The reporter said that communication with Jensen Huang should carry a warning label, because the CEO of Nvidia is so invested in the direction of artificial intelligence that after nearly 90 minutes of heated conversation, I am convinced that the future will be the Nirvana of neural networks. I can see it all: robot revivals, medical godsends, self-driving cars, chatbots with memories. The buildings on the company's Santa Clara campus didn't do anything. Everywhere my eyes fall, I see triangles within triangles, shapes that helped Nvidia make its first fortune.
Huang was the man of the year, maybe even the next decade. Because tech companies really love Nvidia's supercomputing GPU. This isn't Nvidia, it's the vendor of Gen X video game graphics cards that bring images to life by effectively rendering countless triangles. This is Nvidia, whose hardware created a world where we talk to computers, computers talk to us, and ultimately, depending on which technologist you talk to, they outdo us.
At our meeting, Huang, 61, wore his trademark leather jacket and simple black sneakers. He told me that Monday morning that he hated Monday mornings because he worked all day on Sunday and was already tired when he started the official work week.
Huang has developed a model that puts Nvidia ahead of every tech megatrend. In 2012, a small group of researchers released a groundbreaking image recognition system called AlexNet that uses GPUs (rather than CPUs) to process code, ushering in a new era of deep learning. Therefore, Huang immediately commanded the company to pursue artificial intelligence with all its strength. In 2017, when Google released a new neural network architecture called Transformer (T in ChatGPT) and sparked the current AI gold rush, Nvidia was in the perfect position to start selling its AI-focused GPUs to hungry tech companies.
Nvidia currently accounts for more than 70% of sales in the AI chip market and is valued at just over $2 trillion. Revenue for the last quarter of 2023 was $22 billion, up 265% from the previous year. Its share price rose 231% last year. Huang is either surprisingly good at what he does or extremely lucky-or both!- Everyone wants to know how he did it.
But no one can rule forever. He is now at the center of a Sino-American tech war and at the mercy of regulators. Some of Huang's challengers in artificial intelligence chips are household names-Google, Amazon, Meta and Microsoft-and have the deepest pockets in technology. At the end of December, semiconductor company AMD introduced a large processor for artificial intelligence computing, designed to compete with Nvidia. Start-ups are aiming for this goal as well. According to Pitchbook, venture capitalists invested more than $800 million in AI chips in the third quarter of last year alone.
How does Huang Jen-hsun view this?
Let's look at the original interview:
Huang Renxun:You and I are both Stanford graduates.
Lauren Goode: Yes. Well, I majored in journalism, and you didn't.
Huang Renxun:I wish I had.
Lauren Goode: Why?
Huang:Well, as a leader and as an individual, the person I really admire is Adobe CEO Shantanu Narayen. He said he always wanted to be a journalist because he loved telling stories.
Lauren Goode: Being able to effectively tell the story of a business seems to be an important part of building a business.
Huang Renxun:Yes. Strategy is storytelling. Culture is about telling stories.
Lauren Goode: You've said many times that you don't have the idea of marketing Nvidia based on promotional material.
Huang Renxun:That's right. It's actually for storytelling.
Lauren Goode: So I want to start with what another tech executive told me. He noted that Nvidia is a year older than Amazon, but in many ways Nvidia has a more "day one" approach than Amazon. How do you maintain this outlook?
Huang Renxun:Frankly speaking, this is really a good word. I wake up every morning like the first day because we're always doing something we've never done before. It also has a vulnerable side. We'll probably lose. Just now, I was in a meeting and we were doing something completely new to our company, but we didn't know how to do it properly.
Lauren Goode: What's new?
Huang Renxun:We are building a new type of data center. We call it the AI factory. The way data centers are built today, many people share a group of computers and keep their files in this large data center. The AI factory is more like a generator. This is quite unique. We've been building it for the past few years, but now we have to turn it into a product.
Lauren Goode: What are you going to call it?
Hwang:We haven't named it yet. But it will be everywhere. Cloud service providers will build them and we will build them. Every biotech company has it. Every retail company, every logistics company. Every car company of the future will have a factory that makes cars (physical goods, atoms) and a factory that makes artificial intelligence (electronics) for cars. In fact, as we speak, you see Elon Musk doing just that. He was far ahead of most people in thinking about what the future of industrial companies would look like.
Lauren Goode: You've said before that you run a flat organization with 30 to 40 executives reporting directly to you because you want to be part of the information flow. What has piqued your interest lately and made you think,"I might end up betting on Nvidia on this?"
Huang:Information doesn't have to flow from the top of the organization to the bottom like it did in the Neanderthal era, when we didn't have email and text messaging and all that stuff. Information flows much more rapidly these days. Therefore, there is no need for a hierarchical tree that interprets information from top to bottom. Flat networks allow us to adapt faster, which is what we need because our technology is evolving so fast.
