"The number of NVIDIA's AI projects has gotten out of control," said Jensen Huang after downing five glasses of wine, revealing all his long - held secrets.

Recently, Jensen Huang, the founder, president, and CEO of NVIDIA, had an exclusive fireside chat with Chuck Robbins, the chairman and CEO of Cisco. Both of them were very relaxed. After having five glasses of wine, Jensen Huang rarely and unreservedly looked into the future of intelligence and infrastructure, as well as the global transformation that is redefining every industry on Earth.

During the chat, Huang sharply pointed out that programming is just typing, and typing itself is a commodity.

He revealed that the number of AI projects within NVIDIA is almost "out of control", but he still lets them develop and hasn't reined them in prematurely. The flourishing of various projects within the company is a necessary stage for innovation. His first reaction to new AI projects is "yes", rather than "prove it to me first".

He believes that the real opportunity of AI lies not in "smarter software" but in "enhanced labor". For the first time in history, the long - term economic value of digital labor exceeds that of hardware itself. The most valuable intellectual property of an enterprise is not the answers but the questions themselves, and these must be kept locally. In the future, every employee will work "with multiple AIs".

Huang reminded that you don't have to be the first to use AI, but never be the last. Being the last basically means being eliminated. And the real priority for using AI should be the "core and most influential work" of the enterprise, not trivial matters.

Below is the content of their conversation. We have translated and organized it, making some deletions without changing the original meaning. There were a few joking moments between them, which are presented in the form of "mini - dramas" for readers to enjoy.

Mini - drama 1

Jensen Huang: I feel like I'm working for someone else.

Chuck Robbins: When the wine was brought over just now, Jensen reminded me and asked if I knew it was a live broadcast. I said, "Forget it. It's so late anyway."

Jensen Huang: The first principle is not to cause harm.

Chuck Robbins: Don't harm anyone and be aware of how lucky you are.

First of all, thank you all for staying here for so long. We started very early this morning, had one speech after another, then took a break for about two and a half hours, and now you're all back to see us.

Jensen Huang: So I got up at one o'clock in the morning.

Chuck Robbins: He just finished a two - week trip, visiting four or five cities in Asia. He was in Taiwan, China one day and was in Houston last night.

Jensen Huang: And now I'm here.

Jensen Huang: We Need to Reshape Computing

Chuck Robbins: This guy hasn't been home for two weeks. Now the question is, can he sleep in his own bed or will he have to stay in a hotel? So we'll keep it light, have a good chat, and try to let him go as early as possible.

Actually, you don't need to introduce yourself, but still, thank you for coming today. Thanks for our partnership, and I'm proud of you and your team. Let's start with our cooperation. You came up with the concept of the AI factory, and we're working on it together. In the enterprise field, the progress might not be as fast as we hoped, but let's first talk about what an AI factory means to you?

Jensen Huang: First of all, remember that we're undergoing the first reshaping of computing in 60 years. Previously, it was explicit programming. We wrote programs and passed variables through APIs, and everything was very clear. Now it's implicit programming. You tell the computer what you want, and it figures out how to solve the problem on its own.

From explicit to implicit, from general - purpose computing (basically arithmetic) to artificial intelligence, the entire computing stack has been reshaped. When people talk about computing, they often only focus on the processing layer, which is our area, but computing also includes storage, networking, and security, and all of these are being reshaped.

First, we need to develop AI to a level where it's useful to people. Currently, so - called chatbots, when you give them a prompt, they come up with an answer. It's interesting and makes people curious, but it's not very practical. Sometimes it helps me with crossword puzzles, but only within what it has memorized and generalized.

Recall three years ago when ChatGPT first emerged. We thought, "My goodness, it can generate so many words and imitate Shakespeare." But in essence, it's still just memorizing and generalizing existing content. However, we know that real intelligence is about solving problems. Solving problems means, on one hand, knowing what you don't know, and on the other hand, reasoning, that is, how to break down a problem you've never seen before into parts that you can easily solve. Then, through combination, you can solve problems you've never encountered.

In addition, you also need to be able to come up with strategies to execute tasks, which we call planning. The terms we hear now, like Agentic AI, tool invocation, retrieval, fact - based enhanced generation, memory, etc., are essentially about these abilities.

But importantly, to evolve from general - purpose computing, which is explicit programming written in Fortran, C, C++, Cobalt, to the new form, we need to rethink how the entire enterprise can utilize it.

