Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchNVIDIA releases the Physical AI Data Factory — an open architecture for generating, curating, and scaling synthetic training data for robotics and embodied AI, with Microsoft Azure and Nebius as launch partners.
AI, startups & growth insights. No spam.
TL;DR: NVIDIA has released the Physical AI Data Factory — an open reference architecture designed to let robotics companies generate, curate, and scale the synthetic training data that embodied AI systems need. Microsoft Azure and Nebius are the first cloud launch partners, giving the ecosystem on-demand access to NVIDIA's Omniverse-powered simulation pipeline without running on-premises GPU clusters. The announcement lands as the robotics industry is hitting a wall that hardware alone cannot solve: a structural shortage of diverse, labeled physical-world training data.
The Physical AI Data Factory is not a product in the traditional sense. It is an open, composable reference architecture — a set of specifications, APIs, and cloud-deployable pipeline components that any organization can use to build a production-grade pipeline for generating and curating robot training data at scale.
Think of it the way you think of a cloud data warehouse reference architecture. The blueprint tells you which pieces to assemble, how they connect, and what standards to follow. You bring your robot, your task domain, and your simulation assets. NVIDIA supplies the rendering engine (Omniverse), the physics layer (Newton), the model scaffolding (GR00T), and now the factory-floor layout for turning all of it into usable datasets.
The architecture covers four layers:
Scene generation. Procedurally generated 3-D environments built inside Omniverse, covering warehouse floors, kitchen countertops, factory assembly stations, hospital corridors, and other settings where robots will actually operate. Lighting conditions, object placements, surface textures, and clutter levels vary automatically across thousands of scene instances per run.
Motion synthesis. Robot trajectories and manipulation sequences are generated by combining motion-capture references, inverse-kinematics solvers, and reinforcement-learning policies trained in simulation. This produces physically plausible motion data at a volume no human teleoperator team could match.
Sensor simulation. Simulated RGB cameras, depth sensors, LiDAR, and tactile inputs generate the kind of multimodal observation streams that real robots use. Domain randomization — varying sensor noise, lens distortion, and calibration offsets — helps close the sim-to-real gap.
Curation and filtering. Automated quality filters remove physically implausible samples, detect task-failure cases, and enforce dataset balance across object classes and environment conditions. The output is a structured, versioned dataset ready for supervised fine-tuning or reinforcement learning.
The word "open" carries real weight here. NVIDIA is publishing the reference design publicly, not locking it behind a proprietary cloud service. Companies can deploy this pipeline on their own infrastructure, on Microsoft Azure, or on Nebius — and mix components from the stack with their own tooling.
Synthetic data is not new in machine learning. Computer vision teams have used rendered images to augment training sets for years. What is new is the scale, fidelity, and multimodality required for physical AI.
A robot learning to pick up a coffee mug needs more than photographs of coffee mugs. It needs hundreds of thousands of examples showing the mug in different orientations, under different lighting conditions, at different positions relative to the robot's gripper, on surfaces with different friction coefficients — and it needs to know what happens in the sensor stream at each millisecond of the grasp attempt.
Collecting that data in the real world is economically impossible at the scale modern robot foundation models require. A single task policy might need millions of demonstration episodes to generalize reliably. Human teleoperation, even at $10 per episode (optimistic), would cost tens of millions of dollars for one task. Industrial robots doing tens of thousands of repetitive cycles can generate real-world data, but only for the narrow task they were already programmed to do — not for the novel task you are trying to teach.
Simulation solves the economics. A GPU cluster running NVIDIA Omniverse can render physically accurate robot-interaction scenes orders of magnitude faster than real time. A scene that would take a human teleoperator thirty seconds to execute can be simulated, randomized, and logged in milliseconds. The same cluster can run thousands of scenes in parallel.
The challenge is fidelity. Simulations that look convincing to a human eye can still produce training data that fails to transfer to real robots, because subtle physical inaccuracies in contact dynamics, deformable object behavior, or lighting create distributional gaps. NVIDIA's Newton physics engine, announced alongside GR00T N2 at GTC, is specifically designed to close that gap by running GPU-accelerated rigid-body and soft-body simulation at a level of accuracy that previous real-time physics engines could not reach.
The Physical AI Data Factory packages all of this into a repeatable pipeline so every robotics company does not have to build it from scratch.
Two cloud providers are launching as the first infrastructure partners for the Physical AI Data Factory: Microsoft Azure and Nebius.
