NVIDIA Released an 'Open' Physical AI Model. The License Bans You From Using It to Compete With NVIDIA.

On June 1, at GTC Taipei, NVIDIA released the weights for Cosmos 3. It is the first open model to unify scene understanding, physical reasoning, and action generation in a single forward pass. Text in, robot trajectory out. The architecture is a mixture-of-transformers — an autoregressive reasoning transformer working alongside a diffusion generation transformer, running jointly rather than in sequence. Nano variant: 8 billion parameters each. Super variant: 32 billion parameters each. Edge variant: coming.

The practical capability: Cosmos 3 can take a video of a workspace and produce joint angles, gripper positions, and trajectory waypoints that a robot arm can execute. It can generate synthetic video of a robot performing a task it has never physically done. It can simulate what happens when a physical environment changes and produce corrected trajectories. These are not new ideas in robotics; they are new capabilities in a single open model at frontier scale.

The weights are on Hugging Face. The benchmarks against other open models are favorable: NVIDIA reports first place on RoboLab, RoboArena, R-Bench, and Artificial Analysis. Coverage of the launch has treated the "open" framing as the headline. The "open" framing requires a closer read.

What "open" means here — and what it doesn't

The OpenMDW-1.1 license governs Cosmos 3. Under this license: commercial use is permitted; modification is permitted; training your own models using Cosmos 3 as a base is permitted; redistribution is permitted; you must include "Built on NVIDIA Cosmos" attribution in your product.

One restriction: Cosmos 3 cannot be used to develop a competing foundation model product.

Translated into practical terms: a robotics startup can fine-tune Cosmos 3 on their proprietary robot data and sell the resulting robot behavior as a product. That startup cannot take Cosmos 3, improve the world model architecture, and sell the improved world model as a competing physical AI foundation model — a Cosmos alternative.

The distinction is precise and strategic. NVIDIA is open enough for developers to build on the foundation. NVIDIA is not open enough for developers to replace the foundation. Any ecosystem built on Cosmos 3 is dependent on NVIDIA's roadmap for Cosmos 4 and beyond. There is no fork that removes this dependency, because a fork built into a competing foundation model is prohibited by the license.

The matthieupesesse.com analysis of the launch is the clearest explication of this: "Open here means open enough to use, not open enough to replace." Five definitions of open apply to different aspects of AI models: weights (downloadable — yes), commercial use (permitted — yes with restrictions), source code (partial), training data (not released), governance (NVIDIA controls roadmap — no external steering). Cosmos 3 passes the first two; fails the last three.

None of the coverage of the Cosmos 3 launch explained the competing-model prohibition. The "open weights" headline is technically accurate. The strategic implication — that the open model is designed to make you dependent on NVIDIA's foundation layer — requires reading the license, which most coverage did not.

Why NVIDIA can afford to give it away

NVIDIA's revenue is GPU sales. Cosmos 3 does not threaten GPU sales. It generates GPU demand.

Every developer who downloads Cosmos 3 and fine-tunes it on robot data needs GPU compute. Every inference call from a deployed Cosmos 3-based robot policy runs on GPU compute. The Cosmos Coalition members get DGX Cloud access as part of their coalition participation — NVIDIA's cloud compute platform. Nebius is building dedicated cloud infrastructure for Cosmos workloads. The entire ecosystem is architected to convert Cosmos 3 downloads into GPU purchases.

Jensen Huang made this explicit at GTC: NVIDIA gives away the model; developers pay for the compute to train and run it. The same strategy underpins CUDA (free GPU programming framework → GPU sales), Omniverse (free simulation platform → GPU sales), and every NVIDIA open SDK. The open model is the demand-generation mechanism for the hardware business.

Cosmos 3's GPU alignment is even more explicit: the Nano variant runs on an RTX PRO 6000; the Super variant runs on Hopper and Blackwell; the Edge variant (coming) targets Jetson for embedded deployment. The model size variants match NVIDIA's GPU product line. This is not coincidence — NVIDIA designed the model portfolio to ensure that every tier of developer buys NVIDIA hardware to run it.

The "open" model is the most effective GPU sales tool NVIDIA has ever made.

The Cosmos Coalition and the member nobody expected

The six founding members of the Cosmos Coalition: Agile Robots, Black Forest Labs, LTX, Runway, Skild AI, Generalist.

Runway does not build robots. Runway Gen-4.5 is one of the leading AI film and VFX generation platforms. Runway is in a robotics AI coalition because the same world-simulation model useful for robot policy training is useful for generating physically coherent video.

