Databricks' Former AI Chief Launches Startup to Cut AI Inference Power Use by 1,000x

Naveen Rao, who previously headed AI efforts at Databricks, is tackling one of the most pressing challenges in AI infrastructure: power consumption. His new company, Unconventional AI, is pursuing a radical redesign of computing architecture itself.

Instead of relying on conventional chips, Unconventional is building inference systems around oscillator-based computing. The startup recently demonstrated the concept with Un-0, an image-generation model that claims to match the output quality of industry standards like Stable Diffusion, but using fundamentally different hardware logic.

The Architecture Shift#

The oscillator-based approach represents a departure from the silicon-based transistor logic that has dominated computing for decades. Rather than simulate the technology, Rao's team built a software emulation of their proposed hardware and used it to train a fully functional image model—proving the concept is viable in simulation.

"This is the 'hello world' of a new kind of computer," Rao told TechCrunch. The company plans to release chip schematics soon and eventually operate its own inference infrastructure, offering AI compute capacity at a fraction of today's power requirements.

The Scaling Problem#

Energy constraints are becoming the primary bottleneck for AI expansion. With inference demand accelerating, Rao argues that power availability—not compute innovation—will be the limiting factor for industry growth. Unconventional's bet is that oscillator-based systems can solve this by reducing power draw by up to 1,000 times compared to conventional inference setups.

The ambition is striking for a sub-50-person team, but the scale of the problem may justify the risk. As AI inference workloads continue to explode, efficiency innovations at the architectural level could become as valuable as the models themselves.