"Subscribing to an API can never remedy mediocre management!" Jensen Huang stated bluntly, AGI is just like a genius newly graduated PhD, and it is sheer fantasy to sit back with your legs crossed waiting for efficiency to double.

At the just-concluded G20 Innovation Ministers' Meeting, Jensen Huang, CEO of NVIDIA, was invited to deliver a fireside chat. The audience sitting below were not ordinary tech enthusiasts, but industrial decision-makers and officials from economies around the world. Facing this group of people, Jensen Huang did not promote chips, but went straight to the point: When it comes to the computing power business, how exactly do we calculate the economic accounts for a country and an enterprise?

At the beginning of the conversation, Jensen Huang broke through a common inertial perception: many people are used to regarding chips and computers as consumer hardware such as mobile phones and PCs, but today they have long evolved into industrial infrastructure. The essence of a modern data center is actually an "oil refinery" — it processes and purifies the input electric energy into high-value intelligence that can be monetized by the whole society.

He once again put forward the classic "five-layer cake" theory, explaining clearly how ordinary countries can avoid the edge of the United States and cut into local advantageous industries; he also brought the seemingly mysterious AGI (Artificial General Intelligence) back to reality — he made an analogy that AGI is just like a newly recruited Stanford PhD at NVIDIA, who is extremely intelligent, but you can never expect to subscribe to an API interface and wait for the company's efficiency to double. How to present business scenarios and align goals for these "digital geniuses" is the real core competence of the management.

Key Points at a Glance

• "Token" is the "kilowatt-hour" of the AI era: "A hundred years ago, the world sold energy at 'several dollars per kilowatt-hour'; today, the measurement standard has become 'several dollars per million Tokens'. Although intelligence is intangible, just like electricity, it is an infrastructure that can be monetized with real money."
• The real inflection point to break the technical threshold: "Over the past 50 years, computers have only been the privilege of 20 to 30 million full-time engineers around the world. For the first time in human history today, anyone can command a computer as long as they use their native language."
• When an Agent has a physical body, it is a robot: "Put an exoskeleton of calling tools on the large model, and it becomes a digital Agent; once you give it a physical entity, install four wheels on it and it realizes autonomous driving, and embed it into a robotic arm and it enables advanced manufacturing."
• Subscribing to an interface cannot save mediocre management: "The realization of AGI does not mean that all enterprise problems are solved with one click. It is just like when the company recruits a group of top PhDs from prestigious schools, you cannot sit back and enjoy the benefits with your legs crossed. What truly determines the upper limit of output is whether managers can build a good business framework and power and responsibility boundaries for them."
• The biggest risk is being left behind by the times: "Safety is always achieved through technological progress, not by stepping on the brakes to preserve it. The worst outcome in this transformation is that you immerse yourself in fear, fail to enter the game, and are eventually completely left behind by the times."

The full text of the interview is as follows:

Computing power is not a consumer product, it is a "new infrastructure" that refines energy into intelligence

(Editor's Note: When the conversation started, the scene was discussing the underlying logic of computing power centers and AI infrastructure, and Jensen Huang went straight to the point:)

Jensen Huang: It converts energy into high-value, high-output intelligence, and we are now fully commercializing and monetizing this intelligence across the United States.

We have reached a historical inflection point in artificial intelligence investment, which has now become an infrastructure that can truly bring productivity.

In the past, we always thought that chips, technology and computers were just mobile phones and PCs. Today, we must examine them the same way we view energy, the Internet and infrastructure — for policy makers, the entire industry and even every enterprise, this change of thinking and cognition has a fundamental reshaping significance.

Host: Today we hope to take this opportunity to help our G20 partners understand our and your views on future growth. In your opinion, can computing power directly create revenue and economic value for countries? Does the world really have to build large-scale data centers to capture this growth? Can you talk about how computing power is converted into economic benefits?

Jensen Huang: No problem, there is a grand and wonderful concept behind this, which is called Token.

Token is the mathematical product of massive computing, and it represents intelligence itself.

When you aggregate and reorganize these Tokens and these numbers, they can generate an image, write a paragraph, answer questions, solve real-world problems, and even inspire brand new creativity. These Tokens are produced by the computers we manufacture through pure mathematical calculations.

This is so similar to history. Go back a hundred or two hundred years ago, the world discovered a new thing called "energy", and then people monetized it in the way of "several dollars per kilowatt-hour". And today, the measurement standard has become "several dollars per million Tokens".

The logic of the two is exactly the same. It seems intangible and elusive to ordinary people, but it contains inestimable value.

