← Back to feed

Unconventional AI raised $475M at a $4.5B seed valuation on June 25, led by a16z, Lightspeed, and Sequoia. The company also released Un-0, its first AI model, which runs on a software simulation of an oscillator-based computing architecture.

The company, founded by former Databricks AI chief Naveen Rao, has under 50 employees.

Un-0: What It Is and What It Is Not

Un-0 is an image generation model. On class-conditional ImageNet 64×64, it scores FID 6.74. That is comparable to the performance of leading diffusion models when they were first published — not current state of the art, but competitive with the generation that proved the approach.

The model runs entirely on a software simulation. No Unconventional hardware exists yet. The significance of Un-0 is not its image quality; it is the proof that oscillator-based computation can be trained end-to-end to produce a functional AI output at all.

Rao describes it as “the ‘hello world’ of a new kind of computer.”

Model weights, training code, and ablation code are all open.

The Architecture

Conventional computing processes data through transistors performing binary operations. Unconventional’s approach uses coupled ring oscillators arranged in a fabric network. Information is encoded and processed through the physics of the oscillators themselves — timing, phase, and frequency relationships rather than binary switching logic.

The claim is that implementing AI workloads directly in analog physical dynamics, rather than emulating them in digital logic, dramatically reduces energy consumption per operation. Rao’s target is approximately 10,000 Joules per generated image or token. GPU-based inference at scale consumes significantly more per operation.

The 1,000x efficiency figure assumes that the software simulation’s computational behaviour maps faithfully onto physical hardware. That mapping is the central unverified claim.

Hardware Roadmap

Unconventional plans to release schematics for a physical oscillator chip shortly. The current internal timeline targets a first system-on-chip tape-in in 2026 and mass delivery in 2027.

The intended end state is a compute provider model: Unconventional builds and operates inference infrastructure on its own chips, with standard network interfaces. Clients send prompts in, receive inferences out, at a claimed fraction of conventional power cost.

The gap between a software simulation and a deployed inference stack is wide. Tape-in, fabrication, packaging, software tooling, and customer validation are sequential hard problems. Unconventional has not disclosed how many of those steps are funded by this raise or what milestones trigger future rounds.

Founder Track Record

Rao’s prior exits set the context for why investors are paying $4.5B at seed stage.

He founded Nervana Systems, an AI chip startup acquired by Intel in 2016 for approximately $400M. The acquisition was strategically significant at the time — Intel was attempting to build a position in AI silicon before Nvidia established its current dominance.

He co-founded MosaicML, an AI training efficiency company, which Databricks acquired in 2023 for $1.3B. MosaicML developed training optimisation techniques that reduced the compute required to train large models.

Two exits, both in AI infrastructure, both above $400M. That background is doing meaningful work in a $4.5B seed valuation for a company that has released one open-source model running in simulation.

The Energy Constraint Argument

Rao’s framing positions energy as the next hard constraint on AI scaling, harder than parameter count or algorithmic efficiency. Data center power demand from AI workloads has driven nuclear restarts, gigawatt-scale utility agreements, and constrained grid capacity in multiple regions.

If conventional digital silicon is the bottleneck — if transistor switching logic is fundamentally energy-inefficient relative to what physical dynamics can do — then the argument for a new substrate is structural rather than incremental.

That argument is consistent with a long body of research on neuromorphic computing, analog circuits, and physical computing going back decades. What most of that research has not produced is a model that performs competitively on real AI benchmarks. Un-0 is the closest current demonstration that oscillator-based systems can replicate conventional AI outputs. The FID 6.74 score on ImageNet 64×64 is the number that matters here — not because it is state of the art, but because it is real.

Whether 1,000x efficiency survives contact with physical hardware, fabrication tolerances, and production inference workloads is a question only a built chip can answer.