Oscillator-based AI computer by Unconventional AI: status and timeline

Oscillator-based AI computer

Run AI inference on about 1,000 times less energy than GPUs, easing the power limits on data-centre growth.

ResearchStage 1 of 5

Un-0 runs on a GPU simulation; chip schematics promised soon, no chip built or measured yet.

Updated 25 Jun 2026Checked 11 Oct0 updates this week

Milestones

Next · Oscillator chip schematics released

Company foundedOct 2025Complete.
$475M seed at $4.5B valuation9 Dec 2025Complete.
Un-0 image model released (simulated oscillators)25 Jun 2026Complete.
Oscillator chip schematics releasedNowCurrent milestone.

Most important updates

Current obstacles

No hardware yetUn-0 is a GPU simulation. Building real oscillator chips that match the simulated behaviour is still ahead.
Image quality trails the state of the artUn-0's best score matches earlier generators such as BigGAN and iDDPM but falls behind recent methods like EDM, which had years of refinement.
Getting data in and outConverting digital inputs to analog states and back costs energy that can eat into the savings.

Physics limits

Analog noise caps precisionThermal noise and component mismatch blur every analog value, so results carry only a few bits of precision. Models must tolerate that, or errors grow layer by layer.
Manufacturing variationNo two fabricated oscillators are identical. Each chip may need calibration or training-time tolerance, which adds cost and complexity.
Amdahl's law on energySaving 99% of energy in one part still leaves the rest. Reaching 1,000x needs every stage, including memory and conversions, to shrink together.

How it works

4 parts

Oscillators

Phases instead of bits

Each oscillator has a phase and a natural frequency. Learned couplings pull phases together, so the network drifts toward a pattern that encodes the output.

Generation

Un-0: let the system evolve

Start from random phases, nudge a small group with the requested class, let the dynamics run, then a small digital decoder turns final phases into pixels.

Learning

Training sets the couplings

The main learned parameters are the coupling strengths between oscillators; the digital decoder that turns phases into pixels holds under 15% of the parameters.

Efficiency

Compute where data lives

In hardware, the couplings would act as weights stored inside the circuit, avoiding the trips to off-chip memory that use most of a GPU's inference energy.

Spec sheet

Spec	Oscillator-based AI computer
Energy reduction target	~1,000x vs today's AI chips^R (reported)
Largest Un-0 oscillator count	16,384^R (reported)
Largest Un-0 parameters	322M^R (reported)
Un-0 FID, ImageNet 64x64	6.74^R (reported)
Largest model training	640 B200 GPU-hours^R (reported)

^R reported by the company

Papers & demos

1 item

25 Jun 2026code
Un-0: generating images with coupled oscillators (opens unconv.ai)First model whose core is an oscillator network; weights and code released.

Update log

2 updates

Thu 25 Jun

Major: Unconventional AI releases Un-0, an image generator built on simulated coupled oscillatorsBlogResearch

Tue 9 Dec 2025

Major: Unconventional AI confirms $475M seed round at $4.5B valuation, led by a16z and LightspeedPressFunding

About Unconventional AI

The team behind Oscillator-based AI computer

Unconventional AI

Oscillator-based analog AI computers

Unconventional AI, co-founded in October 2025 by Naveen Rao, is designing a new kind of AI computer in which coupled analog oscillators do the maths through their own physics. Its goal is AI inference using about 1,000 times less energy than today's chips.

Raised a $475M seed at a $4.5B valuation in Dec 2025, led by a16z and Lightspeed: the first tranche of up to $1B.
Un-0 (Jun 2026) generates images with up to 16,384 oscillators, simulated on GPUs; no chip exists yet. Chip schematics are promised 'soon'.
Founder Naveen Rao previously sold Nervana to Intel and MosaicML to Databricks.

unconv.ai (opens unconv.ai)Company profile →

Founded: 20251 yrs
Headquarters: United States
Status: Private
Valuation: $4.5BprivateDec 2025
Raised: $475M1 round
Last round: Seed · $475MDec 2025
Works in: ComputingNew kinds of computers
Coverage: 1 program · 2 updateslatest 25 Jun 2026checked 11 Oct
Lead investors: Andreessen HorowitzLightspeed Venture Partners
Investors: Andreessen Horowitz (opens techcrunch.com)Lightspeed Venture Partners (opens techcrunch.com)Lux Capital
People: Naveen RaoCo-founder & CEOMeeLan LeeCo-founder & VP EngineeringMichael CarbinCo-founderSara AchourCo-founder

Spec

Oscillator-based AI computer

Energy reduction target

~1,000x vs today's AI chips^R (reported)

Largest Un-0 oscillator count

16,384^R (reported)

Largest Un-0 parameters

322M^R (reported)

Un-0 FID, ImageNet 64x64

6.74^R (reported)

Largest model training

640 B200 GPU-hours^R (reported)