The news comes from The Information and was picked up by Reuters and Bloomberg Law: Google is working on a chip that doesn't just accelerate an LLM, but embodies it. The project, informally called "Frozen v2," would see Gemini's architecture etched into silicon, turning the hardware into a physical twin of the model.

It's not Google's first custom chip: TPUs (Tensor Processing Units) have been around for years, designed for large-scale inference. But TPUs remain general-purpose programmable platforms onto which you load the model you need each time. What's described here is an ASIC where the network weights are fixed at manufacturing time – a slab of silicon that does one thing only: run Gemini as-is, with no intermediate software layers.

Why do it? The answer lies in the economics of massive-scale inference. When serving billions of requests a day, every milliwatt and nanosecond counts. A dedicated chip eliminates the overhead of model loading, reduces data transfers between memory and compute, and allows the circuit layout to be optimized precisely for Gemini's computation graph. Essentially, it pushes hardware-software co-design concepts already guiding TPU design to their logical extreme.

But the cost is zero flexibility. "Frozen" isn't a whim: the model is locked in its current version. No fine-tuning, no incremental updates, no switching to a more capable variant. If Gemini receives an architectural improvement, the chip becomes obsolete. This makes it suitable only for scenarios where the model version is stable and a long operational lifetime is expected – for example, the backend of Google's chatbot Bard (or whatever it will be called) or some very long-duration API service.

For on-premise deployment watchers, the structural signal is the most interesting part. If Google thinks it's worth building an ASIC for a single model, it means the inference volumes are large enough to justify a huge design and production investment. And if the model is baked into hardware, Google's cloud infrastructure becomes even more efficient, widening the cost gap with those using general-purpose GPUs. For organizations evaluating self-hosted LLMs, the question becomes: when – if ever – will similar chips for open models become available, ready to be placed in one's own racks? Today it's science fiction, but the trend points toward ever more specialized hardware that could make local inference economically unbeatable.

There's also a geopolitical and data sovereignty angle. A chip of this kind, produced and controlled by a single cloud vendor, reinforces dependency on Google for anyone using Gemini via API. Conversely, if the concept were replicated for open-source models (think Llama or Mistral), it could enable completely autonomous inference nodes, free of software dependencies – almost AI appliances.

For now, Google has not officially commented. The "v2" suffix suggests there was already a first iteration, perhaps never out of the lab. And the choice of such a transparent codename – "Frozen" – seems a precise message: we are crystallizing Gemini in hardware, and it's a calculated bet that the current version is already mature enough to justify the cost of a dedicated wafer.