If you look at the way Nvidia technology has evolved, Moore's Law doubles every few years. Well, in the last 10 years, we've advanced AI about a million times. This is many, many times Moore's Law. If you live in an exponential world, you don't want information to travel from one level to the next at a time.
Lauren Goode: But I ask you, what is your Roman Empire? This is a meme. What version of transformer paper is it today? What's going on right now that you think will change everything?
JH:There are a few things. One of them doesn't really have a name, but it's some of the work we've done in basic robotics. If you can generate text, if you can generate images, can you also generate motion? The answer may be yes. Then, if you can generate actions, you can understand intent and generate generic versions of clarity. Humanoid robotics, therefore, should be around the corner.
I think work around state-space models (SSM) could be the next transformer, which allows you to learn extremely long patterns and sequences without having to grow quadratically in computation.
Lauren Goode: What does this lead to? What are real-life examples?
Huang:You can have a conversation with a computer that lasts a long time, but the context is never forgotten. You can even change the theme temporarily and go back to the previous theme, and you can keep that context. You may be able to understand very long chain sequences, such as the human genome. Just look at the genetic code and you can see what it means.
Lauren Goode: How close are we to that goal?
Huang Renxun:From AlexNet to Superman AlexNet, it took only about five years. Basic models of robotics may be coming soon--I'll announce them sometime next year. Five years from now, you'll see some pretty amazing things.
Lauren Goode: Which industry would benefit most from a widely trained robot behavior model?
Huang Renxun:Well, heavy industry represents the largest industry in the world. Moving electrons is not easy, but moving atoms is extremely difficult. Transportation, logistics, moving heavy objects from one place to another, discovering the next drug-all of this requires understanding atoms, molecules and proteins. These are huge and incredible industries that AI has yet to impact.
Lauren Goode: You mentioned Moore's Law. Does this law no longer matter?
Huang Renxun:Moore's Law is now more of a system problem than a chip problem. More about the interconnectivity of multiple chips. About 10, 15 years ago, we started the journey of breaking down computers so that you could connect multiple chips together.
Lauren Goode: That's why you bought Israeli company Mellanox in 2019. Nvidia said at the time that modern computing places huge demands on data centers, and Mellanox's networking technology will make accelerated computing more efficient.
Huang Renxun:Yes, absolutely correct. We bought Mellanox so we could scale our chips and turn the entire data center into a superchip, enabling modern AI supercomputers. It's really to recognize that Moore's Law is over, and if we want to continue scaling compute, we have to do it at the data center scale. We looked at how Moore's Law was formulated and said,"Don't be constrained by it." Moore's Law is not a limitation on computation." We must discard Moore's Law so that we can consider new ways of extending it.
Lauren Goode: Mellanox is now considered a very smart acquisition by Nvidia. Two years ago, you tried to acquire Arm, one of the world's most important chip IP companies, but were blocked by regulators.
Huang Renxun:That's great!（That would’ve been wonderful!）
Lauren Goode: I'm not sure if the U.S. government agrees, but yes, let's be sure of that. What specific areas would you focus on when considering acquisitions now?
Huang Renxun:The operating systems of these large systems are extremely complex. How do you create an operating system in the compute stack to coordinate tens, hundreds, or billions of microprocessors in a GPU? This is a very difficult problem. If there are teams outside our company doing this, we can work with them, or we can do more.
Lauren Goode: Do you mean that it's critical for Nvidia to have an operating system and build it into a platform?
JH:We are already a platform company.
Lauren Goode: The more you become a platform, the more problems you face. People tend to take more responsibility for the platform's output. How autonomous cars behave, how much margin of error health care devices have, whether AI systems are biased. How do you solve this problem?
Hwang:well, we're not an app company. This is probably the easiest way to think. We will serve as little of an industry as we can, but as little as possible. So, in terms of healthcare, drug discovery is not our specialty, computing is. Building cars is not our specialty, but building computers that are extremely good at artificial intelligence for cars is our specialty. Frankly, it's hard for one company to be good at all of these things, but we can be very good at the AI computing part of it.
Lauren Goode: There were reports last year that some of your customers waited months for your AI GPU. How's it going?
Huang Renxun：（Well, I don’t think we’re going to catch up on supply this year. Not this year, and probably not next year.）
Lauren Goode: What is the waiting time right now?
Huang Renxun:I don't know what the delivery time is now. But, you know, this year is also the beginning of a new generation for us.
Lauren Goode: You mean Blackwell, your rumored new GPU?
Huang Renxun:Yes, a new generation of GPUs is coming soon, and Blackwell's performance is beyond imagination. It would be incredible.
Lauren Goode: Does this mean customers need fewer GPUs?
