Unconventional AI

Unconventional AI (legal entity Unconventional, Inc.) is a computing hardware and systems company building a new physical substrate for AI computation. Its stated goal is "biology-scale" energy efficiency, using the human brain's roughly 20-watt power draw as an existence proof that AI computation can be far more efficient than current digital approaches. Rather than simulating physical dynamics on deterministic digital abstractions, the company aims to run neural networks directly on the intrinsic physics of silicon circuits, exposing that physics through a software interface. Press coverage characterizes the effort as neuromorphic or brain-inspired computing and reports the company will produce an analog chip fabricated in silicon, though the specific approach has not been publicly disclosed.

The company's public research output spans analog and mixed-signal circuits, memory hierarchy design (PCM, eDRAM, relaxed error rates), coupled-oscillator and dynamical-system neural architectures, system modeling and digital twins, and an instruction set architecture for dynamical-system hardware (DS-ISA). It frames its central method as "neural co-evolution": co-designing neural network architectures together with the underlying physical systems, targeting roughly 1000x inference efficiency gains, with memory and data movement identified as the primary bottleneck rather than compute. In June 2026 it released Un-0, an image generation model built on a simulated system of coupled oscillators, reporting FID 6.74 on ImageNet 64x64 and publishing weights, training and ablation code openly; follow-up work reported that sparsity removed up to 98.4% of connections while improving FID by 1.9.

Founding story

The company was co-founded by Naveen Rao with Michael Carbin, Sara Achour and MeeLan Lee, a team the company describes as experts in AI systems, analog circuits, computing theory and neuroscience. Rao previously co-founded Nervana Systems, acquired by Intel in 2016 (reported at approximately $350 million, with one report citing more than $400 million), and MosaicML, acquired by Databricks in 2023 for $1.3 billion, after which he led AI at Databricks. The founding thesis is that AI demand will be constrained by global energy supply within a few years, requiring an alternative computational substrate.

Business model

The sources do not describe a commercial business model; the company is at a research and prototyping stage focused on developing a new computing substrate for AI, with the CEO stating that the coming years will be spent testing ideas and prototypes to determine which paradigm scales most efficiently and cost-effectively.

No revenue model is described in the available sources.

Traction

Traction to date is research- and capital-based rather than commercial: a $475 million seed round at a $4.5 billion valuation, an open-weights model (Un-0) reporting FID 6.74 on ImageNet 64x64, a published instruction set architecture for dynamical-system hardware, a $500,000 academic grant program, and a research publication cadence spanning December 2025 to August 2026 with contributions from named researchers including Peter McMahon, Boris Murmann, Giacomo Pedretti, Srenik Mehta, Ang Li, Carlos H. Morales, Sriram Krishnamoorthy, Matthew Storm Bull and Ali Esfahani.

Latest developments

The most recent published work, dated August 13, 2026, reports that sparsity in coupled oscillator networks cuts up to 98.4% of connections in the Un-0 image generation model while improving performance by 1.9 FID on ImageNet 64x64. Earlier in 2026 the company published DS-ISA, an instruction set architecture for dynamical system hardware (June 30, 2026); released Un-0 (June 25, 2026); and closed the Unconventional Grant application cycle, with award notifications scheduled for August 27, 2026.