1,000x Less Power: A New Kind of AI Computer

Naveen Rao, who ran AI at Databricks for years, has started a company called Unconventional AI and unveiled its first working demo: a system called Un0 that generates images using an oscillator-based computer architecture. The headline claim is a 1,000-fold reduction in power consumption for AI inference — running a model to produce an answer — compared to today's GPU-based systems.

The demo works in software simulation. No actual chip exists yet; Rao said the team plans to release chip schematics publicly soon. But the image-generation results are real, and he says Un0 performs comparably to established diffusion models. The company has fewer than 50 employees.

Why energy is now the ceiling for AI

Every few months, the AI industry announces a new generation of models more capable than the last — and each one needs more compute to train and run. That compute needs electricity, and electricity is becoming the limiting factor. Data centers are straining power grids in Virginia, Texas, and Ireland. Rao made the constraint explicit: 'AI scaling is hard because of energy — it's an energy limited problem.' The question isn't whether this wall exists; it's how quickly AI labs run into it.

Unconventional AI's answer is to rethink the chip itself. Standard processors and GPUs use transistors that flip between two states — off and on. Oscillator-based architectures work differently: they use continuously cycling signals that can carry more information with far less switching energy. Rao built a software simulation of this architecture and trained Un0 on it. 'This is the hello world of a new kind of computer,' he said.

What builders should watch

A 1,000x efficiency claim is extraordinary, and hardware roadmaps almost always slip. But the logic behind the bet is solid: the companies most likely to feel AI energy costs first aren't the hyperscalers with dedicated power contracts — they're the builders and startups paying per-token API bills that keep growing regardless of what they build. If oscillator chips eventually move from schematics to silicon, the first application would almost certainly be inference: the part that runs every single time someone sends a message to a model.

For context, one of the main reasons inference APIs have been getting cheaper over the past two years is that chip efficiency keeps improving incrementally — a few percent here, a new architecture variant there. A genuine order-of-magnitude leap, let alone three orders of magnitude, would change the math so completely that it's worth following even at the simulation stage.