Source: There is a new Newin
This morning, Nvidia’s Q3 earnings report was announced after the U.S. stock market hours, with revenue of $18.12 billion for the third quarter ended October 29, 2023, an increase of 206% year-on-year, a quarter-on-quarter increase of 34%, EPS earnings increased nearly 6 times, nearly 13% and 20% higher than analysts’ expectations, respectively, and the data center revenue of the business where the AI chip is located increased nearly 2 times year-on-year, hitting a new high in a single quarter.
![]()
“Our strong growth reflects the transformation of a broad range of industry platforms from general-purpose to accelerated computing and generative AI, with LLM startups, consumer internet companies, and global cloud service providers being the first movers, and the next wave is starting to take shape, with national and regional communications service providers investing in AI clouds to meet local demand, enterprise software companies adding AI Copilot and Assistant to their platforms, and enterprises creating custom AI,” said Huang The era of generative AI is taking off with NVIDIA GPUs, CPUs, networking, AI foundry services, and NVIDIA AI Enterprise software as engines of full-speed growth!"
![]()
PS: In last weekend’s column, we shared John Luttig, Head of Investment at Funders Fund, on his analysis of the current GPU market landscape.
Last month, Jensen Huang, co-founder & CEO of Nvidia, also gave a very dry talk at Columbia Business School (CBS), where Huang spoke with CBS Dean Costis Maglaras to discuss the digital future, including how does NVIDIA do strategy and operations, and how does Huang have entrepreneurial experience and how to become a qualified CEO.
Here are some of the dry goods that Lao Huang shared in the process of CBS for you to try:
Before making a decision, everyone has to figure out what they’re doing, why they’re doing it, and it’s all about choice.
From a personal point of view: There are three things that need to be determined:
-
Difficult but right things;
-
what you are destined to do;
-
Things you like;
From the company’s point of view: Using NVIDIA as an example, Lao Huang’s answer was very straightforward, clearly explaining the market choices, business models, barriers, and flywheel effects involved in NVIDIA’s Pivot:
"The reason we don’t do manufacturing is because TSMC is doing so well and they’re already doing it, why should I go and take their jobs? I like the people at TSMC, they’re good friends to me, just because I have a business, I can get into this space, so what? They’ve done a great job for me, let’s not waste time repeating what they’ve already done, let’s waste time doing something that no one has done, something that no one has done, and that’s how do you build something special, otherwise, you’re just talking about market share. **
We observe two things: accelerated computing is a software problem, it’s an algorithm problem, and AI It’s a data center problem, so we’re the only company that goes out and builds all of these things, and part of what we do is the choice of business model, we could have been a data center company, fully vertically integrated, and yet, we recognize that no matter how successful a computer company is, it’s not going to be the only computer company in the world, and it’s better as a platform computing company because we love developers. Being a platform computing company that serves every computer company in the world is better than being a computer company alone. **
We’ve taken this approach, we’ve taken this data center that’s the size of this room, all the wires, all the switches and the networking, and a lot of software, and we’ve broken it all down and integrated it into other different data centers around the world, and it’s a crazy complexity, and we’ve found a way to have enough standardization when necessary, enough flexibility when needed so that we can work with computer companies all over the world enough.
The result is that Nvidia’s architecture is now implanted into every computer company in the world, which creates a bigger footprint, a bigger installed base, more developers, better applications, which makes customers happier, they buy more chips, which increases the installed base, increases our R&D budget, etc., the flywheel effect, the positive feedback system, that’s how it works, it’s simple and straightforward"
In addition, Lao Huang also made clear his views on AI and labor & workflow in his sharing - **AI will not take away your job, the people who use AI will take away your job, and if a company does not have more ideas to invest in incremental gains, then when the work is replaced by automation, the company has to lay off employees and join those companies that have more ideas but can’t afford to invest money, so that when AI automates their work, of course the situation will change, of course it will change the way of working. **
The following is the full content of the conversation between Lao Huang and CBS President Costis Maglaras, enjoy~
Costis Maglaras:
I want you to first take us back a little bit through the history of Nvidia, and then I want to talk about the leadership issue that we just mentioned, but you started this company 30 years ago and led it through a transformation that launched different applications and product types. Take us through the journey.
Jensen Huang:
One of my proudest moments. I started with one of the proudest moments that happened recently, when I was the CEO of the first company I worked for, Denny’s, and learned that Nvidia was not only my progression from dishwasher and handyman to the top of the company to a waiter at Denny’s, but that they were my first company and that I am still familiar with the menu. By the way, Superbird is great, does anyone know what Superbird is? What kind of college students are you?
![]()
Denny’s is a restaurant in the United States, and Nvidia was founded by me and two other co-founders in San Jose — Denny’s outside our house, so they recently contacted me, and the box we used to sit in is now Nvidia’s box, and it’s called Nvidia, and that’s the birthplace of a trillion-dollar company, and it’s a very proud moment.