Microsoft Azure's participation is the higher-profile of the two. Azure already hosts NVIDIA GPU capacity through its ND-series and NC-series virtual machine families, and the Azure partnership positions the Physical AI Data Factory as a managed capability within the Azure cloud. For enterprises already running workloads on Azure — a category that includes most large manufacturers, logistics operators, and healthcare systems considering robotics deployments — this means they can spin up a synthetic data generation pipeline without leaving their existing cloud environment, procurement contracts, or compliance boundaries.
The Azure integration also connects to Microsoft's broader industrial AI push. Azure Digital Twins, Azure IoT, and Azure Kubernetes Service can all serve as orchestration layers around the Physical AI Data Factory pipeline. A manufacturer building a digital twin of a production line in Azure Digital Twins can feed that environment directly into the Omniverse simulation layer to generate task-specific training data for the robots that will work on that line.
Nebius is a less familiar name outside the European AI infrastructure market. Formerly part of Yandex, Nebius relaunched as an independent AI cloud provider in 2024 and has been building a GPU-dense infrastructure footprint aimed at AI training workloads. For NVIDIA, adding Nebius as a launch partner serves two purposes: it demonstrates that the Physical AI Data Factory is portable across clouds rather than an Azure-exclusive offering, and it gives European robotics companies a regionally compliant option for running synthetic data pipelines.
Both partnerships underscore that NVIDIA is positioning the Physical AI Data Factory as ecosystem infrastructure — not a service NVIDIA itself operates — which is consistent with the open architecture framing.
The large language model revolution of 2022-2025 was, at its core, a data story. Models like GPT-4 and Claude benefited from the extraordinary fact that the internet already contained trillions of words of human-generated text in a form that could be scraped, cleaned, and fed into a transformer. The data problem was not trivial, but it was tractable.
Physical AI faces a fundamentally different situation. There is no internet of robot manipulation data. There is no trillion-episode dataset of robots picking objects off shelves, assembling components, or navigating hospital corridors. Every demonstration has to be either collected in the real world — slowly, expensively, with physical robots and human supervisors — or generated in simulation.
This bottleneck is not theoretical. It is what separates today's narrow industrial robots, which are excellent at the single task they were programmed for, from the general-purpose humanoids and mobile manipulators that the industry is trying to build. GR00T N2, NVIDIA's latest humanoid foundation model, was trained on a combination of real and synthetic data, and the researchers behind it have been explicit that scaling synthetic data generation was the key lever that produced the 2x improvement in task completion rates over prior models.
The same pattern is playing out across the industry. Tesla's Optimus team has described synthetic data generation as a core pillar of their training pipeline. Google DeepMind's RT-X and Open X-Embodiment projects have worked to aggregate real-world demonstration data across institutions precisely because any single organization cannot collect enough on its own. Boston Dynamics uses simulation extensively for locomotion policy training.
The Physical AI Data Factory does not solve the sim-to-real transfer problem by itself — that remains a research challenge — but it dramatically lowers the cost of generating large, diverse, high-quality synthetic datasets. If the transfer problem is the ceiling, the data factory raises the floor.
NVIDIA Omniverse is the backbone of the Physical AI Data Factory, and understanding how it fits helps clarify why NVIDIA's approach is structurally different from building a simulation pipeline on a game engine or a general-purpose 3-D renderer.
Omniverse is built around Universal Scene Description (USD), Pixar's open format for representing 3-D scenes in a way that is physically accurate, modular, and interoperable across tools. USD allows a manufacturer's CAD model of a factory floor to be imported into Omniverse and used directly as a simulation environment, without manual conversion or fidelity loss. The same digital twin that an engineering team uses to plan the layout of a production line can become the environment in which robots are trained to work on that line.
This is not a hypothetical. Companies like BMW and Foxconn have been working with NVIDIA Omniverse to build factory digital twins, and those same environments are candidates for the kind of synthetic data generation the Physical AI Data Factory formalizes. When a manufacturer wants to train a robot for a new assembly task, they can procedurally generate thousands of variants of their existing Omniverse factory scene — different lighting shifts, different part positions, different co-worker locations — and produce a training dataset that is specific to their actual facility rather than a generic simulated environment.
The specificity matters. A robot trained on generic warehouse data will perform worse in a specific facility than one trained on data generated from a digital twin of that facility. This is the same argument that makes domain-specific fine-tuning of language models valuable: general capability plus domain adaptation outperforms general capability alone.
Omniverse also provides the rendering fidelity that domain randomization requires. To close the sim-to-real gap, simulation pipelines randomize visual conditions — lighting angles, shadow intensity, surface reflectance — so the trained policy learns to generalize across visual variation. Omniverse's physically based rendering engine produces variation that is realistic enough to be useful, rather than the kind of obviously artificial variation that a game engine's renderer might produce.
