Google just bet its inference future on a chip built for one model

Google has apparently decided that general-purpose AI chips are a phase, not an endpoint. According to The Information, the company is developing a project internally known as Frozen v2 — a chip that bakes portions of Gemini's architecture directly into the silicon while leaving the model's weights updatable. It's a narrower, more permanent bet than anything Google has shipped before, and it signals just how badly the industry wants inference costs to come down.

The logic here isn't new, even if the execution is. Bitcoin mining went from CPUs to GPUs to ASICs once the workload stopped changing enough to justify flexibility, and inference is starting to look the same way. Training a model benefits from GPUs that can bend to whatever architecture researchers dream up next. But once a model is locked and shipped to a billion users, all that flexibility becomes waste — silicon spending cycles on capabilities nobody in production actually needs. Google's internal projections reportedly put Frozen v2 at six to ten times more tokens per watt than its current TPUs, which is a big enough number to explain why anyone would take on the risk of designing hardware around a single, evolving model.

Google has apparently tried this before and backfired. An earlier effort led by DeepMind chief scientist Jeff Dean reportedly wanted to etch Gemini's actual weights into the chip. That's the purest version of the idea, and it died for an obvious reason: Gemini gets updated constantly, and a chip that only understands one frozen snapshot of the model becomes e-waste the moment a new version ships. Frozen v2 is the compromise — freeze the architecture, keep the weights flexible — which buys Google efficiency without burning the hardware's shelf life every time Gemini gets a refresh.

Google isn't alone in chasing this. Nvidia, of all companies, struck a reported $20 billion licensing deal with Groq last year, hedging against its own GPU dominance. Taalas has gone further than anyone, fusing an entire 8-billion-parameter Llama model onto a chip and claiming around 17,000 tokens per second by never shuttling data off-die. D-Matrix and SambaNova are attacking the same problem from the memory side, rethinking how close compute sits to data rather than hardwiring a specific model. Frozen v2 sits in between all of that — not as extreme as Taalas, not as general as a TPU, but built explicitly around Gemini's shape.

None of this is officially confirmed. Google's public statement is the usual corporate fog about 'full stack' co-design and 'rigorous exploration,' which tells you almost nothing except that the project is real enough to warrant a canned response. If it does ship, the interesting part isn't the efficiency number — it's what it implies for developers building on Gemini: a future where the model and the hardware underneath it are so tightly fused that swapping either one becomes a much bigger decision than it used to be.