![]()
Nvidia was founded at a time when the PC revolution was just beginning, and microprocessors captured the imagination of the entire industry. The world rightly sees how CPUs, microprocessors, are going to reshape the IT industry, how they will reshape the computer industry, and before and after the x86 revolution, successful companies were very different. We started our company during that period, and our view was that as amazing as general-purpose computing is, it can’t be the solution to all problems.
We believe that there’s a way of computing that we call accelerated computing, where you add an expert next to the general-purpose computing. The CPU is a generalist and can do anything, if you will. It can do anything. However, obviously, if you can do anything, then obviously you can’t do anything well.
As a result, we believe that there are some problems that are not suitable for solving by what we call ordinary computers. That’s why we started this accelerated computing company. The problem is, if you want to create a computing platform company, I don’t know how many computer scientists there are, but if you want to create a computing platform company, there hasn’t been a company like that since 1964, and that was the year after I was born, IBM Systems 360 perfectly describes what a computer is.
In 1964, IBM described that 360 had a central processing unit, I/O subsystem, direct memory access, virtual memory, binary compatibility across scalable architectures, and it described everything we have today as computers that we describe today, and 60 years later, we feel like there’s a new form of computing that solves some interesting problems, and it wasn’t entirely clear what we could solve at the time, but we felt there was a future for accelerated computing.
Still, we set out to start this company and made a really good first decision, and frankly, that decision has been incredible to this day, and if somebody comes up to you and says to you, we’re going to invent a new technology that doesn’t exist in the world, everybody wants to build a computer company around CPUs, we want to build a computer company around other things that are connected to CPUs, number one.
Number two, the killer app is a video game, a 1993 3D video game, and that app doesn’t exist, the company that built this company doesn’t exist, the technology that we’re trying to build doesn’t exist. So now you have a company that has both technical challenges and market challenges and ecosystem challenges, so the probability of success for this company is almost 0%, but either way, we’re lucky because of two very important people.
Frankly, the three of us co-founders worked together, they were very important figures in the technology industry at the time, and I called Don Valentine, the most important venture capitalist in the world at the time, and told Don to give this kid some money and then figure out along the way if it would work, and luckily they did, but that business plan, even today, I wouldn’t invest because it had too many dependencies, and each one had a certain probability of success.
When you add that up, multiply, you get 0%, but we imagine that there’s going to be a market called video games, and that’s going to be the biggest entertainment industry in the world, which was 0 at the time, and we speculate that 3D graphics are going to be used to tell the story of almost every sport, game. So, in the virtual world, you can have any game, any sport, and as a result, everyone will become a gamer.
Don Valentine asked me, how big is the market, and I said, everybody is going to be a gamer in the future, and the wrong answer when starting a company, and frankly, these are bad habits, bad skills, and I’m not suggesting that, but anyway, it turned out to be true, and video games became the biggest entertainment industry in the world, 3D Graphics was a success, and we found the first killer application of accelerated computing, which bought us time, using accelerated computing to solve a range of other problems, and finally moved to AI.
Costis Maglaras:
It’s a great story, and before we talk about AI, I want to ask a little bit about the days of cryptocurrency, obviously, gaming was a big step for Nvidia, and then at some point, the killer app became crypto and mining, what was that development?
Jensen Huang:
Accelerated computing solves problems that ordinary computers can’t. All of our GPUs, even if you use it to design cars, architecture, do molecular dynamics studies, play video games, it has a programming model that we invented called CUDA. CUDA is the only computing model that exists today and is as popular as x86, and it is used by developers all over the world.
In any case, CUDA is able to do parallel processing very quickly, and obviously, one of the algorithms that we can handle very well is cryptography. When Bitcoin first came out, there were no ASICs for Bitcoin, and the obvious thing to do was to go to the fastest supercomputer in the world, and the supercomputer with the highest production volume was none other than Nvidia’s GPUs, which were in the homes of millions of players, so by downloading an app, you could mine cryptocurrencies from your home.
The fact that you can buy our GPUs, our computers, plug them in, and the money starts pouring out. That was the day my mom understood what I was doing. One day she called me and said, son, I thought you were doing something about video games, and I finally understood what you were doing, and you bought an Nvidia product, plugged it in, and the money started pouring out.
And I said, yes, that’s what I’m doing, and that’s why so many people buy Bitcoin, which subsequently led to the rise of Ethereum, but you’re going to use a supercomputing system like an Nvidia GPU to encode or compress, or do something to refine the data, and turn it into a valuable token, and you know what that sounds like? ChatGPT that generates valuable tokens.
