NVIDIA opens Cosmos to robotics and driving developers

The shot

NVIDIA announced Cosmos on 6 January 2025, in a press release timed to CEO Jensen Huang's CES keynote. The company described Cosmos as "a platform comprising state-of-the-art generative world foundation models, advanced tokenizers, guardrails and an accelerated video processing pipeline built to advance the development of physical AI systems such as autonomous vehicles (AVs) and robots," and stated the models would be available "under an open model license." Huang is quoted stating NVIDIA "created Cosmos to democratize physical AI and put general robotics in reach of every developer."

What the documents show

NVIDIA's own framing at launch positioned Cosmos toward robotics and autonomous-vehicle development rather than filmmaking, and the press release's specific figures, including a stated ability to "process, curate and label 20 million hours of videos in 14 days" using NVIDIA's Blackwell hardware, are the company's own reported claims. The project's repository, as retrieved on 16 September 2026, shows the platform has since expanded into an omnimodal "Cosmos 3" model family and states that Cosmos source code and models are now released under the OpenMDW-1.1 license rather than the "NVIDIA Open Model License" the January 2025 launch materials named, a change worth flagging precisely because both documents describe themselves as currently governing. The OpenMDW-1.1 text itself grants permission "to deal in the Model Materials without restriction," with no revenue threshold of the kind seen on some other open releases in this pack.

The workflow

For a production, Cosmos is documented as infrastructure for training and evaluating physical-AI systems, including generating synthetic training video for robotics and driving models, rather than a tool for finishing a shot. Where a visual-effects or virtual-production team's own pipeline depends on synthetic environment data, the repository states developers can access post-training scripts and NIM microservices for deployment; the human decision that remains is confirming which license version, and which specific model in the now-larger Cosmos family, applies to a particular downloaded checkpoint, since the terms have changed since launch.

What the tool does not change

NVIDIA's own materials describe Cosmos's target use as robotics, autonomous vehicles, and physical-AI simulation; nothing in the launch release or the current repository documents a stated filmmaking or visual-storytelling use case, and this site's coverage accordingly treats Cosmos as adjacent infrastructure rather than a production tool, a distinction that is editorial where the sources themselves are silent on the point.

• Which Cosmos model version and license actually applies to a checkpoint under evaluation, given the change since the January 2025 launch?
• Does a specific synthetic-data use case fall within NVIDIA's own stated purpose for the platform, or outside it?
• What independent verification exists for NVIDIA's stated data-processing throughput figures?

Cosmos brought a major hardware and platform company into open video-generation infrastructure, but its own documents point that infrastructure at robotics and simulation, not at the shot on a production's timeline.