source: techcrunch ai: databricks’ former ai chief thinks he can cut ai’s power bill by 1,000x
level: technical
naveen rao, previously head of ai at databricks, now leads unconventional ai, a startup building a new kind of computer architecture based on oscillators. the company released its first model, un-0, an image generation system that runs on a software simulation of its planned chips. the model produces output similar to stable diffusion or openai's gpt image 1, but the underlying hardware design is entirely different from conventional gpus or tpus.
the oscillator-based architecture aims to make inference processing vastly more power efficient. rao claims it could reduce energy consumption by up to 1,000 times compared to current systems. the current un-0 model is a proof of concept, running on a simulation, but the company plans to release chip schematics soon and eventually build a full inference stack. the goal is to offer compute capacity where prompts come in and inferences go out, but at a fraction of the power cost.
rao sees energy as the fundamental limit for ai scaling in the coming years. with demand for inference growing rapidly, the available power supply will constrain progress. unconventional ai, with fewer than 50 employees, is tackling this problem by rethinking computing from the ground up. if successful, the approach could change how ai models are deployed, making large-scale inference more sustainable and accessible.
why it matters: reducing inference energy use by 1000x could make large ai models cheaper to run and more environmentally sustainable, directly addressing a key bottleneck in ai scaling.
source: techcrunch ai: databricks’ former ai chief thinks he can cut ai’s power bill by 1,000x