One of the things that has happened so far is that if you stretch your thinking about Ethereum and crypto mining, it makes sense in a sense because we suddenly create this new type of industry, where raw data comes in, you apply energy to this computer, and literally money starts pouring out, and these currencies are of course in the form of tokens, and these tokens are smart tokens Now I’m just describing something else that makes a lot of sense to us today, but it seems strange at the time, you take water in a building, you heat it, and what comes out is a very valuable and invisible thing called electricity. **
Today we’re moving data to data centers, and it’s going to refine and process it, and use its ability to generate a lot of valuable digital tokens, in digital biology, they’re going to be valuable, in physics, in IT and all kinds of computing, social media, all sorts of things, computer games, and so on, they’re coming in the form of tokens, so the future is going to be about AI factories, and Nvidia’s devices are going to power those AI factories.
Costis Maglaras:
So we’ve jumped to neural networks, and I think we’ve talked about parallel computing, like how to render graphics on a monitor, how to play games, how to solve cryptographic problems for Bitcoin. Tell us a little bit about what GPUs are used for training neural networks, and I want to talk to the audience here, what does it take to train a model like ChatGPT, what hardware do you need, what data do you need, how big a cluster do you need, how much does it cost, because those are huge questions, and I think it would be nice for you to give us some idea of scale.
Jensen Huang:
Everyone wants you to think it’s a huge problem and very expensive. Actually, no, let me tell you why, our company spent about $5~$600 million in engineering costs to design a chip, and then one to two years later, I hit enter, send an email to TSMC, send a big file to TSMC via FTP, and they would make it, and the process cost our company about $500 million.
For a total of $5.5 billion, I got a chip that was certainly valuable to us, but it wasn’t a big deal. I’ve been doing this, so if somebody, hey Jensen, you need to create a billion dollar data center, and once you’re plugged in, the money is going to gush out from the other side. I’m going to do it right away, and obviously a lot of people will do it too, because who wouldn’t want to create a factory that generates intelligence?
Now $1 billion isn’t really a lot of money, and frankly, the world is spending about $250 billion a year on infrastructure computing infrastructure, and none of us are generating money, it’s just storing our files, passing our emails, that’s already $250 billion, and one of the reasons we’re growing so fast is that, after 60 years of development, general-purpose computing is declining because of another 2500 It would be unwise to create another general-purpose computing data center with billions of dollars, which is too crude in energy and too slow in computing. **
Now that accelerated computing is here, that $250 billion will go to create accelerated computing data centers, and we’re excited to support customers in doing so. On top of that, accelerated computing, you now have an infrastructure to generate AI, and like all the things that we just talked about, basically the way it works is that you take a lot of data, and then compress it. **
Deep learning is like a compression algorithm, where you’re trying to learn the mathematical representations, patterns, and relationships of the data that you’re working on and compress it into a neural network, so the input is, let’s say, trillions of bytes, trillions of tokens, so let’s say trillions of bytes, and the output is 100GB, so you’ve compressed all that data into this little file, and 100GB is like 2 DVDs that you can download and watch on your phone, right?
So, you can download this huge neural network on your phone. Now, all of this data has been compressed in, and this compressed neural network model is an LLM, meaning you can interact with it, you can ask questions, and it will come back to its memory, understand your intentions, and generate text for you, have a conversation with you, so, the core is that, sounds magical, but to all the computer scientists and scientists in the room, it’s very reasonable, don’t let anybody convince you that it’s going to cost a lot of money, I’ll give you a good discount, everybody go and create AI bar.
Costis Maglaras:
If I were to pursue that scale a little more, you would need a computer that is basically the data center equivalent to estimate these models.
Jensen Huang:
**What is needed to create GPT-4 is 16,000 GPUs, which is the largest model anyone has ever used, worth $1 billion, and this is just a check, not even a big one, don’t be afraid, don’t let anyone dissuade you from starting a business and making your dreams come true. **
Costis Maglaras: Let me ask you a question about the billion dollar check and the growth you’re experiencing. I think you’ve been named the best CEO by the Harvard Business Review, and that’s entertaining. I’m going to keep repeating this, but in a sense, you’re leading a company through extreme growth, super growth, which most companies haven’t experienced in their lifetimes, and I want to ask you to tell us a few details, like doubling the size in a year or managing the supply chain, managing the customers, managing the growth, managing the money, how did you do that?
Jensen Huang:
I love management, and the only part of it, which is counting money, is fun. Wake up in the morning and roll around on all the cash, isn’t that what all of you are here for? I understand that this is the ultimate goal, it’s hard to build a company, there’s nothing easy to do, there’s a lot of pain and suffering, it takes a lot of effort. **
If it’s easy, everybody will do it, and about all companies, big or small, whether it’s ours or other technology companies, you’re always dying, because there’s always someone trying to outdo you, so you’re always on your way to bankruptcy, and if you don’t internalize that feeling, if you don’t believe that, you’re going to go bankrupt. And I originally started at Denny, and as you all know, Nvidia was built in an extremely unlikely situation. It took us a long time to get to where we are today. I mean, we’re a 30-year-old company. When Nvidia was first founded, in 1993, Windows 95 had not yet been launched. At that time, it was the first available PC, and we didn’t have email.
