Two million robots are welding, assembling, and palletizing in factories worldwide — and most still run hand-coded routines written years ago. At GTC 2026, NVIDIA shipped the full stack to rewire them.

NVIDIA released an integrated platform spanning data generation, simulation, training, and edge deployment. The centerpiece: Cosmos 3, the first world foundation model unifying synthetic world generation, vision reasoning, and action simulation. A developer scans one real-world scenario; Cosmos 3 generates thousands of synthetic training variations. Gartner projects synthetic data will hit 90% of AI training data by 2030 — NVIDIA wants Cosmos to be the engine.

Alongside it: Isaac GR00T N1.7, an open vision-language-action model pairing fast-thinking reflexes with slow-thinking reasoning to translate broad instructions into precise joint movements. The GR00T X-Embodiment dataset has crossed 10M downloads on Hugging Face.

NVIDIA reported that combining Cosmos-generated synthetic data with real-world demonstrations improved GR00T N1’s task success rate by 40% over real-data-only training — a concrete operator metric for anyone weighing the sim-to-real gap.

Isaac Lab 3.0 runs thousands of parallel training scenarios. Omniverse NuRec converts sensor data into digital twins via 3D Gaussian splatting. Isaac Teleop captures human demonstrations from XR headsets and gloves. The pipeline: scan a factory, simulate it, train virtually, deploy physically. Training runs in the cloud; execution at the edge. Jetson Thor and Jetson Orin handle real-time inference on production robots.

Why It Matters for Operators.

NVIDIA is not building robots. It is building the platform every robot maker builds on. FANUC — the world’s largest industrial robot supplier, spanning 3kg to 2.3-ton payloads — announced full integration of Jetson edge, Isaac Sim, and Omniverse digital twins across its portfolio. Foxconn and Samsung deploy NVIDIA-powered AI on electronics assembly lines. The Halos safety framework provides end-to-end guardrails from cloud training to factory floor, targeting the compliance barrier that stalls enterprise adoption.

Why It Matters for Investors.

NVIDIA is embedding itself as infrastructure across the Physical AI value chain — Brains (foundation models), Simulation (digital twins), Edge (Jetson compute). Agility, 1X, Figure, and Boston Dynamics build on this stack. The company that controls the development platform captures recurring revenue from every robot deployed on it. At $4.45T market cap, the market is pricing in this ambition.

What Could Go Wrong

Open-source cuts both ways. The stack is composable — competitors can swap in alternative models or edge hardware. If a rival foundation model outperforms GR00T on factory tasks, the ecosystem fragments. And sim-to-real transfer remains the hardest unsolved problem in robotics — digital twins are only as good as the physics they model. The gap between a simulated warehouse and a real one at 2 AM with a wet floor is where deployments stall.