For more on how NVIDIA's simulation stack is evolving, see our coverage of NVIDIA GR00T N2 physical AI robots.
The Physical AI Data Factory did not appear in isolation. It is one piece of a larger robotics infrastructure strategy that NVIDIA has been assembling over the past two years and that came into clearest focus at GTC 2026.
The stack, as NVIDIA has laid it out, runs from silicon to software to data to model. At the hardware layer: Jetson Thor, the edge computing platform for deploying robots. At the simulation layer: Omniverse and the Newton physics engine. At the model layer: GR00T N2, the humanoid foundation model, and Cosmos, the world simulator that can generate video-based training data from text or image prompts. At the infrastructure layer: Isaac Lab for training orchestration and now the Physical AI Data Factory for dataset generation and management.
Each layer feeds the others. Cosmos generates photorealistic video of robot interactions that can be used as training data. The Physical AI Data Factory structures how that data is generated, stored, and versioned. Isaac Lab uses that data to train GR00T N2 policies. The trained policies deploy on Jetson Thor hardware. The real-world performance of those deployed robots generates new data that feeds back into the pipeline.
This is a vertically integrated flywheel — not in the sense of a walled garden, but in the sense of a self-reinforcing system where each component makes the others more valuable. The open architecture framing means that companies can participate at any layer without buying into the entire stack, but the more of the stack they use, the more tightly the integration benefits compound.
The Physical AI Data Factory is, in this framing, the part of the flywheel that was most visibly missing before GTC. NVIDIA had the simulation engine, the physics layer, the model layer, and the deployment hardware. The data factory formalizes the pipeline that connects simulation to training dataset — the step that actually produces the fuel for the entire system.
See also: NVIDIA Vera Rubin DSX AI Factory reference design for how NVIDIA's AI factory concept applies to the broader compute infrastructure story.
NVIDIA is not the only company working on synthetic data infrastructure for robotics, and the competitive landscape illuminates what is at stake.
Tesla has the most vertically integrated approach in the industry. Optimus is trained almost entirely on synthetic data generated from Tesla's internal simulation pipeline, which is built on the same infrastructure Tesla uses for Autopilot. Tesla has the advantage of fleet data from millions of cars that have already learned to perceive the physical world, and it has the compute scale of its Dojo training cluster. The disadvantage is that Tesla's pipeline is entirely proprietary and serves only Optimus — there is no ecosystem play.
Boston Dynamics takes a hybrid approach. Atlas, their humanoid, is trained using a combination of real-world motion capture, human teleoperation, and simulation. Boston Dynamics has deep expertise in locomotion dynamics built over decades of research, which gives their simulation environments high fidelity for movement tasks. What they have historically lacked is the general manipulation capability that requires the kind of large-scale synthetic data generation NVIDIA is now systematizing.
Google DeepMind is pursuing a collaborative data strategy through projects like Open X-Embodiment, which aggregates real-world robot demonstration data across dozens of research institutions. This approach produces real data with no sim-to-real gap, but it is limited by the volume and diversity constraints of human-collected data. DeepMind has also invested heavily in simulation for locomotion tasks, but their large-scale manipulation training has relied more on real data than synthetic generation.
Figure AI, Physical Intelligence (Pi), and Apptronik represent the emerging generation of robotics companies that will be potential customers of NVIDIA's Physical AI Data Factory rather than direct competitors on the data infrastructure layer. These companies need training data at scale and do not have the resources to build their own simulation pipelines from scratch. For them, the open reference architecture and cloud deployment options reduce a major barrier to entry.
The net competitive dynamic favors NVIDIA's ecosystem approach. By open-sourcing the reference architecture and partnering with cloud providers rather than building a proprietary service, NVIDIA positions itself to capture value at the GPU layer — where training and inference run — regardless of which robotics company wins at the application layer.
If you are building a robotics product and this announcement is news to you, the practical next step depends on where you are in the development cycle.
Early-stage teams building manipulation or mobile navigation capabilities should evaluate the Physical AI Data Factory reference architecture before investing engineering time in building a custom simulation pipeline. The architecture is open, which means you can adapt it without licensing costs. Start with the Omniverse foundation — even if you never use any other NVIDIA-specific component, having your environments in USD format keeps your options open as the ecosystem matures.
Mid-stage teams with existing simulation pipelines should assess whether their data generation bottlenecks align with what the factory architecture addresses. If your primary constraint is scene diversity and volume, the procedural generation components of the reference architecture are the most directly applicable. If your constraint is sim-to-real transfer, the Newton physics integration and domain randomization pipeline are worth close examination. The robotics funding environment means capital is available to move fast here.