The problem Runway faces with generative video is physics grounding: generated video looks wrong because the model doesn't understand how physical objects behave — how a fabric hangs under gravity, how a liquid pours, how a mechanical arm moves through space with constraints. Cosmos 3's world model was trained on billions of samples including physical action trajectories. Its physics grounding is accurate because it has to be to generate valid robot actions.

Runway licensing or integrating Cosmos 3's physics layer produces generative video that looks physically real without Runway having to build the world model from scratch. NVIDIA gets a $1.5 billion generative video company as a distribution partner. Runway gets frontier physics simulation it couldn't build in-house.

NVIDIA is betting that the physical AI world model market and the generative video market converge. If Cosmos 3 becomes the standard physics engine for AI-generated film and VFX — not just robot policy training — the addressable market expands by an order of magnitude. Runway's coalition membership is the signal that NVIDIA has already made this bet internally.

The players who aren't in the coalition

Figure AI, Boston Dynamics, Agility Robotics, 1X, Tesla Optimus — the humanoid platforms with actual hardware in production — are not members of the Cosmos Coalition.

Samsung, LG, and Doosan (an industrial arm manufacturer) are confirmed early adopters. Agile Robots, Skild AI, and Generalist are policy-layer companies building on top of the world model. Black Forest Labs, LTX, and Runway are video/media companies.

NVIDIA's "Android for physical AI" thesis — that Cosmos 3 becomes the standard platform layer for robots the way Android became the platform layer for smartphones — requires the hardware OEMs to adopt it. Samsung and LG building home robot products on Cosmos 3 is notable. It is not the same as Figure or Boston Dynamics adopting Cosmos 3 as their policy training foundation.

The benchmark claim — "#1 among open models on RoboLab and RoboArena" — is qualified to "open models." The relevant competition is the closed robotics AI systems that the major humanoid OEMs use internally: Figure's training pipeline, Boston Dynamics' Spot AI, OpenAI's robotics research. NVIDIA has not published a comparison against these closed systems. No coverage has asked for one.

The "Android for robots" thesis is unvalidated until a major hardware OEM with shipping products formally adopts Cosmos 3. Until then, Cosmos 3 is the best open model in a competition that closed systems are not participating in.

The Prometheus question

The physical AI wave produced three significant announcements in June alone: Cosmos 3 (NVIDIA, world model infrastructure), PhysicsX ($300 million, engineering simulation acceleration), and Prometheus ($12 billion, Bezos's "artificial general engineer" for design pre-production). The relationship between these three is the unresolved strategic question no coverage has addressed.

Prometheus is building an AI system for pre-production engineering design — simulation, prototyping, generative CAD. Cosmos 3 generates physics-accurate synthetic video and robot action trajectories. The underlying capability that makes Cosmos 3 useful for robots — physics-grounded world simulation — is the same capability that would be useful for Prometheus's engineering pre-production simulation.

Does Prometheus build on Cosmos 3 as its physics simulation substrate? If yes, Prometheus is part of the Cosmos ecosystem whether or not it's in the coalition — and NVIDIA becomes an infrastructure dependency for the $41 billion startup Jeff Bezos is personally co-CEOing.

If no, Prometheus is building its own world model from scratch — which is the product Cosmos 3 is. Bezos recruited from NVIDIA, including researchers who worked on physical simulation. The strategic relationship between Prometheus and NVIDIA's physical AI stack is one of the most important unanswered questions about both companies.

Neither NVIDIA nor Prometheus has addressed it. NVIDIA didn't include Prometheus in the coalition. Prometheus hasn't disclosed whether it uses or avoids Cosmos 3.

What Cosmos 3 actually is

It is the most sophisticated GPU sales funnel in AI history. The model is technically real, the architecture is a genuine advance over stitched-model pipelines, and the open weights are genuinely accessible. The license is strategically designed to prevent the open model from undermining NVIDIA's foundation layer position.

The "first open omnimodel for physical AI" framing is accurate and incomplete. The competing model prohibition is the clause that defines the actual terms of the ecosystem NVIDIA is building. Developers who commit to Cosmos 3 as their world model foundation are choosing NVIDIA's roadmap for the physical AI layer they're building on. That is a real bet — the same bet developers made on Android, on CUDA, on Omniverse. NVIDIA has delivered on those ecosystems. It is also the same constraint those ecosystems carry: you are building on someone else's foundation, and they control what the next version looks like.