Unconventional AI, led by Naveen Rao who previously headed AI at Databricks, says it can drastically lower the energy required for AI inference by rebuilding computing architecture around oscillator-based hardware.
Un-0: first model demonstrated in software simulation
On Thursday the company unveiled its first model, called Un-0, an image-generation system that demonstrates how the company's architecture can reproduce conventional AI system behavior. In an accompanying research paper, Unconventional's team explains they constructed a fully functional image-generation model using a software simulation of their new architecture, and that the model performs comparably to state-of-the-art diffusion approaches.
Rao told TechCrunch, “This is the ‘hello world’ of a new kind of computer. Over the next year, you’re going to start seeing some pretty interesting news around this.”
How the oscillator-based approach differs
Un-0's outputs are similar to those from models such as Stable Diffusion or OpenAI’s GPT Image 1, but the underlying computation is fundamentally different. The system is built on an oscillator-based architecture rather than the conventional chips that power today’s models. Rao says the benefits of oscillator-based computing are complex, but that they could ultimately reduce power use for inference by as much as 1,000 times.
Hardware and infrastructure are still under development
Currently, Un-0 runs on a software simulation of Unconventional’s oscillator chips; the company plans to publish schematics for a physical chip in the near future. Their longer-term plan is to construct a complete inference stack from the ground up: producing chips, assembling systems that run AI models, and offering compute capacity similarly to other providers.
“We will build a new kind of system composed of our chips,” Rao says. “We will run AI models there, and we will have a network cable where prompts come in and inferences go out, but it’ll be done at 1/1000 of power.”
Ambitious goals from a small team
The objective is highly ambitious, especially for a company that currently employs fewer than 50 people. Given the scale of AI deployment and the expected cost of meeting growing inference demand, Unconventional's approach may be one of the few aiming to address that scale of the problem.
Rao sees available power supply as a hard limit for AI growth: “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. It’s going to be an energy-limited problem, at the end of the day.”
The company’s next steps include releasing chip schematics and building the full inference infrastructure; for now, the Un-0 demonstration serves as a software-simulated proof of concept for the oscillator-based approach.