JH:That's the goal. The goal is to greatly reduce the cost of training the model. Then people can scale up the model they want to train.
Lauren Goode: Nvidia invests in a lot of AI startups. Last year it was reported that you invested in more than 30 startups. Do these startups queue up to buy your hardware?
Huang Renxun:They face the same supply shortage as everyone else, because most of them use public cloud, so they have to negotiate with public cloud service providers themselves. However, what they do get is our AI technology, which means they can use our engineering capabilities and our special techniques to optimize their AI models. We make them more efficient. If you increase throughput fivefold, you actually get five more GPUs. That's what they got from us.
Lauren Goode: Do you consider yourself a kingmaker in that regard?
Huang Renxun:No. We invest in these companies because their work is incredible. It's an honor to invest in them, not the other way around. They are some of the smartest people in the world. They don't need us to bolster their credibility.
Lauren Goode: What happens when machine learning turns more toward reasoning than training (basically, if the computational intensity of AI work decreases)? Will this reduce the need for GPUs?
JH:We like to reason. In fact, I'd say Nvidia's business today is probably 70% reasoning, 30% training, if I guess. And that's a good thing, because then you realize that AI has finally succeeded. If Nvidia's business is 90% training and 10% reasoning, you might say AI is still in the research phase. This was the case seven or eight years ago. But today, whenever you type a prompt in the cloud, it generates something-it can be video, it can be images, it can be 2D, it can be 3D, it can be text, it can be graphics-and it's most likely that there's an Nvidia GPU behind it.
Lauren Goode: Do you think the demand for AI GPUs will wane at any point?
Huang Renxun:I think we are at the beginning of the generative AI revolution. Today, most of the computing done in the world is still based on retrieval. Retrieval means that you touch something on your phone and it sends a signal to the cloud to retrieve a piece of information. It might compose a response with a few different things and use Java to render it on your phone's pretty screen. In the future, computation will be more RAG based (Retrieval-augmented generation, a framework that allows large language models to extract data from outside their usual parameters), with less retrieval and a much higher personalized generation.
That generation will be done by GPU. So I think we're at the beginning of this retrieval-enhanced generative computing revolution, where generative AI is going to be an integral part of almost everything.
Lauren Goode: The latest news is that you have been working with the U.S. government to develop sanctions-compliant chips that can be shipped to China. My understanding is that these are not state of the art. How closely do you work with the government to ensure you can still do business in China?
Huang:Well, it's actually export control, not sanctions. The United States has determined that NVIDIA's technology and AI computing infrastructure are strategic to the nation and will apply export controls to them. We first complied with export controls when--
Lauren Goode: August 2022.
Huang Renxun:Yes. And the United States added more provisions to export controls in 2023, causing us to redesign our products again. So that's what we did. We are developing a new suite of products that comply with today's export control rules. We work closely with governments to ensure that our proposals are consistent with their thinking.
Lauren Goode: How worried are you that these restrictions will spur China to launch competitive AI chips?
Huang Renxun:China has some competitive things.
Lauren Goode: Exactly. That's not data center scale yet, but Huawei's Mate 60 smartphone, launched last year, has attracted some attention because of its self-developed chips.
A:Really, really good company. They are limited by the semiconductor processing technology they have, but they are still able to build very large systems by aggregating many chips together.
Lauren Goode: Are you worried about whether China can catch up with the U.S. in the field of generative AI in general?
Huang Renxun:This regulation will limit China's ability to acquire the most advanced technology, which means that the Western world, that is, countries that are not subject to export control restrictions, will be able to acquire better technology and develop it quite quickly. So I think this restriction imposes a great cost burden on China. Technically, you can always aggregate more chip fabrication systems to do the job. But that only increases the unit cost of those products. This is probably the easiest way to think.
Lauren Goode: Does the fact that you are making compliant chips to continue selling in China affect your relationship with TSMC, the pride and joy of Taiwan Semiconductor?
Huang Renxun:No. Regulations are specific. It's no different than speed limits.
Lauren Goode: You've said many times that eight of the 35,000 components in your supercomputer come from TSMC. When I heard this, I thought it must be a small part. Are you downplaying your dependence on TSMC?
Huang Renxun:Not at all. Not at all.
Lauren Goode: So what point do you want to make?
Huang Renxun:I just emphasize that in order to build an artificial intelligence supercomputer, there are many other components involved. In fact, in our AI supercomputers, almost the entire semiconductor industry collaborates with us. We have worked closely with Samsung, SK Hynix, Intel, AMD, Broadcom, Marvell and others. In our AI supercomputers, when we succeed, a whole bunch of companies will succeed with us, and we're happy about that.
Lauren Goode: How often do you talk to TSMC's Zhang Zhongmou or Liu Deyin?
Huang Renxun:Every moment. Constantly. Yes, it is. Constantly.
