Nebius Launches 2026 Physical AI Awards With $750K Compute Prizes

Nebius, in partnership with NVIDIA, has launched the 2026 Nebius Physical AI Awards, a global competition designed to support startups that have moved beyond the research phase and have a demonstrable physical AI product in the field. The awards specifically target companies facing the challenges of scaling their operations, including managing large volumes of fleet data, validating retraining cycles for hardware deployment, and optimizing the sim-to-real loop for production environments. Applications for the competition are now open and will close on October 25, 2026, with finalists and winners to be announced in mid-November 2026. The program features five distinct categories, and each winner will receive a substantial prize of $150,000 in Nebius compute credits. This prize is intended to directly fund future training runs for the winning companies' AI models. In addition to the compute credits, winners will benefit from joint promotional support across Nebius and NVIDIA's communication channels, executive mentorship sessions with Nebius C-level leaders focused on scaling strategies, and a "Winners Circle" feature that will showcase their team and product through content published by Nebius. Furthermore, each winner will receive two invitations to an exclusive executive dinner next year, providing networking opportunities with other leaders in the Physical AI ecosystem. There are no entry fees associated with the competition, and no categories are sponsored, ensuring a level playing field for all participants. The value of the $150,000 compute credit prize is made concrete by Nebius's published GPU-hour rates. For instance, at on-demand rates, this prize can translate into significant compute resources. For an 8-GPU NVIDIA H100 node, $150,000 provides approximately 38,960 GPU-hours, equivalent to about 203 node-days. With the newer H200 GPU, the prize offers around 33,330 GPU-hours, or 174 node-days. For the upcoming B200 GPU, it amounts to roughly 20,980 GPU-hours (109 node-days), and for the B300 GPU, approximately 19,100 GPU-hours (99 node-days). To illustrate the scale of compute, $150,000 in H200 credits on a 64-GPU cluster equates to about 521 hours, or roughly three weeks of continuous operation. Utilizing preemptible H200 rates, which are priced at $2.45 per GPU-hour, the compute time extends to approximately 61,200 GPU-hours, a duration suitable for checkpointable reinforcement learning tasks and data generation workloads. Nebius bundles vCPU and RAM resources into these GPU rates, simplifying the calculation of compute costs for participants.