Nvidia didn't announce a new chip this time. Instead, it released a model — and it's open-source.
Cosmos 3, which CEO Jensen Huang calls an "open frontier foundation model for Physical AI," aims to give robots and self-driving cars the ability to "imagine" what will happen next in the physical world and then decide how to move.
A single model for seeing, thinking, and acting
Cosmos 3 is built on a hybrid transformer architecture: one transformer for reasoning and another for generation. The reasoning transformer first understands object interactions, motion, and spatial-temporal relationships; the generation transformer then produces videos and action trajectories.
What sets it apart is its ability to directly output "action data" — joint angles, gripper openings, and movement paths. Previously, such data had to be collected manually in real environments; now the model can synthesize it.
Nvidia says this reduces the training and evaluation cycle for Physical AI from months to days.
Three sizes, smallest fits in a workstation
Nvidia released three versions at once:
- Cosmos 3 Nano (16B parameters) — runs on a workstation with a single RTX PRO 6000
- Cosmos 3 Super (64B parameters) — for data centers with Hopper/Blackwell
- Cosmos 3 Edge (to be released) — for edge real-time inference
The 16B size is notable: small robotics teams can run it on their own workstations without renting cloud compute.
In benchmarks, Nvidia claims Cosmos 3 tops multiple leaderboards: Physics-IQ, PAI-Bench, R-Bench for world generation; RoboArena for action policy; and several visual understanding benchmarks. Of course, these are Nvidia's own selections; real-world performance awaits community testing.
Nvidia also forms a Cosmos Alliance
Beyond open-sourcing, Nvidia launched the "Cosmos Alliance," whose founding members include Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI — a mix of robotics and video-generation companies.
Jensen Huang said: "The Cosmos 3 family of open, frontier omnimodels gives developers a generational leap in ability to build robots, autonomous vehicles and vision AI."
The strategic choice is interesting. The Physical AI space is crowded, but most companies keep their models closed-source. Nvidia is doing the opposite: open-sourcing the base model and building an alliance. Its bet isn't on any single robot selling well — it's on everyone using its models and chips to train robots. The "selling shovels" strategy, once again.
Sources: NVIDIA Newsroom; Tech Startups; Axios; CocoLoop