This is just like when we realize that "intelligence is essentially valuable". How do you embrace intelligence? How do you measure it? When it appears in front of you and helps you solve practical problems, you can immediately feel its existence. But it is intangible, as ethereal as mist. Intelligence is not made of atoms, it is a conceptual existence. That is why many people still find artificial intelligence hard to understand.

However, once you really start using these AI applications, you will instantly realize its extraordinary value. It makes you more productive, multiplies the capabilities of ordinary workers, and gives engineers an extra edge.

For countries with a huge young population, it will greatly accelerate the popularization of education; for countries committed to developing scientific and technological strength, it can help you achieve leapfrog upgrading in a short period of time.

Just as energy production popularized electricity to every country, artificial intelligence will also popularize intelligence and top capabilities to all G20 countries. For countries that dare to invest in it, the speed at which it promotes social progress and industrial upgrading will far exceed everyone's previous imagination.

From many perspectives, it is the ultimate equalizer.

Looking back at the 50-year history of computer development, computers were essentially exclusive tools for 10 to 20 million people around the world. The only way to take advantage of this computing power was to master computer programming and design. That is also my lifelong career: designing computers and writing programs. But this ability has always been limited to 20 to 30 million people around the world at most.

And today, for the first time in human history, anyone in the world can program a computer.

You only need to use your own human language to command it. We have completely popularized this computing technology, which is crucial to society, into the hands of everyone, enabling all people to achieve a qualitative leap in their human capabilities.

This is also why I fully call on everyone to embrace artificial intelligence. It can empower you, support you, and fulfill you. It is the ultimate equalizer.

Disassemble AI's "five-layer cake": You don't need to be perfect in every aspect, but you must be fully involved

Host: What advice do you have for the global partners present? How should countries seize this historical opportunity in the future? What must they build within their own borders? What key investments should they make so that when they go back to their countries to plan and participate in this AI revolution, they have a clear optimal implementation path in mind?

Jensen Huang: The top priority is to see the essence of AI. I have been using the word "infrastructure", but if you break it down completely, AI is nothing more than a "five-layer cake".

The bottom layer is energy. Without energy, you can't produce anything. This layer is responsible for converting electricity into mathematics. So, the first layer is energy.

The second layer is chips. This is also the field I have been deeply involved in.

The third layer is infrastructure, that is, large data centers: including land, power supply, and the shell and plant that accommodate all computing technologies. When energy is input into it, value gushes out.

Going up to the fourth layer, it is where the AI model is located. Most people mistakenly think that the model is the whole of AI, but in fact, AI is a complete five-layer cake.

Nowadays, I have carried out cooperation in many countries of the representatives present, and I have come into contact with start-ups, mature large companies and scientific research institutions. All regions around the world have their own strengths in certain key areas of artificial intelligence, because the connotation of AI is far more than just language.

Any structured information can be learned by artificial intelligence to understand and predict. Biology has its structure, chemical molecules have their structure, and physical laws also have their structure. Different regions around the world have their own specialties, and AI spans all these fields: different languages, different technologies, different information theories.

The fifth layer, which is also a crucial part of AI, is data and applications.

The United States has invested the most deeply in inventing, building and developing the entire five-layer cake. However, every region and every country should clearly decide which layer to concentrate their investment on. You don't have to take the lead in every layer, nor do you need to develop all layers comprehensively.

The most critical thing is that every country must actively advocate the widespread popularization of AI and promote the deep integration of AI into thousands of industries in its own country. The industries mentioned here cover almost all fields from education, medical care, manufacturing to basic science.

The United States will maintain its leading position in building this five-layer technology stack, but we are eager to cooperate with everyone so that all countries in the world can share the dividends.

The technology platform built by NVIDIA can support the operation of every model in all scientific fields and all modalities. It can run models from the United States, as well as models from all countries around the world, supporting biological models, chemical models, physical models and robotics technologies.

Put an exoskeleton on the digital brain: When an Agent has a physical body, it is a robot

Host: Whenever I go to a dinner party and talk about "AI", everyone takes it for granted that it is just software. But what we are seeing now is edge AI, hardware, robotics and advanced manufacturing. In your opinion, what is the relationship between physical physical infrastructure such as factories and the stereotyped cognition of the public that "AI is just software"?

Jensen Huang: AI is essentially intelligence, and its starting point is indeed software.

Looking back at the AI process over the past year, we initially talked about Large Language Models (LLMs). The whole world now knows what "LLM" means, which is a remarkable thing in itself.