Chuck Robbins: Cobalt

Jensen Huang: That's good stuff, Chuck. That's good stuff.

Chuck Robbins: That's my backup job.

Jensen Huang: Yes, those are still valuable skills. I know, I know they're still valuable. I've received many job offers. Dinosaur skills are always valuable.

Chuck Robbins: We just confirmed that you're older than me.

Jensen Huang: I know. I'm prehistoric. I might not look like it, but I am. Well, I might be the oldest one in this room.

Chuck Robbins: So let's talk about it. When you think about this topic...

Jensen Huang: So, I found Chuck and said, "Listen, we need to reshape computing, and Cisco has to play an important role." We have a brand - new computing stack, Vera Rubin, and Cisco will promote it with us. But that's the computing layer, and there's also the networking layer. Cisco will integrate our AI networking technology and put it into the control layer of Cisco Nexus. In this way, from your perspective, you can get all the performance of AI while having Cisco's controllability, security, and manageability. We'll do the same thing in terms of security. Every pillar needs to be reshaped so that enterprise computing can fully utilize it.

But ultimately, we still come back to the question: Why wasn't enterprise AI ready three years ago? Why do you have to get involved as soon as possible now? Don't fall behind. I think you don't have to be the first company to adopt AI, but never be the last.

NVIDIA's "AI Project Count Is Out of Control"

Chuck Robbins: If you're an enterprise now, what's your advice? What steps should they take to prepare?

Jensen Huang: I'm often asked questions about ROI, but I don't recommend starting from there. The reason is that in the early stage of deploying any technology, it's difficult to quantify the return on investment of new tools and technologies using spreadsheets. What I recommend is to figure out what the essence of the company is and what the most influential work we do is. Don't mess around and don't waste time on peripheral matters.

In our company, AI projects are blooming like a thousand flowers. The number of AI projects in the company is out of control, but that's good. Innovation is not always controllable. If you want to control it, first, you need to see a psychiatrist. Second, control is just an illusion. You can't control the company. If you want the company to succeed, you can't control it. You can influence it, but not control it.

So I think, first of all, too many companies and too many people want a clear, specific, and provable ROI. But it's difficult to prove the value of something worth doing in the early stage.

I'd say, let a hundred flowers bloom. Let people try, and let them try safely. We try all kinds of things in the company. We use Anthropic, Codex, Gemini. We use everything. When our team says they're interested in an AI project, my first reaction is "yes", and then I ask why. I don't ask why first and then say "yes".

The reason is simple. I hope my company can explore life just like I hope my children to. When they say they want to try something, my answer is "yes", and then when they ask "why", I won't let them prove it to me. I won't let them prove that it will bring economic success or some kind of happiness in the future. But we often do this in the workplace. Don't you think it's strange? It's beyond my comprehension.

So our attitude towards AI is the same as our attitude towards the Internet and cloud computing in the past. Let it bloom. Then at some point, you need to use your judgment to decide when to start tidying up the garden because a thousand flowers will make the garden messy.

At some point, you have to start "pruning" to find the most suitable flowers, that is, the best method or the best platform, so that you can concentrate all your resources in one direction. But you don't want to concentrate your resources too early. What if you choose the wrong direction? So first, let a thousand flowers bloom, and then tidy up at some point.

To clarify, I haven't started tidying up yet. Flowers are blooming everywhere. But I encourage everyone to try. However, I clearly know what's most important for our company, what the essence of our company is, and what our most important work is. I make sure that we have a lot of expertise and capabilities focused on using AI to revolutionize these works. For us, it's chip design, software engineering, and system engineering.

You've also noticed that we're collaborating with companies like Synopsys, Cadence, and Siemens to embed our technology. Whatever they need and want to use, I'll give it to them to the fullest. Because only in this way can I completely revolutionize the tools we rely on to design our next - generation products.

This tells you something about my attitude, what's most important to me, and how I'll revolutionize my own work.

Think about what AI can do. AI reduces the cost of intelligence or creates an abundance of intelligence by an order of magnitude. In other words, what used to take us a year to complete, now might take only a day. What used to take a year might now take only an hour. It can even be done in real - time.

The reason is that we're in a world of abundance. Moore's Law? My goodness, it's so slow, like a snail. Remember, according to Moore's Law, things double every 18 months, increase ten - fold every five years, and a hundred - fold every ten years. But now? It increases a million - fold every ten years. In the past ten years, we've advanced AI so far that engineers are saying, "Hey, why not train an AI model on all the data in the world?"