There were no laptops or smartphones at that time. All of these things don’t exist, so you can imagine how different the world we had when we first started and how different it is now. We don’t have an LCD screen. All are cathode ray tubes (CRTs). In those days, even CD-ROMs didn’t exist. In short, these things are the context of the times when we were founded, and it took us so long for the company to be recognized as the first company to reinvent computing in 60 years. Rapid growth depends on people.
Obviously, the company is all about people, and if you have the right system and you have people like me around you, the company will have the skills. It doesn’t matter if you’re selling $100 billion or $200 billion.
Now the truth is, the supply chain is not simple, does anyone know what a G-Force graphics card looks like? Raise your hand, does anyone know what an Nvidia graphics card looks like, so you would think that a graphics card is like a cartridge that plugs into the PC Express slot of a PC, but the graphics chips that we have now, used in these deep learning systems, have 35,000 parts and weigh up to 70 pounds, because they’re so heavy, they need robots to assemble, they need a supercomputer to test because it’s a supercomputer in its own right, and it costs $200,000, and with that $200,000, you can buy a computer like this, and it can replace hundreds of general-purpose processors, and those processors cost up to millions of dollars, and for every $200,000 spent on buying at Nvidia, you save $250 $10,000 to calculate the cost, that’s why I’m telling you, the more you buy, the more you save; obviously, this strategy is very successful, people are really lining up to buy, that’s what we do;supply chains are very complex, we make the most complex computers in the world, but how hard is that? It’s actually very hard, and the core is that if you’re surrounded by great people, the simple truth is, it’s all about people; I’m lucky to have a great management team around, and then the CEO will say something like" Make it number one", such as “let it work”.
Costis Maglaras:
I want to go back to AI trends and your vision for the future, but you mentioned the word “platform” earlier, and you mentioned your software environment. So you have the hardware infrastructure, you have a software environment that is currently ubiquitous in terms of training neural networks. Are you building data centers, or creating environments within data centers that are made up of Nvidia’s hardware, software, and communication clusters between those resources, how important is it to be a complete platform solution and just hardware involvement? How central is that to Nvidia’s strategy?
Jensen Huang:
I think, first of all, before you can create something, you have to know what you’re creating and why you’re creating it, what are the first principles of its existence. **
Accelerated computing is not a chip, that’s why it’s not called an accelerator, accelerated computing is about understanding how you can accelerate everything in life. If you can accelerate every application, that’s called really fast computing, so accelerated computing is first about understanding which domains, which applications are important to you, and understanding the algorithms, computing systems, and architectures needed to accelerate those applications.
It turns out that general-purpose computing is a reasonable idea, as is speeding up an application. As an example, you have a DVD decoder. You use your phone to play a DVD or h.264 decoder. It does one thing, and it does it very well. No one knows how to do it better.
Accelerated computing is a bit like this weird intermediate state. There are many apps that you can speed up. For example, we can accelerate various image processing, particle physics, and more, including linear algebra. We can speed up a lot of applications, and that’s a challenge, it’s usually easy to speed up one thing, and it’s easy to run everything with a C compiler.
Accelerate enough domains that if you accelerate too many domains, you’re back on general-purpose processors, right? Why can’t they make a faster chip? On the other hand, if you only accelerate one application, then the market isn’t big enough to support your R&D.
So we have to find the middle point of that switch, and that’s the strategic journey of our company, and this is where strategy meets reality, and this is where Nvidia gets it right, and it’s where no other company in the history of computing gets it right; find a way to have a large enough application area that we can accelerate, which is still 100~500 times faster than CPUs, so that the economic flywheel effect can scale the number of applications, expand the number of customers, expand the number of markets, Increasing sales, and thus creating larger R&D budgets, allows us to create more amazing things and stay far ahead of the CPU,**Does that make sense?
It’s very hard to create this flywheel effect, no one has done it before, only once, and that’s ability. In order to do that, you have to understand the algorithm, you have to understand the application domain very well, you have to choose right, you have to create the right architecture for it**, and then the last thing that we do right is, we realize that in order to have a computing platform, the application that you develop for Nvidia should run on all Nvidias, and you shouldn’t think about, is it running on this chip? Will it run on that chip? It should run on every computer that has Nvidia on it.
That’s why every GPU our company creates, even if no customer used CUDA a long time ago, we’re committed to it. We were determined to create this computing platform from the very beginning. Customers are not, it’s a 10-year, multi-billion dollar hardship for the company. If it weren’t for all the video gamers here, we wouldn’t be here. You’re our day-to-day job, and in the evenings we can go and solve digital biology, help people solve quantum chemistry, help people with AI and robotics, and so on.
We realized that, first of all, accelerated computing is a software problem, and secondly, AI is a data center infrastructure problem, which is very obvious because you can’t train an AI model on a laptop, you can’t train on a phone because it’s not a big enough computer, the amount of data is calculated in terabytes, and you have to deal with those trillion bytes, billions of times, so obviously, it’s going to be a massive computer, and the problem is spread across millions of GPUs.
