Unconventional AI Targets 1,000x Power Cut for AI
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Unconventional AI released Un-0, its first model — an image-generation system that performs as well as state-of-the-art diffusion models like Stable Diffusion, built on a software simulation of an oscillator-based architecture rather than conventional chips.
- Naveen Rao, formerly head of AI at Databricks, believes the oscillator-based computing will ultimately reduce AI inference power consumption by as much as 1,000 times compared to current hardware.
- Unconventional AI plans to release schematics for an actual physical chip soon, then build a complete inference stack and supply compute capacity 'just like any other provider,' with prompts in and inferences out at 1/1000 of current power.
- The company has fewer than 50 employees, making its ambition to rebuild computing architecture from the ground up unusually aggressive for its headcount — Rao calls this the 'hello world' of a new kind of computer.
- Rao frames energy as the hard ceiling on AI's future growth: 'AI scaling is hard because of energy... it's going to be an energy-limited problem, at the end of the day.'
Why it matters: Rao himself identifies energy as the fundamental limit on AI scaling, which makes even a partial efficiency breakthrough commercially significant for inference providers facing grid and power constraints. But the gap between the 1,000x promise and current reality is enormous — fewer than 50 employees, no physical chip yet, and only a software simulation demonstrating parity with existing diffusion models.
Ask SkimNews