Enterprise teams at manufacturers or logistics operators who are deploying or planning to deploy robots should look at the Azure integration path. If you already have an Azure relationship and are thinking about digital twins for your facilities, the Physical AI Data Factory's Azure deployment option gives you a way to connect those digital twins to robot training pipelines without standing up specialized on-premises GPU infrastructure.
Research teams should watch the open architecture documentation closely. The reference design will evolve, and the curation and filtering components — the quality control layer of the data factory — are where the most interesting research questions live. What constitutes a high-quality synthetic demonstration? How do you measure and optimize the diversity of a synthetic dataset? These are open problems that the architecture exposes without solving.
The broader signal to take from the Physical AI Data Factory announcement is that the robotics industry is entering a phase where data infrastructure is becoming a first-class engineering discipline — not a side task for researchers, but a core production capability that serious robotics organizations need to build or buy. NVIDIA is offering a version of "build" that is substantially cheaper than starting from scratch. That is a meaningful shift.
Is the Physical AI Data Factory a cloud service NVIDIA operates, or a reference architecture companies deploy themselves?
It is a reference architecture — an open blueprint for building a synthetic data pipeline — not a managed service. NVIDIA provides the specifications, the component integrations (Omniverse, Newton, GR00T), and the cloud deployment patterns. Microsoft Azure and Nebius are infrastructure partners where you can deploy the pipeline, but NVIDIA is not operating it as a service you sign up for. This is a deliberate choice that reflects NVIDIA's ecosystem strategy: capture value at the GPU infrastructure layer, keep the software open to maximize adoption.
How does the sim-to-real transfer problem affect the value of synthetic data generated by this pipeline?
Sim-to-real transfer — the gap between how a robot behaves in simulation and how it behaves in the real world — remains the central technical challenge for synthetic data-based training. The Physical AI Data Factory addresses this through domain randomization (varying simulated conditions to force generalization), high-fidelity physics simulation via Newton (reducing systematic inaccuracies), and photorealistic rendering via Omniverse. These techniques reduce the gap but do not eliminate it. For manipulation tasks especially, contact dynamics in simulation remain imperfect. The pipeline should be understood as raising the ceiling on how much useful training data you can generate synthetically, not as a complete replacement for real-world validation.
What is the relationship between the Physical AI Data Factory and NVIDIA Cosmos?
Cosmos is a world foundation model — a generative model that can produce physically plausible video of robot interactions from text or image prompts, or from a policy trajectory. It is one possible source of synthetic training data. The Physical AI Data Factory is the broader pipeline architecture that manages how data is generated (Cosmos is one generation option), curated, filtered, versioned, and fed into training. Cosmos generates data; the factory decides which data is worth keeping and how to structure it for training.
Are there open-source components I can start using today?
NVIDIA has published the reference architecture documentation and is releasing compatible tooling through its Isaac platform. Isaac Lab, the robot training framework, is available on GitHub. Omniverse's core components are available through NVIDIA's developer program. The Newton physics engine is being released open-source. The full pipeline integration, including the cloud deployment configurations for Azure and Nebius, is being rolled out in phases following the GTC announcement. Check NVIDIA's developer portal for the current availability status of each component.
How does this announcement affect smaller robotics startups that cannot afford large-scale GPU infrastructure?
The cloud deployment options with Azure and Nebius make the pipeline accessible to organizations that cannot afford to buy or lease on-premises GPU clusters. You pay for compute time on demand rather than capital expenditure on hardware. For a startup running a few thousand synthetic episodes to validate a task policy, this significantly lowers the barrier. For a startup trying to generate the tens of millions of episodes needed to train a general-purpose foundation model, cloud costs at scale are still substantial — but the cost-per-episode of synthetic data is orders of magnitude lower than the cost-per-episode of real-world teleoperation data, which makes the economics work even at large scale compared to the alternative.
NVIDIA launches GR00T N2 robots that complete new tasks 2x more often than leading models, Cosmos 3 world simulator, and Newton physics engine — with Disney's Olaf and Boston Dynamics among first partners.
Disney debuts a free-walking Olaf robot at Disneyland — the first untethered, AI-powered character robot from a major entertainment company, marking the beginning of physical AI in theme parks and live entertainment.
NVIDIA GTC 2026 runs March 16-19 in San Jose with the Rubin platform, Jetson T4000, physical AI breakthroughs, and robotics partnerships with Boston Dynamics and more.