But what really makes AI generate practical value is putting an exoskeleton on the LLM. This exoskeleton is what we call the agent harness. It fundamentally endows this "brain" with the complete components necessary to retrieve knowledge, retain working memory, call tools, collaborate and solve practical problems.

The pure digital form of AI is called an agent. If you no longer let this digital agent only operate digital tools such as spreadsheets, PPTs or browsers, but guide it to manipulate mechanical equipment —

Once this agent is given a physical entity (embodiment), it instantly turns into a robot.

Put this agent into a machine with four wheels, and you have an autonomous driving car; put this agent into a robotic arm, and it becomes an industrial robotic arm that can perform high-precision sorting, assembly or advanced manufacturing tasks; implant this agent into a logistics vehicle, surgical robot or automated drug R&D laboratory, and it can perform professional tasks.

The underlying principles of all these forms are completely the same: the large language model is wrapped in an agentic system, commanding and manipulating digital tools or physical machinery.

In the final analysis, all countries must clearly realize that this is an indispensable infrastructure, just like water conservancy, roads, power grids and the Internet. You must build your own digital infrastructure to support the local economy. It is the digital intelligence reserve exclusive to your researchers, students, the public and local start-ups.

Once we assist in building infrastructure locally, local start-up teams and scientific researchers can be completely activated immediately. This is an excellent investment with amazing returns.

In the United States, thanks to policies that encourage energy growth, execution speed and regulatory environment, we have built one of the largest computing power infrastructures in the world. As a result, AI is creating jobs in all fields: wafer manufacturing plants, computer assembly plants and AI data centers are no exception.

This year alone, the total investment in AI infrastructure across the United States will approach one trillion US dollars, directly driving huge economic growth and creating hundreds of thousands of jobs.

The biggest risk is never out-of-control technology, but that you are completely left behind by the times

Host: Any major era transformation will cause people's anxiety, and concerns from all sides come in droves. If we compare this AI revolution with all previous major transformations in human history (all the way back to the Industrial Revolution), how do you define it? What historical coordinate do you place it on, and how should leaders from all walks of life embrace it in the best posture?

Jensen Huang: Just like when all new technologies first came out, it initially carried a layer of magical color.

In the early days of the popularization of electricity, people pressed a switch far away, and the room was instantly illuminated, which seemed incredible and like magic at that time. The beginning of every disruptive technology is like magic. But for us builders in it, the truth is: this is engineering science. It is built step by step by computers, software, mathematics and probability theory.

Just like the aviation or automotive industry, ensuring the safe and responsible operation of technology is the basic bounden duty of technology builders. And the only way to make technology safer is to continuously promote the self-iteration and progress of technology.

I would obviously prefer to drive a modern car equipped with all modern active safety features than a vintage car from a hundred years ago; I would also prefer to take today's civil aviation airliner than a plane from a hundred years ago. We must accelerate the evolution of technology to make its functions more and more perfect, and more and more able to fulfill its original promise, and this capability itself will make it safer in use.

The discussion around technology security and protection is extremely necessary and precious. However, from the perspective of sovereign countries, what is the worst possible outcome in this new industrial revolution?

The worst outcome is that you don't participate in it at all, thus being completely left behind by the times.

This tragedy will happen if leaders only immerse themselves in fear and uncertainty when talking about technology, and focus all their attention on potential dangers rather than enabling potential. Discussions about risks must be balanced with conversations about prosperity and empowerment. This is not an either-or single-choice question, and we must maintain a good balance between the two.

Let's review the course of the aviation industry. Once upon a time, airlines spent huge sums of money on advertising claiming: "My plane is safer than the competitors". But it turns out that passengers don't want to hear that at all. From the perspective of consumers, ensuring flight safety is the unshirkable bottom-line responsibility of the entire industry.

What passengers really want to hear is that flights can take them to brand new destinations, broaden their horizons, bring the world closer, let them experience diverse cultures and open up new economic opportunities. They want safety to be our job, not a burden that worries them all the time.

The same is true in the automotive, power generation, and aviation fields, and it is also true for AI today. Ensuring safety compliance and solid and reliable engineering design is the common responsibility of us as industry leaders and builders.

At the policy level, we need precise and pragmatic regulation. Regulation should target clear and measurable actual hazards in reality, rather than unfounded fears based on speculation and assumptions. This technology is still in the early stage of the innovation S-curve, and it must be given sufficient space to develop rapidly forward.

Three years ago, if we had frozen or slowed down the R&D pace due to excessive concerns, we would not be able to solve the problem of model "hallucinations" today. It is the continuous refinement of technology that makes these models based on verifiable objective