beyond general-purpose AI hardware toward something far more specialized. Instead of running a flexible model on a flexible chip, this design locks in Gemini’s specific structure, potentially offering massive gains in efficiency, speed, and power consumption.

The implications are significant. If Google succeeds, it could create a processor that is essentially Gemini in silicon, bypassing the overhead of loading and managing software models. This would reduce latency dramatically and could make on-device AI far more capable, especially for tasks like real-time translation, image recognition, and voice processing. It also raises questions about flexibility: once the model is frozen in hardware, updating it would require a new chip, not just a software patch.

Alphabet has not confirmed the project publicly, but the reporting suggests the company is investing heavily in custom silicon as it races to dominate the AI hardware market. With competitors like NVIDIA, AMD, and Amazon all pushing their own chips, Google’s “Frozen v2” could represent a bet on hardwired AI intelligence over flexible, software-driven models.