AI’s next constraint may not be model ideas, data, or demand. Naveen Rao, formerly the head of AI at Databricks, argues that energy will become one of the defining limits for AI systems, especially as inference demand grows.
His company, Unconventional AI, is trying to answer that problem with a computing architecture that looks very different from the chips behind conventional computing and traditional LLMs. Its first public proof point is Un-0, an image-generation model that runs on a software simulation of oscillator-based chips.
A different route to AI inference
Unconventional AI’s core claim is straightforward but ambitious: inference processing can become vastly more power efficient if the underlying computer architecture is rebuilt. The company is not presenting Un-0 as just another image-generation product. It is using the model to show that its architecture can reproduce the behavior of familiar AI systems.
Un-0 is described as an image-generation system tool. According to the source article, its output is similar to image-generation models like Stable Diffusion or OpenAI’s GPT Image 1. The important distinction is not only what the model produces, but how it produces it.
The company’s research team says it built a fully functional image-generation model using a software simulation of the new architecture. That model, according to the article, performs just as well as state-of-the-art diffusion models.
Rao described the release in foundational terms: “This is the ‘hello world’ of a new kind of computer,” Rao told TechCrunch. “Over the next year, you’re going to start seeing some pretty interesting news around this.”
Why oscillator-based computing matters here
The architecture behind Un-0 is oscillator-based computing. The source article does not describe the full technical mechanism, but it makes clear that this design is completely different from the chips used in conventional computing and traditional LLMs.
That difference is the reason Unconventional AI sees a path to much lower power use. Rao believes the approach could ultimately reduce power use by as much as 1,000 times.
For AI infrastructure, that claim matters because inference is the work of running models after they have been built. When people send prompts and receive generated outputs, inference systems do the processing. If demand for those outputs keeps growing, the power cost of inference becomes a larger part of the AI buildout.
Unconventional AI is trying to address that cost at the level of computing architecture, rather than only optimizing software around existing chips. That is why Un-0 is positioned as a proof of concept for a broader system, not merely as a standalone model launch.
The hardware is not here yet
The current version of Un-0 does not run on finished Unconventional AI chips. It runs on a software simulation of the company’s oscillator chips. That distinction is important because the central energy-efficiency promise depends on hardware that is still being developed.
The company plans to release schematics for an actual chip soon. After that, its stated goal is to build an entire inference stack from the ground up.
That stack would eventually let Unconventional AI provide compute capacity in the way other providers do. Rao described the intended system this way: “We will build a new kind of system composed of our chips,” says Rao. “We will run AI models there, and we will have a network cable where p rompts come in and inferences go out, but it’ll be done at 1/1000 of power.”
The path from simulation to real infrastructure remains the central question. The company has shown an image-generation model on a simulated architecture, but the broader plan requires chips, systems, and operational capacity.
A small company targeting a large bottleneck
The scale of the goal stands out because Unconventional AI still counts less than 50 employees. It is attempting to rethink a core part of AI infrastructure while the wider market continues to build around conventional compute systems.
That makes the project risky and ambitious, but the source article frames it against a real pressure point: the anticipated cost of meeting growing demand for inference. If AI systems keep expanding, power supply may become a hard limit.
Rao is direct about that concern. “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,” he says.
Un-0, then, is best understood as an opening demonstration. It shows that Unconventional AI’s simulated oscillator-based architecture can support an image-generation model with output comparable to well-known diffusion systems. The larger promise is that a future chip-based version could make AI inference far less power intensive.
What to watch next
The next steps are practical. Unconventional AI needs to move from software simulation toward actual chip schematics, then toward systems that can run AI models and serve inference requests.
The key signals to watch are:
- whether the company releases schematics for an actual oscillator chip;
- whether the hardware can support the same kind of AI inference shown in simulation;
- whether the claimed 1,000 times power reduction can be demonstrated in real systems;
- whether Unconventional AI can build a full inference stack around its chips.
For now, Un-0 is a first public marker. It does not prove that AI’s power problem is solved, but it does show how one company is trying to solve it by changing the computer itself.