Unconventional AI has launched its initial model, which is designed on an oscillator architecture that has the potential to reduce AI power consumption by a factor of 1000.

      TL;DR: Unconventional AI has introduced Un-0, an image generation model based on a simulated oscillator architecture, which founder Naveen Rao claims could reduce AI power consumption by 1000 times.

      Unconventional AI, a startup established by former Databricks AI chief Naveen Rao, has launched its first AI model, an image generation system named Un-0, that operates on a novel computing architecture. According to an accompanying research paper, the model produces results similar to leading diffusion models like Stable Diffusion. However, it operates on a software simulation of hardware that is not yet available.

      The company is developing an oscillator-based computer architecture that moves away from the digital logic that underpins almost all modern computing. Rather than processing information through transistors executing binary operations, Unconventional employs coupled ring oscillators within a network, encoding and processing data through the oscillators' physical properties. Rao shared with TechCrunch that this method could potentially cut power consumption by a factor of a thousand compared to traditional chips.

      This claim is ambitious. US utility companies are projected to spend about $1.5 trillion by 2030 on infrastructure primarily driven by AI data center needs, so any technology that could significantly alleviate that strain would be immensely valuable. However, Unconventional has yet to create a physical chip, and the thousand-fold improvement remains solely a theoretical estimate.

      What Un-0 does showcase is that the architecture can mimic the functions of standard AI systems. The research team successfully developed a fully functional image generation model using a software simulation of the oscillator architecture, with the findings indicating performance on par with existing diffusion models. “This is the ‘hello world’ of a new kind of computer,” Rao stated to TechCrunch.

      Rao's history lends credibility to investors' willingness to support the venture. He co-founded Nervana Systems, a deep learning chip startup that Intel purchased for around $400 million in 2016. He later founded MosaicML, which Databricks acquired for about $1.3 billion in 2023.

      With a PhD in neuroscience from Brown and electrical engineering studies at Stanford, Rao's unique background, combining chip design and neuroscience, is key to his argument that computing architecture needs a transformation.

      His experience helped attract $475 million in seed funding at a $4.5 billion valuation in December 2025, led by Lightspeed and Andreessen Horowitz, with contributions from Sequoia, Lux Capital, DCVC, and Jeff Bezos. Rao also invested $10 million of his personal funds under the same terms. While Unconventional is not the only startup pursuing AI efficiency through new architectures, its method is one of the most ambitious.

      The company aims to release schematics for a physical chip soon and plans to construct a comprehensive inference stack from the ground up. The ultimate goal is to serve as a compute provider, with Unconventional delivering inference capacity via its chips. “We will create a new kind of system made of our chips,” Rao explained, noting that inputs would come in and outputs would go out via a standard network connection, but with significantly reduced power requirements.

      The ambition of this project is substantial considering the company's small size. Unconventional employs fewer than 50 people and is seeking to replace the von Neumann stored-program computer architecture, which has been dominant for about 80 years. The push to decrease AI’s energy consumption has prompted a surge of startups, though most focus on cooling, efficiency software, or gradual hardware enhancements rather than a complete overhaul of the computing framework.

      Rao argues that incremental solutions will not suffice. “AI scaling is challenging due to energy,” he remarked to TechCrunch, pointing out that power will become a crucial limit in the coming years. The International Energy Agency forecasts that global data center electricity usage will surpass 1,000 terawatt-hours by the end of 2026.

      The disparity between Un-0’s software simulation and an operational chip capable of real-world inference at scale is significant, and the company has not provided a timeline for when the physical hardware will be ready for commercial application. Nevertheless, the demonstration that oscillator-based computing can yield functional AI output offers the first tangible proof that this approach extends beyond theory. Whether it can fulfill the promise of a thousand-fold efficiency remains a question that only hardware can resolve.