AMSTERDAM, March 17 — Nebius (NASDAQ: NBIS), the AI cloud company, today announced it is collaborating with NVIDIA to accelerate physical AI development with an end-to-end platform purpose-built for the full robotics lifecycle, from simulation and training to real-world deployment at scale.
Combining Nebius’s global AI cloud infrastructure with the NVIDIA Physical AI Data Factory Blueprint, an open reference architecture for massive data generation and evaluation, Nebius will provide robotics developers and enterprises an agent-driven environment that addresses the two fundamental barriers to physical AI at scale: infrastructure and tooling fragmentation, and the lack of high-quality training data for rare, unpredictable scenarios that determine real-world success.
“Physical AI is going to be one of the defining technology shifts of this decade, and the teams building it today are being held back by infrastructure and tooling that was never designed for those workloads,” said Evan Helda, Head of Physical AI at Nebius. “Working with NVIDIA, we are building the execution layer for the entire physical AI ecosystem — so that any team, anywhere, can go from idea to deployed robot at the speed the market demands.”
“Physical AI is the next phase of computing — where intelligence is trained, tested and validated in simulation before it operates in the real world,” said Rev Lebaredian, VP of Omniverse and simulation technologies at NVIDIA. “That demands tightly integrated systems connecting large-scale AI training with physically accurate simulation to create a continuous data flywheel. By integrating the NVIDIA Physical AI Data Factory Blueprint, Nebius is enabling developers to generate physics-grounded synthetic data and build safe, robust autonomous machines at scale.”
Solving physical AI’s three-computer problem
Building physical AI at scale means operating across three distinct environments — large-scale GPU training, simulation testing, and edge deployment — each with its own infrastructure and tooling. Engineering teams routinely spend 30–40% of their time on integration work rather than improving robot behaviour.
Real-world training data compounds the challenge: it is expensive and dangerous to collect, inconsistent across companies, and never sufficient to cover the long-tail edge cases that determine whether a robot succeeds or fails in the field.
The Nebius cloud solution for physical AI addresses both challenges. NVIDIA OSMO — delivered as an easy-to-consume managed service — provides unified, agentic orchestration across the entire pipeline. NVIDIA Cosmos open world foundation models generate large-scale, physics-consistent synthetic data that bridges the gap that real-world collection cannot close.
The whole stack runs on Nebius AI Cloud — purpose-built infrastructure combining NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, high-throughput object storage, integrated data management and labelling, serverless features and managed inference directly within the platform — so teams can consume it as a service, without having to provision clusters or manage integrations.
Beyond large-scale simulation and training, Nebius extends the robotics lifecycle into production with serverless and managed inference services, including Nebius Token Factory, enabling teams to deploy and scale trained policies with low-latency execution from cloud to edge.
The result is a complete managed physical AI runtime, from synthetic data generation to real-world inference, delivered through a tightly integrated platform that can be consumed as a service.
Leading physical AI companies building the future with Nebius and NVIDIA
RoboForce builds AI robots for unstructured outdoor environments — solar farms, construction sites, agricultural fields — where encountering rare edge cases is a daily reality. Using NVIDIA Cosmos open world foundation models on the Nebius cloud, RoboForce cut pipeline setup time by more than 70% and significantly accelerated the rate at which new policies reach production.
“Manual handoffs between data generation, simulation, and training means our GPUs can sit idle — costing us both time and money,” said Calvin Zhou, co-founder of RoboForce. “Using OSMO agentic orchestration, our engineers can push a single configuration file and run the entire pipeline end-to-end. We’re generating thousands of scenario variations with NVIDIA Cosmos on Nebius AI Cloud, powering our AI data flywheel and accelerating the development of our robot foundation model. This allows us to push hardened robot models straight to the edge and cut our iteration cycles from weeks to days.”
Voxel51, a physical AI data platform and key technology partner of Nebius, provides powerful data visualization, curation, annotation, and analysis capabilities for teams to build high-quality datasets for model training and simulations. By running FiftyOne workflows on Nebius GPU clusters, Voxel51 customers can curate, augment, and quality-check visual datasets at scale—reducing the time between data collection and model deployment.
“Data is the biggest determinant of computer vision success. As vision AI systems become more capable, the limiting factor is no longer algorithmic innovation, but the quality, coverage, and observability of the data used to train models,” said Brian Moore, CEO and co-founder of Voxel51. “Nebius gives our users the compute infrastructure for running complex data tasks such as auto labeling and generating novel scenes at the speed and scale needed by physical AI systems.”
Together with Nebius cloud for physical AI and NVIDIA technologies, Voxel51 is delivering a synthetic data generation pipeline for its customer, Porsche Engineering, to accelerate autonomous driving data augmentation workflows.
Milestone Systems, a global leader in intelligent video management software and the company behind the Hafnia platform for computer vision, selected Nebius to fine-tune its next-generation Vision-Language Models (VLMs). Milestone curates real-world video footage into compliant, annotated training data, then uses it to fine-tune NVIDIA Cosmos Reason into highly accurate, use-case specific VLMs. For this computationally intensive work Nebius provides sustained access to large GPU clusters, high-throughput data pipelines, and managed workflow orchestration that keeps training runs stable and cost-efficient.
“We evaluated several cloud providers, and Nebius offered the best combination of GPU availability, price-performance, and hands-on engineering support for our physical AI and VLM training workloads,” said Edward Mauser, Director of Hafnia at Milestone Systems. “We chose Nebius not just for their tech, but also for their commitment to data sovereignty — guaranteeing that European customers’ data can remain within Europe.”