I say millions because there are tens of thousands inside 16000. As a result, we’re spreading the workload across millions of processors. No application in the world today can be spread across millions of processors; Excel runs on a single processor. So this computer science problem of distributed computing is a huge breakthrough, definitely a huge breakthrough, and that’s why it’s able to enable generative AI, enable LLMs.
We observe two things: accelerated computing is a software problem, it’s an algorithm problem, and AI It’s a data center problem, so we’re the only company that goes out and builds all of these things, and part of what we do is the choice of business model, we could have been a data center company, fully vertically integrated, and yet, we recognize that no matter how successful a computer company is, it’s not going to be the only computer company in the world, and it’s better as a platform computing company because we love developers. Being a platform computing company that serves every computer company in the world is better than being a computer company alone. **
We’ve taken this approach, we’ve taken this data center that’s the size of this room, all the wires, all the switches and the networking, and a lot of software, and we’ve broken it all down and integrated it into other different data centers around the world, and it’s a crazy complexity, and we’ve found a way to have enough standardization when necessary, enough flexibility when needed so that we can work with computer companies all over the world enough.
The result is that Nvidia’s architecture is now implanted into every computer company in the world, and that creates a bigger footprint, a bigger installed base, more developers, better applications, which makes customers happier, they buy more chips, which increases the installed base, increases our R&D budget, and so on, the flywheel effect, the positive feedback system, and that’s how it works, it’s simple and straightforward. **
Costis Maglaras:
One of the things that you didn’t do, and I want you to explain, is that you didn’t invest in making your own chips.
Jensen Huang:
That’s a good question, and the reason is that as a strategic choice, our company’s core values, my personal core values, our company’s core values are all about choice.
The most important thing in life is choice. How do you choose? Well, everything is, how do you choose what to do tonight? How do you choose? Our company decided to choose the project with only one fundamental goal, and my goal is to create an environment, an environment where the best people in the world come and work here. An amazing environment for the best minds in the world, who want to pursue the fields of computer computing, computer science, and AI, to create the conditions for them to come here and do their life’s work. **
So, if I say that, the question now is, how do you achieve this? Let me give you an example of how you don’t have to do this. No one I know who wakes up in the morning and says, you know, my neighbor is doing that. What I want to do is, I want to take it from them. I can do that too. I want to take it from them. I want to grab their market share. I want to suppress them on price, I want to kick them, I want to take their share.
It turns out that no great person does this, and everyone wakes up in the morning and says, I want to do something that has never been done before, which is very hard to do. If you succeed, you can make a huge impact in the world, and that’s NVIDIA’s core values.
Number one, how do we choose to do something that has never been done before in the world? By the way, the reason why you choose to do something incredibly difficult is because you have a lot of time to learn it if something is easy to do, like TikTok Dance, I’m not going to bother about it, obviously the reason is that there is a lot of competition, so you have to choose something that is really hard to do, and those hard things in themselves will stop many others, because the one who is willing to endure the longest will eventually win, so we choose something that is very difficult to do, and you have heard me say many times that pain and suffering, and that is actually a positive trait and the one who is able to endure ends up being the most successful.
Number two, you should choose something that you are destined to do, whether it is your personality traits, your expertise, or the environment you’re in, your size, whatever you have, your perspective, what you’re meant to do. **
Third, you’d better enjoy doing that thing very much, because unless the pain and suffering is too great. Now what I’ve just described to you are NVIDIA’s core values. It’s as simple as that. If that’s the case, why would I make phone chips? How many companies in the world can make mobile phones? a lot. Why do I need CPUs? Do we need more CPUs? Is that reasonable? We don’t need all these things.
As a result, we naturally exclude ourselves from the mass market. We naturally excluded ourselves from the mass market because we chose amazing markets, we chose really hard things to do, amazing people joined us because amazing people joined us because we had the patience to make them successful and do something amazing. Have the patience to let them do something amazing, and they will do something amazing.
Is it reasonable that the equation is actually that simple, but it takes incredible character to do? That’s why learning it is the most important thing, great success and greatness are all about character. The reason we don’t do manufacturing is because TSMC is doing so well, and they’re already doing it, so why should I take their job? I like TSMC people, they’re good friends of mine, and just because I have a business, I can get into this field, so what? They’ve done a great job for me, let’s not waste time repeating what they’ve already done, let’s waste time doing something that no one has done, something that no one has done, and that’s how do you build something special, otherwise, you’re just talking about market share. **
Costis Maglaras:
Thinking about the future, when we think about these 10 years.
Jensen Huang:
Correct answer?By the way, I know I don’t have an MBA, I don’t have a degree in finance, I read some books, I watch a lot of Youtube videos, and I have to tell you, no one watches more business YouTube videos than I do, so I can tell you guys, you guys are nothing good for me, but these are the correct answers, Professor Maglaras?
