Nebius Opens 2026 Physical AI Awards: Five $150K Compute Credit Prizes

Nebius has opened applications for its 2026 Physical AI Awards, a global contest for startups already shipping or testing a real product. The program runs with NVIDIA, closes on October 25, 2026, and will name finalists and winners in mid-November. There are five categories, and each one comes with $150,000 in Nebius compute credits. No entry fee, no sponsored categories, no fog machine.

That compute is the point. Nebius says the credits go straight into the next training run, and it publishes GPU-hour pricing so the award can be translated into actual capacity. At on-demand rates, $150,000 buys roughly 38,960 GPU-hours on H100s, 33,330 on H200s, 20,980 on B200s, or 19,100 on B300s. For a team already running multi-node work, Nebius says that works out to about 521 hours on a 64-GPU H200 cluster. At preemptible H200 pricing, the same prize stretches much further.

The categories cover the full stack of physical AI work: models, perception, simulation and synthetic data, systems and deployment, and software plus tooling and orchestration. Nebius wants applicants to choose the single best fit, even if they have more than one product. To qualify, a company needs a clear physical AI use case, an MVP in active use or testing, a registered legal entity, and a live public website.

Judging starts in-house for quality and completeness, then moves to a panel picked by Nebius and NVIDIA that scores innovation, implementation, market potential, and real-world impact. The listed judges skew toward operators who know the mess behind the demos: people from Nebius, NVIDIA, Foxglove, Voxel51, Encord, Antioch, and RoboForce. Nebius says more judges may be added before the deadline. The winners also get joint promotion, executive mentorship, a Winners Circle feature, and two seats at an exclusive dinner next year.

There’s a reason this kind of award matters now. Physical AI is past the novelty stage, where a pretty prototype can hide weak infrastructure. Once a robot, sensor system, or fleet product is out in the wild, the real constraint is compute, validation, and the ability to keep retraining without breaking what already works.