Databricks' former AI chief launches startup promising to cut AI inference power consumption by 1,000x using a completely new chip architecture.

Key takeaway

• Unconventional AI, led by former Databricks executive Naveen Rao, released its first AI image-generation model built on a new oscillator-based chip architecture designed to reduce power consumption by as much as 1,000 times compared to conventional systems.

• The startup believes energy constraints will become the fundamental limit to AI scaling and is developing actual chips and a complete inference infrastructure to address this bottleneck.

3 Key Points

FAQ

How does Unconventional's approach differ from conventional AI systems?

The company uses an oscillator-based computer architecture that is completely different from the chips powering conventional computing and traditional LLMs. Un-0 was built on a software simulation of this new architecture and achieves performance comparable to state-of-the-art diffusion models like Stable Diffusion.

When will the actual hardware be available?

The current Un-0 model runs on a software simulation; the company plans to release schematics for actual oscillator chips soon and will build an entire inference stack from the ground up to eventually supply compute capacity over the network.