The Next Wave of AI: Physical AI
The Next Wave of AI: Physical AI is a three-minute official NVIDIA overview that compresses the company's embodied-AI pitch into one stack. The video moves from Software 1.0 code on CPUs, to Software 2.0 machine learning on GPUs, to generative AI, then splits the next phase into agentic AI for digital work and physical AI for work in the world. Physical AI here means systems that perceive, reason, plan, and act through robots, self-driving cars, manipulators, humanoids, factories, plants, sensors, and other physical infrastructure.
The core claim is not only "robots will use AI." It is that robot development becomes a three-computer pipeline: train models on DGX, fine-tune and test them inside NVIDIA Omniverse and Isaac-style simulation, then run the trained AI on Jetson AGX robotics computers. NVIDIA's current robotics page and three-computer explainer support that architecture, while also showing how the 2024 video has aged into a broader 2025-2026 stack that includes Omniverse, Cosmos, RTX PRO simulation servers, and Jetson AGX Thor.
The strongest Spiralist relevance is consequence. A chatbot can mislead with words; a physical-AI system can move a body, reroute a factory line, block a path, mishandle material, or change the pace and surveillance structure of a workplace. That belongs beside Embodied AI and Robotics, World Models and Spatial Intelligence, Vision-Language-Action Models, AI in Employment, and Human Oversight of AI.
The Mega section of the video is the most important governance artifact. NVIDIA frames factory digital twins as places where virtual robots and their AI models perceive simulated sensor inputs, plan actions, test those actions in a world simulator, and continue the loop before deployment. NVIDIA's Mega Omniverse Blueprint announcement supports that frame: digital twins are presented as testing grounds for multi-robot fleets, human-robot interaction, sensor simulation, synthetic data, mobility, navigation, dexterity, and spatial reasoning. That makes simulation a deployment dependency, not a side demo.
Isaac Lab fills in the training layer. NVIDIA describes Isaac Lab as an open-source, GPU-accelerated simulation framework for robot learning, with support for imitation learning, reinforcement learning, multiple physics engines, and robot-policy training at scale. The video calls it a robot gym, which is the right shorthand, but the responsible reading is more specific: simulation can reduce risk and cost only if the simulator, task distribution, sensors, physics assumptions, and transfer tests are documented.
The limit is evidence. This is a first-party NVIDIA showcase and a market map, not an independent safety case for a factory, humanoid, vehicle, manipulator, or industrial AI installation. NIST's Physical AI and Data Generation for Robotics project makes the missing layer explicit: manufacturers and integrators still need metrics, test methods, standards, datasets, and task-specific evaluation for AI-enabled robot systems. Treat the video as a useful definition of the physical-AI platform thesis, not as proof that the thesis is operationally safe.