Naveen Rao's New Chip Startup Released Its First Model. The 1,000x Energy Claim Runs on a Simulation.

Naveen Rao has spent the last decade arguing that the way we build AI hardware is wrong. On Thursday his new company, Unconventional AI, put out the first thing you can actually run to test the idea: a model called Un-0 that generates images using coupled oscillators instead of the matrix multiplications that run on a GPU. The weights and training code are open. The headline number attached to it, a 1,000x cut in energy use, is not something Un-0 did. It is something the company hopes a chip it has not built yet will eventually do.

Both of those things are true at once, and keeping them straight is the whole story.

What Un-0 actually is

Rao ran AI at Databricks after it bought his last startup, MosaicML, and before that founded the chip company Nervana, which Intel acquired. So when he says the bottleneck for AI is energy, he has been circling the problem for a while. Unconventional AI, which TechCrunch reports has fewer than 50 employees, is his bet that you fix it by changing the computer, not the model.

Un-0 is an image generator. On the ImageNet 64x64 benchmark, its largest version (about 322 million parameters) reaches an FID of 6.74. FID measures how close generated images are to real ones, lower is better, and 6.74 is a respectable score. It is the quality early diffusion models and GANs hit when they were first published several years ago. The company is careful about this in its own write-up: "Un-0's quality matches where today's leading generative methods began." It trails current state of the art like EDM and GDD, and ImageNet 64x64 is a small, old benchmark, not the megapixel image generation people actually use.

So the interesting part is not the pictures. It is how the model produces them.

Computing with oscillators, translated once

Picture two metronomes on the same table. Through the shared surface they nudge each other until they either fall into the same rhythm or settle into opposite ones. Each metronome has a phase, the position of its arm in the swing. That coupled push and pull is the entire primitive Un-0 is built on, scaled from two oscillators to tens of thousands.

The math is the Kuramoto model, a 1970s description of how populations of oscillators synchronize. Each oscillator rotates at its own natural frequency and gets pulled toward or away from its neighbors by a coupling strength. In Un-0 that grid of coupling strengths is what the model learns during training. To make a picture, you set every oscillator to a random starting angle (the equivalent of the noise a diffusion model starts from), bias a small second group of oscillators toward a class like "volcano" or "daisy," let the system evolve under its own dynamics for a fixed time, then read out the final phases and hand them to a small conventional decoder that turns numbers into pixels.

That decoder is the catch worth watching, and to their credit Unconventional went after it directly. The decoder is an ordinary neural network, under 15 percent of the model's parameters. If it were quietly doing all the real work, the oscillators would just be expensive decoration. So the team ran ablations: train the decoder alone, freeze the oscillator dynamics at random values, vary how long the system runs. The dynamics measurably helped, and helped more the longer the system evolved. On the company's own evidence the oscillators are computing, not decorating. That is the genuinely good piece of engineering here, and it is checkable, because the code is public.

The number to interrogate

Now the 1,000x. The standard question for any benchmark is what it was measured against, and here the answer is nothing yet.

Un-0 does not run on an oscillator chip. It runs on a software simulation of one, and that simulation was trained on Nvidia's B200 GPUs, the most power-hungry datacenter silicon currently shipping. The largest ImageNet model took 640 B200-hours to train. The energy-efficient computer at the center of the pitch does not exist. The company says it will release chip schematics "soon" and eventually build an inference service out of its own hardware, with prompts going in one end and images coming out the other "at 1/1000 of power," in Rao's words. That is a roadmap, not a result.

This is where the press framing drifted ahead of the work. The version going around says Un-0 "performs just as well as state-of-the-art diffusion models." The company's own paper says it matches where those methods started, and trails where they are now. Both can be repeated honestly. Only one of them is the company's actual claim.

To be fair to Rao, he is not hiding the gap. "This is the 'hello world' of a new kind of computer," he told TechCrunch, which is roughly the right amount of caution for a model that proves a method works in simulation and nothing more. His larger argument is also hard to dismiss: "AI scaling is hard because of energy. It's going to be the fundamental limit in the next few years. You just can't go past it." Plenty of people building datacenters right now would agree with that sentence.

What to actually take from this

The science is real and the paper is honest about its limits. Coupled oscillators can be trained to generate images at a scale nobody had pushed them to before, the dynamics do measurable work, and you can download the whole thing and check. That is a legitimate result and a good day for the small research community working on physics-based computing.

The 1,000x energy claim is a thesis about hardware that has not been taped out, validated in silicon, or benchmarked against a GPU doing the same job. Analog and neuromorphic computing have a long history of demos that look great in simulation and run into noise, yield, and precision problems the moment they hit a real chip. Maybe Unconventional clears those hurdles. Until there is a chip and a measured joules-per-image number next to an Nvidia part running the same model, the efficiency figure stays in the same box as every other pre-silicon projection: interesting, unproven, and not the headline.

Watch for the schematics. That is when this gets real or stays a paper.