Jensen Huang's Ultimate Gamble: The Next Wave of AI, to Conquer the Physical World

In Silicon Valley's narrative, large models have already dominated the world of Bits, but Jensen Huang's sights have long been set on the world of Atoms.

"The next wave of artificial intelligence is Physical AI, which will sweep across the largest industries on the planet." In a recent public address, Jensen Huang, CEO of NVIDIA, defined the track for the second half of AI development. Even with a hint of frustration at the industry's underperformance, he pointed out the fatal critical flaw of current large models:

"Most of today's artificial intelligence does not understand the laws of physics, friction, inertia or causality. These are physical reasoning capabilities that even children and pets possess."

This statement is like a basin of cold water poured over the AI craze that is obsessed with "text generation", "code writing" and "multimodal dialogue". It reveals an extremely harsh reality: the large models we take pride in today are essentially more like a "Brain in a Vat" locked in servers. They have read all human libraries, yet cannot even stack a few building blocks steadily.

Jensen Huang's judgment is not only a prediction of technological trends, but also a declaration concerning the reshaping of a trillion-dollar commercial landscape.

The Dilemma of the Brain in a Vat: Why Doesn't AI Understand Friction?

To get AI out of the screen and into factories, roads and households, the biggest obstacle lies in the fact that the success of Large Language Models (LLMs) cannot be directly replicated in the physical world.

The underlying logic of large models is probability. It generates text by predicting the next word, and it does not need to understand universal gravitation, only needs to know that "apple" is often followed by "falling to the ground". In the virtual world of language and images, this "pattern matching" based on massive amounts of data is extremely effective.

But in the physical world, the fault tolerance rate is zero.

When a two-ton self-driving car travels on a wet and slippery road in rainy weather, or when a humanoid robot carries glass in a cluttered factory workshop, the variables they face are infinite and real-time. Friction changes with temperature, inertia will cause the robotic arm to tremble when it stops suddenly, and causality can collapse in an instant — this is not a problem that can be solved by "predicting the next word".

As Jensen Huang said, these physical reasoning capabilities are innate to children and pets. A cat jumping off a table can accurately calculate the landing point and adjust its body posture in a fraction of a second. But even the most advanced robot in the world often falls over completely when facing unknown terrain.

The lack of deep understanding of the real laws of physics is a fatal roadblock for AI to move into the real world.

Physical AI: The "Real Brain" of Embodied Intelligence

If large models endow AI with "generalized common sense", then Physical AI requires AI to have "intuition of space and time".

The Physical AI depicted by Jensen Huang is not just an AI installed in an iron shell, but a complete new architecture based on the World Model. It requires AI to be able to simulate real physical laws in the digital world, conduct low-cost trial and error and reinforcement learning, and then deploy this capability to the real world with low latency.

This exactly echoes the hottest track in the current venture capital circle — Embodied AI.

As we discussed in previous articles, a large number of Huawei-affiliated executives are now entering the embodied intelligence sector, and capital from Silicon Valley and Wudaokou is pouring frantically into the robotics track. But Jensen Huang's perspective is more grand: Physical AI is not just about building a few humanoid robots. It points to the world's largest industries — manufacturing, logistics, heavy industry, agriculture and healthcare.

Multiple institutions have issued trillion-dollar-level forecasts for the long-term economic value of Physical AI and related robotic applications, but the specific figures, time nodes and statistical calibers vary greatly: some focus on the related market around 2030, while others place the large-scale economic value of robots and Physical AI around 2040, and the geographical scope and whether workflow reconstruction is included are not unified. Therefore, a more prudent understanding is: this is a trillion-dollar directional judgment, rather than a unified measurement that can be accurate to a specific year or specific region. The scale of the software industry has an upper limit, and gradually transforming the world's physical infrastructure with AI does have huge commercial imagination space.

NVIDIA's Open Strategy: "Training" the Earth in the Virtual World

Faced with such a huge market cake, NVIDIA has already laid out its layout.

Since the cost of collecting data in the real world is extremely high (it is too costly to let robots break down millions of times), the perfect training ground for Physical AI must be in the virtual world.

This also explains why NVIDIA strongly promotes "Digital Twin" and the Omniverse platform. In the future, factories and cities will all have a "digital clone" in the virtual space, where Newton's laws and hydrodynamics are simulated. AI agents can conduct simulation training in this digital world at a speed far exceeding that of the real world. One night of simulation may be equivalent to a very long period of trial and error in the real world.

To fill the severe lack of real interaction data for Physical AI, NVIDIA launched Project GR00T — an open reference platform for humanoid robots, including foundation models, simulation frameworks and deployment toolchains. Combined with the Isaac robotics simulation platform, NVIDIA is trying to build an ecosystem from "virtual trial and error" to "real-world deployment".

However, there is a well-known industry challenge here: the "Sim-to-Real" gap. The virtual environment is after all a simplification of reality. Even the top physics engine cannot perfectly simulate every trace of light refraction in the real environment, or the complex deformation of a piece of fabric in reality. The unparalleled skills the model has mastered in the virtual world may be completely invalidated in the real world due to a tiny sensor noise. Overcoming this gap will be the lifeline that all Physical AI players must cross.

In this process, NVIDIA will not only sell computing power chips for training large models, but also sell edge computing chips for real-time physical reasoning. From the cloud to the end side, it is trying to capture every profit in the Physical AI era.

From Digital to Physical: The Testing Ground for Capital and Bubbles

When capital senses the certainty of "Physical AI", fanaticism follows.

As we mentioned when discussing the embodied intelligence track, Silicon Valley is striving to demonstrate hard-core technology, while some domestic startups are creating myths with PPTs, taking moving a bottle of mineral water as a breakthrough of the World Model.

The threshold of Physical AI is much higher than that of Large Language Models. It not only requires top AI algorithm talents, but also extremely profound capabilities in mechanical engineering, material science and supply chain management. This is a typical software-hardware integrated track, and any player who tries to use "Internet thinking" to achieve dimensionality reduction attacks will eventually be worn down by the physical laws of the manufacturing industry.

The real opportunity in China lies in the large-scale manufacturing and cost control of the hardware supply chain. When the "brain" of AI is gradually open-sourced or monopolized by giants, Chinese enterprises (local hardware companies represented by Unitree, and entrepreneurial teams with Huawei background such as Agibot) have advantages in the large-scale supply and cost control of joint motors, reducers, sensors and complete machine manufacturing; but it should also be noted that core components still rely heavily on the division of labor and collaboration of professional suppliers, and not all of them are independently developed and produced. This supply chain capability will become one of the indispensable foundations for the implementation of Physical AI worldwide.

Conclusion: The Singularity of Civilization Lies in AI Stepping Out of the Screen

In the past two years, we have been immersed in the intellectual carnival brought by large models. Text, code, images — AI is omnipotent in cyberspace.

But Jensen Huang's warning reminds us: the real industrial revolution is not completed by pixels on the screen, but by machines in factories and cars on roads. When AI truly understands friction, inertia and causality, and when it can interact with the real physical world like a child, human civilization will usher in the real singularity.

That singularity does not happen in servers, but on every atom around us. In this journey to the physical world, all old industries will be rewritten, and the winners will be those who not only know how to revere the laws of physics, but also dare to embrace massive computing power.

This article is from the WeChat official account "Tech Business", written by Tech Business, and published with authorization from 36Kr.