They're not talking about collecting data from my hard drive, but pulling down all the data in the world to train a model. That's the definition of "abundance". Abundance means facing a problem so huge that you say, "I'm going to solve it all." "I'm going to cure all diseases in all fields. I won't just do cancer research." Is it a joke? It's crazy. We're going to solve all human sufferings. That's abundance.

Recently, when I think about a problem, I usually make this assumption: My technology, my tools, my instruments, and my spaceship are infinitely fast. How long does it take to get to New York? One second. So if I can get to New York in one second, what different things would I do? If what used to take a year can now be done in real - time, what different things would I do? If something that used to be heavy now becomes as light as anti - gravity, what different things would I do? When you approach everything with this attitude, you're using AI thinking.

For example, many of the companies we're collaborating with are doing graph analysis, dealing with various dependencies and relationships. These graphs have countless nodes and edges, up to trillions. In the past, the approach was to calculate in small chunks. Now? Just give me the whole graph. It doesn't matter how big it is. This way of thinking is being applied in various fields. If you're not thinking this way, you're basically doing it wrong.

Does speed matter? No, you've reached the speed of light. Does weight matter? You're in zero - gravity and have zero mass. If you still have a hard time with things that used to be extremely difficult for you, and your attitude is "it doesn't matter", then you're not applying this logic, and you're not doing it right.

Now, try to apply this way of thinking to the most difficult and crucial problems in your company. That's the real way to drive change. If you haven't thought this way yet, think about it: Is your competitor already thinking this way? Or, even scarier, a newly - established company is already thinking this way. It will change everything.

So, my advice is to find the most influential work in your company, apply the concepts of "infinity", "zero cost", and "speed of light", and then ask Chuck how to really make it happen.

"Most Valuable Things Are Called Intuition and Wisdom"

Chuck Robbins: Now let's talk about how to implement it. You have the metaphor of the "five - layer cake" because everyone is talking about infrastructure, models, and applications. How should I start? Can you talk about this?

Jensen Huang: First of all, one of the things successful people do is to think about the essence of things.

About 15 years ago, an algorithm and two engineers solved the computer vision problem. Intelligence includes perception, reasoning, and planning. Computer vision is basically the first part of intelligence: perception. Perception is "Who am I? What's going on? What's my environment?" Reasoning is "How do I compare this with my goal?" And then come up with a plan to achieve it. For example, in the case of a jet fighter, first, it's about perception and positioning, and then action.

Intelligence is about these three things. You can't have the second and third parts without perception. You can't understand or decide what to do without understanding the environment. And the environment is highly multimodal. Sometimes it's a PDF, sometimes a table, sometimes information, and sometimes even sensory, like smell, the environment. "Where are we? What are we doing? Who are we facing?" We often say "read the atmosphere" or "size up the situation", which is about perception.

About 13 or 14 years ago, we made a huge leap in computer vision, which is the first layer of the perception problem. It's extremely difficult. How to solve computer vision? AlexNet was the first breakthrough we saw. It's like the movie I like, "First Contact". This was our first contact with AI.

At that time, we were thinking: What does this mean? Why could two engineers, with just a few GPUs, surpass all the algorithms we'd accumulated over thirty years?

I was talking to Ilya Sutskever and Alex Krizhevsky about this problem yesterday. How did two young people do it? We completely dissected the problem. Ten years ago, I came to the conclusion that most of the truly difficult and valuable problems in the world can actually be solved in this way.

The reason is that there's no so - called "principled algorithm" for these problems. There's no F = ma, no Maxwell's equations, no Schrödinger's equation, no Ohm's law, and no laws of thermodynamics. They're not precise. Most valuable things are called intuition and wisdom.

The intuition and wisdom we talk about, and the problems we face every day, often have only one answer: It depends. If the answer is 3, that's great. If it's 3.14, that's even better. But in reality, most of the most valuable and difficult problems are almost all "it depends" because they're highly context - dependent.

It was about twelve or thirteen years ago that computer vision was conquered. We realized that with deep learning, this path could be continuously expanded, and the models could become larger and larger. The only problem was: How to train? And the real breakthrough came from self - supervised learning and unsupervised learning, letting AI learn on its own. Until today, we're almost no longer limited by manual annotation.

This breakthrough completely opened the floodgates, allowing models to expand from hundreds of parameters, hundreds of millions of parameters, all the way to billions and trillions of parameters. There was an