The former AI head at Databricks believes he can cut his AI energy bill by 1,000 times

The drive to discover the next big thing in AI has led to funding for some very ambitious projects — but one company is using it as an opportunity to rebuild computing architecture from the ground up.

Led by Naveen Rao, former Head of AI at Databricks. Unconventional artificial intelligence It promises to make inference processing significantly more energy efficient. The secret weapon: a new type of oscillator-based computer architecture.

The company on Thursday released its first AI prototype — called Un-0 — an image-generating system tool that shows for the first time how the company’s technology can replicate traditional AI systems. In an accompanying new paper, the company’s research team demonstrates how to build a complete image generation model using a software simulation of the new structure — a model that performs as well as modern diffusion models.

“This is a welcoming world for a new kind of computer,” Rao told TechCrunch. “Over the next year, you’ll start to see some interesting news about this.”

The output of the new Un-0 model is similar to the output of image generation models such as Stable Diffusion or GPT Image 1 from OpenAI. The impressive part is how he reached this performance. The model is built on an oscillator-based architecture that is very different from the chips that power traditional computing and traditional LLMs. The benefits of oscillator-based computing are complex, but Rao believes it will ultimately reduce energy use by up to 1,000 times.

Much of the infrastructure to get there is still under construction. The current version of Un-0 runs a software simulation of unconventional oscillator chips, but the company plans to release schematics of an actual chip soon. From here, the plan is to build a complete inferences suite from the ground up, with non-traditional AI eventually providing the computing power just like any other provider.

“We will build a new type of system composed of our chips,” says Rao. “We’ll be running AI models there, and we’ll have a network cable where the prompts come in and the inferences go out, but it will be done at 1/1,000 power.”

It’s an amazingly ambitious goal, especially for a company that still has fewer than 50 employees. But given the scale of AI construction and the expected cost of meeting the growing demand for inference, this may be one of the few efforts to meet the scale of the problem. As Rao sees it, available energy supplies will be one of the tough constraints AI faces in the coming years – and the unconventional project is one of the few capable of addressing it.

“Scaling up AI is difficult because of energy,” he says. “It will be the fundamental limit in the next few years. You can’t go beyond it. It will be an energy-limited problem, eventually.”