[DigitalToday reporter Jinju Hong] Unconventional AI, founded by Nabeen Rao (나빈 라오), who led the artificial intelligence (AI) division at Databricks, unveiled a new computing structure that it said can dramatically cut power use required for AI inference. The company set a long-term goal of reducing power consumption to as low as one-thousandth of existing AI systems.

On June 25 local time, IT outlet TechCrunch reported that Unconventional AI released its first AI model, the image generation system "Un-0," along with a related research paper. The company claimed it can deliver performance similar to existing diffusion models by using a new oscillator-based computing architecture.

The core of the announcement is the computation method that runs the system, rather than the image generation model itself. Un-0 aims for image generation results at a level similar to Stable Diffusion or OpenAI's "GPT Image 1," but its biggest difference is that it applies an oscillator-based structure instead of typical GPUs and large-scale neural network computation.

Unconventional AI expects that this structure could, over the long term, reduce power consumption in AI inference to as low as one-thousandth of existing levels.

The current stage is software simulation, not an actual semiconductor. The research team implemented the new architecture in a software environment to complete the image generation model, and explained that its performance is similar to that of the latest diffusion models. The company plans to publish actual chip designs later and then push ahead with development of dedicated AI semiconductors.

Unconventional AI Chief Executive Officer Nabeen Rao (나빈 라오) described the release as "Hello World for a new kind of computer." He said he plans to disclose additional technical development results within the next year, adding that while this is the stage of proving feasibility, it will ultimately lead to implementation in real hardware.

Its business strategy is not limited to selling chips. Unconventional AI plans to build a new AI system with its own chip design, then expand into an infrastructure business that provides AI inference services based on that system. Users would enter prompts and receive results the same way as existing AI services, but the goal is to handle the same tasks with far less power internally.

The company pointed to energy as the biggest bottleneck in the AI industry. Unconventional AI currently has fewer than 50 employees, but Rao stressed that securing power needed for AI adoption is a bigger challenge than company size. "Over the next few years, the most fundamental factor limiting AI progress is energy," he said, adding that innovation in computing efficiency will be key to next-generation AI competitiveness.

The industry is watching whether this technology can be implemented in actual semiconductors. If Unconventional AI can reproduce in real chips the performance and power efficiency it confirmed in simulation, some assess it could become a new computing architecture that significantly reduces the power consumption of AI data centres. Others expect that mass hardware production and ecosystem building will take considerable time, making the release of chip designs and tangible product development results a watershed for the technology's competitiveness.