A startup bets oscillator chips can cut AI inference power 1,000x

Naveen Rao, who ran AI at Databricks, has started a company called Unconventional AI with a bet that the chips running today's models are the wrong design. His pitch is oscillator-based computing, an architecture that does inference with networks of coupled oscillators rather than the silicon logic that powers current GPUs. Rao claims it can cut the power a model burns during inference by up to 1,000 times. TechCrunch reports the company has fewer than 50 people.

The proof so far is a model, not a chip. Unconventional released Un-0, an image generator that Rao says performs on par with established diffusion models like Stable Diffusion. The catch is that Un-0 runs on a software simulation of the oscillator hardware, not the hardware itself, which does not exist yet. Rao says the schematics for the actual chips are coming soon.

A 1,000x claim from a company still running in simulation deserves caution. Inference power is where the money goes once a model is deployed, so the prize is real, and Rao has built large AI systems before. But simulation results and silicon that ships are different things, and the gap between them is where most novel-hardware bets fail.

Why it matters

If you pay for inference at scale, the cost of serving a model, not training it, is what dominates your bill over time. A working oscillator chip would change that math, but nothing here has run on real hardware yet, so the thing to watch is whether the promised schematics turn into silicon that matches the simulation.