Nvidia’s Cosmos 3 Wants to Become the CUDA of Physical AI

Nvidia shipped Cosmos 3 on May 31, a free foundation model trained on 20 trillion tokens of data built to teach robots and self-driving cars how the physical world behaves. It reasons about a scene, predicts what happens next, then outputs the numbers that steer a machine. The release folds four separate Nvidia products into one open system.

Giving away a frontier model costs Nvidia nothing in license fees. It buys something bigger: developers who build their robots, cars and factory systems on Nvidia’s architecture instead of a rival’s.

One Model Replaces Four Separate Systems

The big bang of physical AI is just around the corner thanks to breakthroughs in multimodal reasoning language, vision and world models.

Jensen Huang, Nvidia’s founder and chief executive, said that when Cosmos 3 launched at GTC Taipei. Until now, building on Nvidia’s Cosmos platform meant juggling separate models: Cosmos Predict for generating a simulated world, Cosmos Transfer for controlled scene generation, Cosmos Reason for making sense of what a camera sees, and Cosmos Policy for turning that understanding into a robot’s next move.

Cosmos 3 folds all four into a single mixture-of-transformers (MoT) architecture, pairing a reasoning transformer with a generation transformer. The reasoning half studies motion, physics and cause and effect. The generation half turns that understanding into pixels, sound or a robot’s next move.

Nvidia shipped two sizes immediately. The 16 billion parameter Nano variant runs on a single workstation GPU and produces results in fractions of a second. Super, at 64 billion parameters, targets data center hardware for the highest physics accuracy. A third size, Edge, is coming for real-time inference on the device itself.

Every size shares the same architecture, so a developer can prototype on Nano and move to Super without rebuilding a pipeline.

Why Can’t Robots Learn the Way Chatbots Did?

Robots and self-driving cars can’t learn from scraped text the way chatbots did. The physical mistakes that matter most, a pedestrian stepping between parked cars, a forklift’s blind turn, are too rare or dangerous to collect at scale. World foundation models fill that gap with synthetic scenarios that still obey real physics.

An autonomous vehicle developer would once have needed years of test-fleet driving to encounter enough fog, black ice and near misses. Nvidia says a developer can instead run millions of simulated scenarios in days, generating rare or dangerous events, a robot collision, an unusual road event, without repeating them on a real street or factory floor.

• 20 trillion tokens of multimodal training data, spanning text, image, video, audio and action.
• Nearly a billion images and 400 million real and synthetic videos feed the model, according to Nvidia.
• Months to days: how far Nvidia says Cosmos 3 compresses physical AI training and evaluation cycles.

The hardware Nvidia sells underneath all of this is not incidental. Nvidia says peak performance on NVIDIA Blackwell GB200 is reserved for the heaviest Cosmos workloads, meaning the fastest path through Cosmos 3 still runs through Nvidia’s own chips.

The Free Price Tag Is the Strategy

Axios reported plainly that Nvidia is continuing its move beyond chips into AI models and software, positioning itself to become a foundational platform for physical AI development. Ming-Yu Liu, vice president of Nvidia’s Cosmos Lab, told the outlet that a world model exists to help “physical agents to become more generalizable.”

Turing Post, an AI research newsletter, still credited the ambition even while flagging rough edges elsewhere in the release. Automotive World described the underlying strategy in blunter terms: Nvidia is betting that infrastructure pull-through matters more than model licensing revenue. The more of the physical AI industry that standardizes on Cosmos, the outlet argued, the harder it becomes for rivals to build a credible alternative at comparable scale.

A similar point came from Silicon Report: developers and hardware makers who start from Nvidia’s base model keep their customization and future alignment work inside an Nvidia-centered stack. That pattern echoes Nvidia’s own climb to become the world’s most valuable company, built less on any single chip than on making its software the default layer underneath an entire industry.

Nvidia has run a version of this playbook before. Its revenue-sharing arrangements with AI startups already tie compute access to continued use of Nvidia’s stack. Cosmos 3 extends the same logic to physical AI: the model is free, but the fastest, most capable path through it runs on Nvidia silicon.

Nvidia released Cosmos 3 under the Linux Foundation’s OpenMDW 1.1 open license, which permits commercial use and covers weights, code, datasets and benchmarks. Nvidia also formed the Cosmos Coalition, a group of AI labs and robotics companies developing the platform jointly.

The Companies Already Running on Cosmos

Nvidia named working partners at launch, not a hypothetical list.

