Databricks’ former AI chief thinks he can cut AI’s power bill by 1,000x

In the quest for groundbreaking advancements in artificial intelligence, one startup is daring to rethink the very foundations of computing architecture. Unconventional AI, spearheaded by Naveen Rao, the former AI chief at Databricks, aims to drastically enhance power efficiency in AI inference processing. Their innovative approach centers around a novel oscillator-based computer architecture. On Thursday, the company unveiled its inaugural AI model, dubbed Un0, an image-generation tool that exemplifies how their technology can emulate traditional AI systems. Accompanied by a detailed research paper, the Unconventional AI team revealed that they successfully constructed a fully operational image generation model through software simulations of their unique architecture. This model matches the performance of cutting-edge diffusion models currently in use. "This is the 'hello world' of a new kind of computer," Rao shared with TechCrunch. He anticipates that the coming year will bring significant developments in this field. The outputs produced by the Un0 model closely resemble those from popular image generation systems like Stable Diffusion and OpenAI’s GPT Image 1. What sets this model apart is its underlying oscillator-based architecture, which diverges fundamentally from the conventional chips that power existing computing and traditional large language models (LLMs). Rao is confident that this innovative architecture could potentially cut energy consumption by an astonishing factor of 1,000. Although much of the infrastructure necessary for this transformation is still under construction, the present version of Un0 operates on a software simulation of Unconventional’s oscillator chips. The company is poised to release actual chip schematics soon, with plans to establish a comprehensive inference stack from scratch. "We envision creating a system built around our chips," Rao explained. "AI models will run on this system, with prompts input through a network cable, and inferences output, all while consuming just a fraction of the power—1/1000, to be exact." This ambitious goal is particularly striking for a company with fewer than 50 employees. However, given the immense demand for AI and the associated costs of inference, Unconventional's initiative could be one of the few capable of addressing these challenges effectively. Rao emphasizes that the availability of power will soon become a critical constraint for AI development. "AI scaling is challenging due to energy limitations. This will be a fundamental barrier in the coming years," he stated, underlining the urgency of their mission.