Costis Maglaras:
You’re asking the wrong person, and I haven’t studied business either, but they’re the right answer haha~ What do you think AI, when you think AI applications and the changes that we’re going to see in the next three, five, seven years, and what might be affected in our daily lives?
Jensen Huang:
First of all, I’m going to jump straight to the conclusion, AI doesn’t take your job, people who use AI take your job. Do you agree with that?Well, use AI as soon as possible so you can maintain beneficial employment.
The second thing I ask you guys is, when productivity goes up, it means that we’re fully embedded in AI at NVIDIA, and NVIDIA is going to be a huge AI entity, and we’re already designing our chips with AI, and we can’t design our chips, and we can’t write our optimized compilers without AI, so we’re using AI everywhere.
When AI increases your company’s productivity, what’s next? layoff or more people? you’re going to hire more people. Profitable growth was due to increased productivity.
Why do people think about losing their jobs? If you think you don’t have a new idea, it doesn’t make sense. If you don’t have more ideas to invest in your incremental gains, what do you do when jobs are replaced by automation? You’re going to lay off people and join companies that have more ideas and can’t afford to invest the money so that when AI automates their work, of course things change, of course change the way things work. **
AI will soon be targeting CEOs, department chairs and CEOs, we’re done, sounds good, I think first the CEO, then the department chairs, but you’re close, so you join companies that have more ideas and don’t have enough money to invest, and naturally, when the earnings go up, you hire more people. First of all, it’s a huge breakthrough, we somehow taught computers how to learn and represent information digitally, okay? So, have any of you ever heard of this thing called Word2vec? It’s one of the best things ever, Word2vec, you take a word and you learn by studying each word and how it relates to every other word, you study all of our sentences and paragraphs, and you try to figure out what is the vector of numbers that are most relevant to that word, what numbers are most relevant to that word, so “mother” and “father” are close to each other numerically, “orange” and “apple” are close to each other numerically, but they are far from “mom” and “dad”, “dog” and “cat” Far from “Mom” and “Dad”, but probably closer than they are to “oranges” and “apples”, chairs and tables, it is difficult to say exactly where they are, but these two figures are close to each other, away from “Mom” and “Daddy”, “King” and “Queen”, closer to “Mom” and “Daddy”.
Is that reasonable? Imagine doing this for every number, and every time you test it, you’re like, gosh, this is great. It makes sense when you subtract something from another. Well, that’s basically the representation of learning information. Imagine doing this to the English language. Imagine doing this for every language. Imagine doing this to anything that has structure, meaning anything that is predictable.
The image has structure, because if there is no structure, it would be white noise, in fact, white noise, so there must be structure, and that’s why you see a cat, I see a cat, you see a tree, I see a tree, you can identify where the tree is, you can identify where the coastline is, where the mountains are, where the clouds are, right? We can learn all of that, obviously you can convert that image into a vector, you can convert the video into a vector, 3D Converted into vectors, proteins into vectors, because proteins obviously have structures, chemicals are converted into vectors, genes are eventually converted into vectors, and we can learn the vectors of everything.
If you can learn everything into numbers, and it makes sense, then obviously you can convert the word cat “cat” into an image, which is obviously not an image of a cat, it’s the same meaning, if you can convert from words to images, that’s called intermediate journey steady diffusion, if you can convert from images to words, it’s called subtitling, subtitles under YouTube videos, so if you go from, what do you call it? If you convert from amino acids to proteins, it’s called the Nobel Prize, because it’s an alpha fold, an incredible breakthrough.
So, this is an amazing moment in computer science, where we can really transform one type of information into another kind of information, so you can do text-to-text, lots of text, PDFs to small amounts of text, aggregate archives, which is what I really like, right?
We can ask it to aggregate this paper, and instead of reading every single paper, it has to understand the images, because in the archive, the paper has a lot of images, charts, and things like that, so you can put all of that together, so you can now imagine all the productivity benefits, and actually the ability to do it without it, so in the near future, you’re going to do that.
You can say, hey, I want to design, give me some options for cars. I work at Mercedes and I care a lot about the brand, it’s the style of the brand, let me give you a couple of sketches, maybe a couple of photos of the model that I want to build, which is a four-wheel drive SUV, let’s say, and then all of a sudden, it came up with 2010, 200 full 3D design CAD; now, the reason you want this and not just finish this car is because you might want to pick one of them and say iteration 10 on top of that Second, you may end up choosing one and then making your own modifications, so the future of design will be very different. Everything is going to be very different in the future, and now if you give designers this ability, they will go crazy. They will love you very much, and that’s why we do it.
