“The next wave requires us to know things like the laws of physics, friction, inertia“ The next wave requires us to understand things like the laws of physics, friction, inertia, and cause and effect. By this statement, Nvidia CEO Jensen Huang has tossed a curveball at the conventional wisdom that software proficiency is the key to AI success. In a year in which Nvidia has become the world’s most valuable company, Huang’s advice to 2025 graduates is direct: aim for the physical sciences if you want to ride a wave of the coming AI revolution.
Huang’s journey, from electrical engineering student to the architect of the GPU-powered AI era, mirrors the evolution of artificial intelligence itself. He traces AI’s ascent through four waves: Perception AI, Generative AI, Reasoning AI, and now, the emerging era of Physical AI. Each wave, he argues, has demanded a new set of technical skills and the next will belong to those who can bridge the digital and physical worlds.
The first seismic shift came in 2012, when AlexNet, a deep convolutional neural network, shattered records on the ImageNet Large Scale Visual Recognition Challenge. The creators of AlexNet, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, were able to take advantage of the power of GPUs to train an image model on millions of labeled images, achieving a top-5 error rate of 15.3%, nearly halving the best previously achieved. This moment of revolution, as told in the story of AlexNet, was the start of Perception AI computers that can “see” and interpret the world.
Huang credits this innovation with kicking off the modern era of deep learning.“Modern AI really came into consciousness about 12 to 14 years ago, when AlexNet came out and computer vision saw its big, giant breakthrough,” he told a Washington, D.C. audience. The technological achievement was driven by GPU acceleration, large-scale datasets like ImageNet, and new neural network architectures acombination of hardware and algorithmic innovation.
The second, Generative AI, saw models that could generate text, images, and code, while the third Reasoning AI dropped the bombshell agentic AI: digital work robots that were capable of reasoning, planning, and acting. These AI agents, as detailed in reports from the industry, are transforming manufacturing by individually optimizing production timetables, forecasting maintenance needs, and managing quality control. Whereas earlier automation was limited to adapting with occasional software upgrades, agentic AI adapts in real-time from massive data streams and collaborates with people as well as other AI systems.
But it’s the fourth wave Physical AI that Huang believes will define the next decade. Here, AI must learn not just about information, but the messy, unorganized nature of the physical world. That entails thinking about object permanence, force, friction, and causality skills rooted in physics and engineering. As Huang says, “It’s not just about data anymore it’s about understanding how data interacts with the real world.” Embodied in robots, these skills become the foundation forfactory automation and next-generation robotics, addressing global labor shortages and transforming whole industries.
To hasten this process, Nvidia has introduced platforms such as Omniverse and Cosmos. These technologies enable the creation of virtual twins of robots and factories so that simulations can be run, tested, and optimized in virtual space for physical AI before it is actually deployed. The process involves three basic steps: building photorealistic 3D worlds, annotating them with physical attributes, and generating massive amounts of synthetic data to train AI. It converges in a synthetic data multiplication system, doubling the scope and efficacy of AI models for robotics, autonomous driving, and industrial automation.
Huang looks beyond software. “Every robotics company will ultimately have to build three computers,” he told us at CES 2025: one to train AI , one to do edge inference, and one digital twin to bring together and iterate the system in a virtual world. Nvidia Cosmos, launched recently a world foundation model trained on 20 million hours of dynamic physical environments demonstrates the technical maturity of this new frontier.
To would-be engineers and technologists, Huang’s message is a call to add to their toolkit. It is as much a matter of learning the fundamentals of physics, mechanics, and materials science as it is learning to program. The future AI will be built by people who can create systems that not only calculate, but also sense, move, and adapt in the natural world a world that is governed by the laws of nature and also by the code itself.
