NVIDIA is making the brain for robots free. Cosmos 3, released June 1, is the first fully open foundation model for physical AI, with launch partners Runway, Black Forest Labs, and Skild AI. Just as Meta goes closed on text AI, NVIDIA goes open on robots.
NVIDIA shipped the first fully open foundation model for physical AI on June 1. It targets robotics, autonomous driving, and any system reasoning about the physical world.
The model handles text, images, video, ambient sound, and physical actions in one system. NVIDIA also launched the Cosmos Coalition with six partners including Runway and Black Forest Labs. Cosmos 3 reduces robot training cycles from months to days, per NVIDIA.
The strategic contrast with Meta is sharp - closed for text, open for physical. NVIDIA bets the open standard for robot models is worth more than a proprietary one.
⚡ Why this matters
- First open frontier model for physical AI. Sets the de facto standard for the next robotics generation.
- The Cosmos Coalition (Runway, Black Forest Labs, Generalist, LTX, Agile Robots, Skild AI) signals industry alignment around NVIDIA's stack. This is positioning, not just product.
- NVIDIA's open bet on physical AI directly contradicts Meta's closed bet on text AI (Muse Spark, April 8). Two opposite plays on category economics, same year.
🔍 What happened
- June 1, 2026. NVIDIA launches Cosmos 3 - built on a mixture-of-transformers architecture combining vision, world generation, and action prediction in one system.
- Native modalities: text, image, video, ambient sound, AND physical actions (joint angles, gripper positions, trajectory points).
- Variants: Cosmos 3 Super (high-fidelity post-training), Cosmos 3 Nano (sub-second inference), Cosmos 3 Edge (real-time, coming soon).
- Released as fully open: models, training scripts, deployment tools, and datasets all available.
- Cosmos Coalition launched alongside: Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI. Members contribute models and research while using Cosmos 3 plus NVIDIA DGX Cloud.
- NVIDIA claim: physical AI training and evaluation cycles drop from months to days.
💬 Smart takes
- NVIDIA newsroom: 'the world's first fully open omnimodel with native vision reasoning and multimodal generation.'
- WinBuzzer: framed it as NVIDIA betting physical AI needs an open standard like CUDA was for compute.
- Skeptic: the same Cosmos Coalition partners (Black Forest Labs, Runway) compete with each other and with NVIDIA-owned products. 'Open' may not survive contact with real revenue pressure.
🧭 Where this goes
- LikelyCosmos 3 becomes the default starting point for robotics teams without an in-house foundation model - replacing one-off training stacks within 12 months.
- LikelyTesla, Boston Dynamics, and Waymo respond by either joining the coalition or doubling down on closed proprietary stacks within 6 months.
- PossibleGoogle releases a competing open physical AI model under Gemini Robotics branding within 12 months.
- Possibleagent-runtime cloud platforms (Anthropic, OpenAI) extend their managed agent offerings to include physical-AI action loops, using Cosmos 3 as the substrate.
- Wild Carda Cosmos Coalition partner gets acquired by NVIDIA, breaking the 'open coalition' framing and triggering a competing coalition led by AMD or Cerebras.
🥄 The Spoon Take
NVIDIA just did to robotics what CUDA did to compute - set the open default, then sell the GPUs underneath. Every Cosmos Coalition partner who trains on the model locks into NVIDIA infrastructure for years. Meta closed its model; NVIDIA opened theirs and won the customer.
🤔 Pushback
'Open' often means free-as-in-puppy - Cosmos 3 only counts if non-NVIDIA teams can train it on non-NVIDIA hardware, and the license doesn't yet say.