Lauren Goode: What was your conversation like?
Huang Renxun:These days we talk about advanced packaging, planning for the next few years of production capacity, advanced computing power. CoWoS [TSMC's proprietary method of cramming chips and memory modules into a single package] requires new factories, new production lines and new equipment. So their support is really, really important.
Lauren Goode: I recently spoke to a CEO who is focused on generating AI. I asked who Nvidia's competitors might be, and this person suggested Google's TPU. Others mentioned AMD. I guess it's not a binary question for you, but who do you think is your biggest competitor? Who keeps you awake at night?
JH:Lauren, they all do that. The TPU team is excellent. Most importantly, the TPU team is really good, the AWS Trainium team and the AWS Infentia team are really good, very good. Microsoft is working on an internal ASIC called Maia. Every cloud service provider in China is building in-house chips, and there are a whole bunch of startups and existing semiconductor companies building great chips. Everyone is building chips.
This shouldn't keep me up all night-because I should make sure I'm so exhausted from work that no one can keep me up all night. It's really the only thing I can control.
But what certainly woke me up in the morning was that we had to keep delivering on our promise, which was that we were the only full-stack company in the world where everyone could collaborate to build a data center-scale AI supercomputer.
Lauren Goode: I have some personal questions for you.
Huang Renxun:[Huang to PR rep] She has done her homework. Not to mention, I'm just enjoying this conversation.
Lauren Goode: I'm glad. Me too. I do want to--
Huang Renxun:By the way, whenever Zhang Zhongmou or someone I know for a long time asks me to be the host of an interview, the reason is that I don't sit there and interview them by asking questions. I was just talking to them. You have to have empathy for the audience and what they might want to hear.
Lauren Goode: So I asked ChatGPT a question about you. I was wondering if you had a tattoo because I was going to propose it to you at the next party.
Huang Renxun:If you get a tattoo, I'll get one too.
Lauren Goode: I already have one, but I've been looking for extensions.
JH:I have one too.
Lauren Goode: Yes. This is what I learned from ChatGPT. Huang is said to have tattooed the company logo when the stock price hit $100. It then said,"However, Huang indicated that he was unlikely to get another tattoo and noted that the pain was more intense than he expected." They say you cried. Did you cry?
Huang Renxun:A little bit. My suggestion is that you should have a glass of whiskey before doing this. Or take Advil. I also think women can take more pain because my daughter has a sizable tattoo.
Lauren Goode: So, if you want a tattoo, I think triangles might be nice because who doesn't like triangles? They're perfect geometric shapes.
Huang Renxun:Or the outline of Nvidia Building! It's made up of triangles.
Lauren Goode: It's a promise. I wonder how often you personally use ChatGPT or Bard, etc.?
Huang Renxun:I have been using Bard, and I also like ChatGPT. I use both almost every day.
Lauren Goode: For what?
Huang Renxun:Research. Computer-assisted drug discovery, for example. Maybe you want to know about the latest advances in computer-assisted drug discovery. So you want to structure the whole topic so that you can have a framework and from that framework you can ask more and more specific questions. I really like these big language models.
Lauren Goode: I heard you used to lift weights. Do you still do that?
Huang:No, I try to do 40 push-ups a day. This should not take more than a few minutes. I am a lazy exerciser. I do squats while I brush my teeth.
Lauren Goode: Recently, you commented on the Acquired podcast, which went viral. If you were 30 years old today and thinking about starting a company, what would you start? And you said you couldn't even start a company. How do you modify this?
Huang Renxun:This question can be answered in two ways. I answer it like this: If I knew everything I know now, I would not dare to do it because I was afraid. I'd be too scared. I wouldn't do that.
Lauren Goode: You have to have certain delusions to start a business.
Huang Renxun:This is the advantage of ignorance. You don't know how hard it's going to be, you don't know how much pain and suffering it's going to involve. When I meet entrepreneurs these days, they tell me how easy it will be, I'm very supportive of them, and I'm not actually trying to pop their bubble. But I knew deep down,"Oh my God, things aren't going to be the way they think."
Lauren Goode: What do you think is the biggest sacrifice you have to make to run Nvidia?
Huang Renxun:Other entrepreneurs have made the same sacrifice. You work really hard. For a long time, no one thought you would succeed. You're the only one who believes you can succeed. The feelings of insecurity, vulnerability, and sometimes humiliation are all real. No one talks about it, but it's all true. CEOs and entrepreneurs are human beings just like everyone else. When they fail publicly, it's embarrassing.
So when someone says,"Jensen, with everything you have today, won't you start it?" It's like,"No, no, no, of course not."
In fact, if I had known Nvidia would be where it is today, you ask, would I have started this company? Are you kidding me? I am willing to sacrifice everything to do this.
