NVIDIA argues open “world models” are the next battleground for physical AI

NVIDIA is using an open-model push to frame the next phase of artificial intelligence development as a physical one, where systems must understand consequences, not just generate convincing images. In a new post in its “Into the Omniverse” series, the company ties progress in “open world models” to the practical needs of teams building robots, self-driving systems and computer-vision agents, where data collection is expensive and failures can be hard to reproduce safely.

The company also points to a July industry effort, joining more than 200 companies and organizations in signing an open letter titled “Open Weights and American AI Leadership.” NVIDIA’s argument is that leadership will be measured not by a single top “frontier model,” but by whether an open ecosystem can reach every sector. The post emphasizes that open models, which are downloadable and can be inspected, modified and run on a customer’s own infrastructure, are what make that ecosystem possible.

For “physical AI,” NVIDIA says the key bottleneck is data. The company describes physical AI deployments as specialization problems, where a general model has not seen a specific team’s robot, sensors, environment or operating conditions. Bridging that gap, it argues, requires access to model weights, a license that permits adaptation, and tools for post-training, so teams can tailor capabilities without rebuilding everything from scratch.

NVIDIA’s solution is presented as an end-to-end foundation for world modeling and downstream specialization. The company describes “NVIDIA Cosmos 3” as an open model family built on a mixture-of-transformers architecture, combining vision reasoning, world generation and action prediction. It says teams can use Cosmos 3 as a vision-language model, as a physics-grounded world simulator that predicts future states and generates synthetic training data, and as a backbone for “world action models,” so developers can avoid assembling and maintaining separate models for each capability.

The post details three Cosmos 3 variants aimed at different compute and deployment scenarios. “Cosmos 3 Super (64B)” is positioned for high-fidelity world modeling, “Cosmos 3 Nano (16B)” for efficient reasoning and post-training, and “Cosmos 3 Edge (4B)” for on-device vision reasoning and robot policy deployment. NVIDIA says Cosmos 3 Edge is designed to run on edge GPUs and can be deployed across NVIDIA RTX GPUs, NVIDIA DGX systems and NVIDIA Jetson, including Jetson Thor platforms.

On performance, NVIDIA cites benchmark rankings for Cosmos 3 across multiple tasks. It says Cosmos 3 is No. 1 on Artificial Analysis for open-weights text-to-image and image-to-video generation, No. 1 on PAI-Bench for world generation and No. 1 in the image-to-video category of Physics-IQ. For robot policy, it says Cosmos 3 ranks No. 1 on RoboLab. The company also states Cosmos 3 Super is the highest-ranked open model on VANTAGE-Bench for vision understanding, and it frames those results as evidence that open weights do not prevent competitive quality.

Beyond the model itself, NVIDIA links specialization to simulation and data generation workflows. The company says its Omniverse libraries, part of NVIDIA Agent Toolkit, provide prebuilt capabilities for building simulation-ready worlds, and that OpenUSD acts as an open framework for composing, reusing and exchanging complex 3D data across “digital twins,” simulations and synthetic data generation. NVIDIA argues that pairing Omniverse with OpenUSD can reduce duplicated work when teams change assets, sensor setups or environmental conditions.

NVIDIA also lays out licensing terms intended to make post-training feasible. It says “NVIDIA Cosmos world foundation models are available under the Linux Foundation’s OpenMDW 1.1 license,” which it describes as enabling teams to post-train models on their own data and hardware. NVIDIA’s broader physical AI stack is also referenced, including Isaac GR00T for robotics, Alpamayo for autonomous vehicles, and Metropolis for vision AI.

The company’s post does not provide specific revenue impact or customer adoption metrics, and it does not describe pricing or any binding commitments behind the open letter. It also relies on benchmark results and adoption examples without detailing how frequently customers choose open weights versus proprietary model access, or how performance holds up in real-world deployments where edge cases and rare events can dominate outcomes.

What to watch next is whether NVIDIA’s open-model approach spreads beyond software experimentation into large-scale production pipelines. The post says NVIDIA has expanded a “Cosmos Coalition” and that robotics and manufacturing leaders in Japan intend to join to develop open world models for sectors including factories, logistics, agriculture, construction, healthcare and transportation. If those collaborations translate into shared models and repeatable simulation workflows, the market could begin treating licensing and integration capability as much as raw model scores. Meanwhile, customers will likely look for clearer evidence on how easily open weights can be specialized to their own robots, sensors and operating constraints.