So, what are the implications for the long-term effects? One of my favorite areas is that if you can describe a protein in words, and you can find out how to synthesize a protein in words, then the future of protein engineering is right now. As you know, protein engineering involves making enzymes to break down plastics, making enzymes to capture carbon, making all kinds of enzymes to grow vegetables better, your generation can create all kinds of different enzymes, so the next 10 years are going to be incredible, we’re the generation of computer chip engineering, you’re going to be the generation of protein engineering, which we couldn’t have imagined a few years ago. **
Costis Maglaras:
Okay, I think we’re going to open up the question session to the audience, so if there’s a question, maybe I’ll point, we’ll have some microphones coming through, okay, over there we’ll start first.
Spectator:
Thank you for being here tonight, are you worried about whether Moore’s Law will catch up with the GP industry as it did with Intel? Can you explain the difference between Moore’s Law and Huang’s Law? Jensen Huang: I didn’t bring up Huang’s Law, and it’s not like something I’d do. Moore’s Law is that performance doubles every year and a half, and the easier way to calculate is to grow by 10 times every 5 years, so it’s about 100 times every 10 years. If so, if general-purpose computing is a microprocessor, why change the calculation method if general-purpose computing grows by 5 times every 10 years, and every 100 years by 100 times? Isn’t that fast enough? Are you kidding? Wouldn’t life be good if cars were 5 times faster every 100 years?
So the answer is, actually, Moore’s Law is very good, and I benefited from it. The whole industry has benefited from this, and the computer industry exists because of it, but ultimately Moore’s Law of Universal Computing, it’s not about the number of transistors in computing, it’s about the number of transistors, how you use it for the CPU, how do you end up translating that into performance, that curve is no longer 10 times every 5 years. If you’re lucky, that curve is two to four times every 10 years. The problem is that the curve is 2~4 times every 10 years.
Computing needs and our vision of using computers to solve problems, our imagination, the imagination of using computers to solve problems is not more than 4 times every 10 years? So our imagination, our needs, the world’s consumption of all of this is beyond that limit, and you can solve this problem by buying more CPUs, you can buy more, but the problem is these CPUs Consumes too much energy because they’re generic, like a generalist, a generalist isn’t as efficient as a specialist, their craft isn’t as good as an expert, they’re not as productive as an expert; if I’m going to have a thoracotomy and don’t get me a generalist, you know what I mean? If you’re around, call an expert, so the way the journalist is too brute force, so now it’s making the world consume too much energy, making the world spend too much, just to brutally force universal computing.
Now luckily, we’ve been working on accelerated computing for a long time, and as I mentioned, accelerated computing isn’t just about processors, it’s really about understanding the application domain, and then creating the necessary software, algorithms, architectures, and chips, and we somehow found a way to do that with an architecture, and that’s the genius of what we’ve done, and we’ve somehow found this architecture, which is both very fast, sometimes to speed up CPUs by 100*500 times, even sometimes 1000 times, but it’s not that specific, it’s only for a single activity, is that reasonable? And you need to be broad enough so that you have a big market, but you need to be narrow enough so that you can accelerate the application, and this delicate line, this razor’s edge, is the reason Nvidia exists. If I had explained this 30 years ago, no one would have believed it, in fact, if you were honest now, no one would have believed it either.
It took us a long time, we stuck with it, we started with seismic processing, molecular dynamics, image processing, and of course computer graphics, and we worked on and on, and on and on, and then one day deep learning, and then with transformers, and then there was some form of reinforcement learning transformer, and then there would be some multi-step inference system, so all of those things, we’re just an application.
Somehow, we found a way, we created an architecture that solved all these problems, and will this new law end? I don’t think so. The reason is this, it doesn’t replace the CPU, it complements the CPU, so the question is, what will be there next to complement us?
We’re just hooking it up next to it, so when the time comes, we’ll know that we should use another tool to solve the problem because we’re serving the problem we’re trying to solve. We’re not trying to make a knife and get everyone to use it. We’re not trying to make a pair of pliers for everyone to use. We’re here to accelerate computing to serve the problem, so that’s one thing for all of you to learn. Make sure your mission is right. Is it reasonable to make sure that your mission is not to make trains, but to facilitate transportation? Our mission is to speed up applications and solve problems that ordinary computers can’t solve. If your mission is well-articulated and you focus on the right things, it will last forever. **
Spectator:
Again, thankfully, there is now a push to localize the semiconductor supply chain, and then there are restrictions on the export of high-tech products from certain countries. What impact do you think this will have on NVIDIA in the short term, and what impact will it have on our consumers in the long term?
Jensen Huang:
That’s a good question. You’ve all heard this, and I repeat, it’s about geopolitics and geopolitical tensions, etc. Geopolitical tensions, geopolitical challenges will affect every industry, affect everyone. Our company believes in national security, we are here because our country is safe, and we also believe in economic security.
The truth is, most families wake up in the morning and don’t say, oh my God, I feel so vulnerable, because of the lack of military force, they feel vulnerable, because of economic viability, so we also believe in human rights, and being able to create a prosperous life is part of human rights. As you know, the United States believes in the human rights of those who live here as well as those who don’t, so this country believes in all these things at the same time. So do we.
