Physical AI

The term Physical AI was systematically articulated by NVIDIA CEO Jensen Huang in his keynote speech at CES 2024, referring to AI systems capable of understanding and acting upon the physical world.

In the AI safety field, the speed at which research results are translated into actual protective measures is clearly slower than the iteration speed of model capabilities.

Common friction points in AI feature acceptance include lack of consensus on evaluation criteria, high cost of constructing test data, and difficulty quantifying subjective quality.

AI Engineering differs from traditional machine learning engineering in that the former focuses on integrating existing foundation models into real products and workflows, while the latter emphasizes model training, hyperparameter tuning, and deployment.

Physical AI refers to the technological direction of taking AI beyond screens and into the real world to interact with the physical environment, with typical deployment scenarios including robots, automated equipment, and edge sensing systems.

Embodied reasoning refers to the capability of an AI model, given a physical body, to perceive, understand, and plan within its environment and execute actions.

Work that requires deep creativity and fundamental breakthroughs is harder to automate with AI, while standardized, highly repetitive application-layer engineering is more likely to be replaced by AI tools

Jensen Huang has repeatedly emphasized in recent years that robotics and Physical AI represent the next trillion-dollar market.

The partnership focuses on faster simulation speed, open benchmarks and datasets for physical AI, and evaluation systems for agentic AI in the CAE/EDA domains.

Physical AI emphasizes embedding physical constraints such as mass conservation, momentum equations and boundary conditions into machine learning models, and common approaches include physics-informed neural networks (PINN), operator learning (such as the Fourier Neural Operator), and using numerical solvers as differentiable layers.

The barrier to physical AI is shifting from "can it be done" to "how to assemble it quickly"; depth cameras, edge computing modules, integrated kits, and agentic frameworks each have relatively mature options.

There is still a gap between a demo and a product; real deployment requires consideration of exception handling, multi-person concurrency, long-term stability, and safety compliance, especially for products aimed at children.

Vishwanath believes physical AI lacks large-scale, high-quality robot interaction data; the success of large language models benefited from massive internet text data, whereas physical AI has no equivalent conditions.

The payback period for physical AI may be longer than for software, but once a breakthrough is made, the barriers are also higher.

Former Nvidia executive and now venture capitalist Amulya Vishwanath proposed on TechCrunch's Equity podcast that physical AI is in its "Lego era."

The supply chain for physical AI is far more complex than traditional software, involving mechatronics, materials science, and functional safety; these fields have longer innovation cycles but deeper moats.

Vishwanath's investment perspective leans toward the infrastructure layer, focusing more on teams that provide "Lego blocks" for physical AI, such as high-performance actuators, tactile sensors, real-time motion control software, and safety certification solutions.

Robotics is only a subset of physical AI, and equating physical AI with humanoid robots or robotic arms limits understanding of the technology wave and investment positioning.

The core difference between physical AI and software AI is that the system must complete closed-loop control in a noisy, not fully predictable physical environment, and is highly sensitive to real-time performance, safety and hardware cost.

The moats of physical AI startups usually come simultaneously from three levels: algorithms, data collection and the hardware supply chain.