Naveen Rao proposes an oscillator-based architecture to bypass the AI power limit

The energy requirement for artificial intelligence is approaching a physical ceiling. As data centers consume a growing share of the global power grid, the industry’s trajectory depends less on algorithmic elegance and more on the raw availability of electricity. Naveen Rao, the former head of AI at Databricks, is wagering that the solution lies not in refining the current chip paradigm, but in abandoning it for an oscillator-based architecture.

His new venture, Unconventional AI, has released its first baseline model, Un-0. It is an image-generation system designed to show that this alternative hardware logic can match the performance of state-of-the-art diffusion models like Stable Diffusion. While the current iteration of Un-0 runs on a software simulation of the architecture, it serves as a proof of concept for a stack that Rao claims will eventually operate at one-thousandth the power of traditional silicon.

The logic differs fundamentally from the digital chips that have defined computing for decades. By utilizing oscillators, the system moves away from the binary switching costs that drive heat and power consumption in modern GPUs. Rao’s objective is to build a vertically integrated inference provider, supplying compute capacity where prompts enter a system and results emerge with significantly lower overhead.

With a team of fewer than 50, Unconventional AI is pursuing a structural shift in how the industry handles inference. The company plans to release chip schematics in the near future, moving from simulation toward physical hardware. If the physical implementation replicates the efficiency of the software models, it would transform energy from a hard constraint into a manageable variable, shifting the economics of the entire AI buildout.