The challenge of geopolitical tensions is that if we decide too unilaterally, we decide on the prosperity of others, then there will be a backlash. There will be unintended consequences, but I’m optimistic. I hope to be able to hope that those who think about this issue have considered all the consequences and unintended ones, but this has led to the deep internalization of the sovereign rights of each country. Every country is talking about their own sovereignty, which is another way to say that everyone is thinking about themselves.
As far as we’re concerned, on the one hand, it could limit the use of our technology in China, as well as export controls there, and on the other hand, since sovereignty and every country wants to build its own sovereign AI infrastructure, and most of them are not enemies of the United States and don’t have a difficult relationship with the United States, we will help them build AI infrastructure around the world.
So in a lot of ways, this weird thing about geopolitics, it kind of limits our market opportunities. On the other hand, it opens up market opportunities for us in other ways, but for people, I am, I really want to.
I really hope that we don’t let our tensions with China evolve into tensions with the Chinese, we don’t let our tensions with the Middle East evolve into tensions with Muslims, and we can’t allow ourselves to fall into that trap, and I’m a little worried that it’s a slippery slope.
One of the largest sources of intellectual property in our country, as you know, is foreign students, and I see a lot here. I want you to stay here, this is one of the greatest strengths of our country. If we don’t allow the world’s brightest minds to come to Colombia and stay in New York City, we won’t be able to keep the world’s greatest intellectual property, so that’s our fundamental core strength, and I really hope we don’t undermine it.
You can see that geopolitical challenges are real, national security issues are real, but economic, market, social, technological issues are just as real, technology leadership is important, market leadership is important, all of these are important, the world is just a complex place, I don’t have a simple answer, all of us will be affected.
Spectator:
I started out as an engineer in a semiconductor company, working as an entrepreneur, and in the case of me like you, as a technologist and engineer at heart, successfully started a company, and I learned about finance from YouTube videos, what do you think about an MBA?
Jensen Huang:
I think that’s pretty awesome. First of all, you’re probably going to live to be 100 years old, so the question is, how are you going to spend the last 7 or 60 years? It’s not what I told you, it’s what I told everybody, care about education as much as you can.
When you come here and you’re forced to get an education, how good can that be? After leaving, like me, I have to go around the globe in search of knowledge, I have to go through a lot of garbage to find something good, and in the school, you have these amazing professors who sift through the knowledge for you and present it to you like a plate, my God, if I could do it again, I’d stay here for as long as I could and absorb a lot of knowledge. **
I’ll be sitting here with the dean. I’m the oldest student here. I’m just preparing for a huge leap when I graduate, and I’m going to be successful right after graduation, but I’m just kidding. You’re going to have to leave someday. Your parents will appreciate it, but don’t rush it. I think learn as much as you can. There is no single right answer to get there.
Obviously, I have friends who have never graduated from college but are very successful, so there are multiple ways to get there, but statistically I still think it’s the best way to get there, so if you believe in statistics and math, just stay in school and go through the whole process, so**I got a virtual MBA by working hard, not because of choice, but because when I first graduated from school, I thought I was going to be an engineer, no one would say," Hey, Jensen, give you a diploma and you’re going to be the CEO. "I didn’t know, so when I got there, I had to go and study. **
There are a lot of ways to get an MBA and learn business strategy, obviously business issues are very different things, financial issues as well, so you have to learn all these different things to build a company, but if you’re surrounded by amazing people like me, they’re going to teach you along the way, so some things, depending on the role you want to play, are crucial, and there are some things that are not just my job, but they’re critical, and I’m going to head with that. That’s character, there’s something about your character that matters about the choices you make, how you deal with success, how you deal with failure and big setbacks, how you make choices. **
Now, in terms of skills and craftsmanship, the most important thing for a CEO is strategic thinking, and there is no other choice. Companies need you to think strategically because you see the most, you should be able to see the future better than anybody, you should be able to connect the dots better than anybody, you should be able to mobilize, remember what strategy is – action! So, the CEO is uniquely placed in the right position to be a chief strategy officer, if you will. From my point of view, these two are the most important things, and the rest have a lot of skills and stuff that you will learn skills.
If I may add one more thing, I do believe that a company is about a particular craft, you make some unique contribution to society, you make something. If you make something, you should be good at it, you should appreciate the craft, you should love the craft, you should know something about the craft, where does it come from, where is it now, where is it going in the future, you should try to show your passion for this craft.
I hope that today I did something that exemplifies the passion and expertise of this craft, that I know a lot about my field, and that the CEO should know about this craft if possible. You don’t have to create this craft, but it’s better to be this craft, you can learn a lot, so you, you want to be an expert in this field, but these are some things. You can learn that here. Ideally, you can learn this at work, you can learn this from your friends, and you can learn this by doing a lot of different things.
