Nvidia is turning its gaze from the digital to the tangible. At its annual GTC conference in San Jose, Jensen Huang laid out a vision that moves beyond large language models and data center dominance. The next wave, he explained, will involve machines that can interact with the physical world. This is the era of physical AI, a term the company is pushing to describe robots and autonomous systems that learn by doing.
Robots that learn from human examples
The company revealed a new suite of tools designed to accelerate robotics development. Central to this effort is Nvidia Isaac, a platform that now includes simulation and training capabilities. Developers can use it to train robots in virtual environments before deploying them in factories or warehouses. Huang emphasized that imitation learning, where robots copy human actions from recorded data, will be a key method for teaching these systems. The goal is to reduce the time and cost of programming complex movements from scratch.
Nvidia also announced partnerships with several robotics firms. These collaborations aim to build general-purpose robots that can handle multiple tasks, rather than single-purpose machines. Huang predicted that the market for robotics will eventually surpass the current size of the PC and smartphone industries. He pointed to logistics, manufacturing, and healthcare as early adoption areas.
The hardware fueling the shift
Underpinning this push is a new chip architecture called Blackwell Ultra. Huang described it as a step up from the current Hopper generation, with improvements in memory bandwidth and processing speed for AI inference. These chips are designed to handle the complex simulations that physical AI requires. Nvidia also introduced a compact computer called Jetson Thor, aimed at powering robotic systems directly on edge devices.
Huang noted that the demand for computing power in this sector is growing faster than expected. He cited the example of a single autonomous vehicle generating terabytes of sensor data per hour. Processing that data in real time, he said, requires a shift in how chips are designed and deployed. The company is positioning its hardware as the backbone for this new class of applications.
The announcement also touched on the convergence of generative AI and physical AI. Huang argued that language models alone cannot control a robot arm or navigate a drone. The next frontier, he said, is combining large language understanding with spatial reasoning and motor control. Nvidia is building software tools that bridge this gap, allowing developers to use natural language prompts to program robot behaviors.
Some analysts at the conference expressed caution. They noted that physical AI faces challenges in safety, regulation, and energy consumption. A robot operating in an uncontrolled environment, one analyst pointed out, has very different failure modes than a chatbot. Nvidia acknowledged these concerns and said it is working with safety researchers to develop standards for testing and deployment.
Huang closed his keynote by framing physical AI as a long term bet. He compared it to the early days of the internet when infrastructure had to be built before applications could flourish. Nvidia, he said, is investing billions in the tools that will enable this future. The company is banking on a world where chips and code move not just bits but objects.
For more on how Nvidia and its partners are shaping the AI hardware landscape, read our analysis of the latest trends in machine learning infrastructure at Mylistingo.
