Jensen Huang’s CES 2026 announcement crystallized the Physical AI infrastructure stack as three connected computers working in continuous loops.
The Physical AI Development Pipeline
1. Training (Cloud/Data Center)
| Function | NVIDIA Platform |
|---|---|
| Build foundation models | DGX GB300 |
| Generate synthetic data | Blackwell + Grace supercomputer pods |
| Train VLA architectures | NVLink multi-GPU fabric |
Physical AI Role: Generates robotic policies, trains VLA models (GR00T, OpenVLA, Octo)
2. Simulation (Workstation/Cloud)
| Function | NVIDIA Platform |
|---|---|
| Digital twin creation | RTX Pro Blackwell |
| Physics-based testing | Isaac Sim robotics simulator |
| Synthetic data generation | Cosmos World Foundation Models |
Physical AI Role: Tests millions of scenarios before real-world deployment, data multiplication + cost reduction
3. Inference (Edge/On-Device)
| Function | NVIDIA Platform |
|---|---|
| On-device decision making | Jetson Thor |
| Real-time perception | 1 PFLOP on-device, no competitor |
| Sub-millisecond response | Edge chips + low latency |
Physical AI Role: Real-time perception, reasoning, and action
The Key Insight
Physical AI requires all three computers working in continuous loops—not sequential handoffs. A warehouse robot doesn’t just run inference; it generates operational data that feeds back to training, while simulation validates policy updates before deployment.
This analysis is part of a comprehensive report. Read the full analysis: Physical AI Is Crossing the Manufacturing Chasm on The Business Engineer.









