Why Jensen Huang Says Future AI Builders Need Physics

Jensen Huang’s advice to favor physical sciences has become more concrete as NVIDIA expands its work on models, simulators and computing systems for physical AI. What remains unchanged is the underlying point: machines that act in the real world must obey physical constraints, however capable their software becomes.

This is not a reason for beginners to abandon computer science, and it does not mean autonomous robots are a solved problem. The practical conclusion is narrower: anyone interested in robotics, autonomous vehicles or intelligent industrial systems benefits from combining computation with mechanics, control and an understanding of uncertainty.

What Huang actually recommended

In a CNBC video published July 22, 2025, Huang answered a question about what he would study if he were graduating again by saying that his younger self would probably choose more of the physical sciences than the software sciences.

His answer did not amount to a universal instruction to avoid programming or pursue one particular degree. It pointed instead to the scientific knowledge AI systems need when their outputs stop at neither text nor images, but become movements with real consequences.

A robot cannot succeed by producing a merely plausible description of an action. It must estimate motion, apply suitable force, maintain balance and adjust when contact with an object differs from its prediction. A weak answer from a brainstorming tool may be inconvenient; an incorrect movement by a machine can damage equipment or injure someone.

Why physical AI needs more than pattern recognition

A physical system operates under inertia, friction, gravity, delay, uncertainty and limited energy. Its information is also incomplete: cameras can be obstructed, surfaces can slip, objects can deform and sensors can produce noisy measurements. These conditions shape the task itself rather than appearing as minor complications after model training.

An Associated Press account of the March 18, 2025 keynote records Huang framing robotics as AI’s next wave and connecting physical AI with friction, inertia, cause and effect, and object permanence. The same account describes NVIDIA’s emphasis on synthetic training data, simulation and the Newton physics-engine project.

Simulation matters because gathering every relevant experience with real machines is slow, expensive and sometimes unsafe. A simulated robot can repeat a task, encounter controlled variations and generate training experience without damaging a physical prototype after every failure. The difficult question is whether the virtual dynamics represent reality closely enough for learned behavior to transfer.

Physics changes how that transfer problem is understood. Rather than treating a robot as an AI model with motors attached, a developer must consider coordinate systems, forces, torque, momentum, collisions, actuator limits and measurement error as interacting parts of one system.

What changed after Huang’s remark

The physical-AI thesis has developed into a broader collection of tools for world modeling, simulation and robot learning. NVIDIA’s March 16, 2026 release introduced Cosmos 3, placed Isaac Lab 3.0 and GR00T N1.7 in early access, and presented GR00T N2 as planned for later in the year.

Those availability labels are important. They show sustained development, but they do not establish that every component is generally available or that general-purpose robots can already work reliably in every unstructured environment. Partner integrations likewise demonstrate active deployment work, not universal autonomy.

The update strengthens the rationale behind Huang’s study advice without proving every part of his forecast. Physics is already embedded in the tools used to simulate movement, evaluate robot policies and connect learned behavior to hardware. At the same time, early-access and planned components show that the supporting technology is still developing.

Which physics topics matter first

A beginner does not need to master every branch of theoretical physics before working with a robot. The most useful starting subjects connect observable motion with mathematical models and controllable behavior.

Classical mechanics: Position, velocity, acceleration, force, energy and momentum provide the basic language for predicting movement.
Rotational motion: Torque, angular velocity and moments of inertia matter when working with joints, wheels, arms and balancing machines.
Contact and friction: Grasping, walking and moving objects depend on interactions between surfaces, not geometry alone.
Electricity and signals: Motors, sensors, power limits and noisy measurements connect software decisions to physical hardware.
Probability and estimation: Because sensors provide uncertain observations, a machine must combine evidence instead of assuming perfect measurements.

Mathematics connects these subjects to implementation. Linear algebra represents transformations and robot poses, calculus describes change, differential equations model evolving systems, and probability handles uncertainty. The goal is not simply to collect formulas, but to understand the assumptions behind a model and recognize when they fail.

How to combine physics with computation

The most useful learning sequence joins physical reasoning and programming from the beginning. Mechanics can feel detached when nothing is modeled, while robot code written without dynamics often produces trial-and-error fixes that fail under new conditions.

Study vectors, introductory calculus and basic mechanics together, using numerical exercises to predict motion.
Build simple simulations of a falling object, a pendulum or a wheeled vehicle, then compare their behavior with the governing equations.
Add feedback control by measuring the difference between a desired state and the simulated state, adjusting an input and observing stability or overshoot.
Introduce imperfect measurements, delays and changing friction so the system must handle uncertainty rather than an idealized world.
Connect perception or machine learning only after the physical task and its measurable failure conditions are clear.

A small simulated balancing task can reveal the central issue. Success depends on estimating the object’s state, applying a correction at the right moment and respecting force limits—not simply recognizing what the object is. A failure can then be traced to a physical assumption, a measurement error, a controller or a learned policy.

Physics and software are complementary

Huang’s recommendation is most useful as an argument for deeper knowledge of the problem domain, not as a contest between physics and computer science. Physical AI still depends on algorithms, data structures, numerical methods, machine learning and reliable systems engineering. Physical constraints determine whether the resulting action can work outside a digital environment.

For a newcomer, the durable combination is therefore physics plus computation. Code provides the means to model, simulate and deploy a system; physics provides the means to judge whether its behavior is feasible. NVIDIA’s robotics push makes that combination more relevant, while the field’s unresolved reliability problems make strong foundations more valuable rather than less.

Jensen Huang Said Learn Physics. NVIDIA’s Robot Push Shows Why