• Agile Robots, Doosan Robotics, LG Electronics and Samsung Electronics are building humanoid and industrial robotics applications on the platform.
• Li Auto is using Cosmos 3 for autonomous vehicle development.
• Waabi, Wayve and Foretellix use Nvidia’s Cosmos models to simulate traffic, weather and pedestrian behavior without physical test drives, according to AI Multiple.
• Centific, Fogsphere, Linker Vision and Milestone Systems build vision AI agents for warehouses, traffic monitoring and public safety.

Mercedes-Benz launched its first premium robotaxi service on the Uber network in January using Nvidia’s physical AI stack built on Cosmos, 4D Pipeline reported. Outside partners can also pull the model apart themselves: Nvidia has posted the underlying training and inference code on GitHub, letting outside teams fine-tune Cosmos 3 on their own robots, cameras or driving logs instead of starting from scratch.

Three Well-Funded Rivals Want the Same Prize

Cosmos 3 is not the only bid to become the operating layer under physical AI. Axios named Fei-Fei Li’s World Labs and Yann LeCun’s AMI Labs among the hot startups chasing the same category. PYMNTS has reported that World Labs is in talks at a five billion dollar valuation and AMI Labs is seeking three billion, while Google DeepMind’s Genie 3 is positioned as a direct rival.

The split so far is about focus, not just funding. Genie 3 excels at generating novel environments from text prompts. Cosmos 3 holds itself to strict physical consistency for industrial use, the kind of accuracy a robot arm or a delivery van actually needs.

Every one of them is chasing the same customer: a robotics or automotive team that needs synthetic data it can trust. Nvidia’s advantage is that its version is free and already has a coalition of labs feeding it.

Where the Physics Still Breaks

Nvidia’s own numbers are aggressive. Among open models, Cosmos 3 ranks first on Artificial Analysis, Physics-IQ, PAI-Bench and R-Bench for world generation, first on RoboLab and RoboArena for action policy quality, and first on VANTAGE-Bench and TAR for vision understanding.

Outside observers are less sweeping. Turing Post wrote that Cosmos 3 is “still early and uneven” in places. Nvidia’s own developer blog built a separate human evaluation framework, called HUE, because automated leaderboards have grown saturated: score gaps between rival video models are often too narrow for meaningful comparison, so Nvidia now checks fact by fact whether a generated video actually obeys physical laws.

The deeper caution comes from Nvidia’s own research team. Its original Cosmos technical paper acknowledged that world foundation models suffer from a lack of object permanence and inaccuracies in contact-rich dynamics, and that generated video does not always reflect gravity, light or fluid behavior correctly. Cosmos 3 is a new generation built on top of that same unsolved category of problem.

Nvidia has not said when Cosmos 3 Edge, the on-device version, will actually ship. Until then, the heaviest physics testing still happens inside a data center.

Frequently Asked Questions

What Is a World Foundation Model?

A world foundation model is an AI system trained to simulate how the physical world behaves, rather than just describe it in text. Instead of predicting the next word in a sentence, it predicts the next state of a scene: where an object will be, how a robot arm should move, or what a pedestrian is likely to do next. Cosmos 3 is Nvidia’s version, but the category also includes Google DeepMind’s Genie 3 and startups like World Labs and AMI Labs.

Is Nvidia Cosmos 3 Actually Free to Use?

The weights, code and tooling are free to download and commercially licensable under the OpenMDW 1.1 license. Nvidia has not published the complete 20 trillion token training dataset itself, so outside teams can build on Cosmos 3 without being able to fully reproduce it from scratch.

What Makes Cosmos 3 Different From a Video Generator?

A standard video generator takes a text prompt and outputs a clip. Cosmos 3 also outputs action data, joint angles, gripper positions and trajectory waypoints, that tells a real robot arm or vehicle what to do next, then can run the result forward or backward to check whether the predicted motion actually holds up physically.

What Is the Cosmos Coalition?

The Cosmos Coalition is a group Nvidia formed alongside the Cosmos 3 launch to develop open world models jointly. Its six founding members are Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI, each contributing models, research or evaluation techniques back to the shared platform.

How Does Cosmos 3 Compare to Google’s Genie 3?

Genie 3 is built to generate novel environments from a text description, useful for creating new worlds on demand. Cosmos 3 is tuned instead for strict physical consistency, the kind of accuracy that matters when the output trains an actual robot or vehicle rather than a game or a simulation built for its own sake.