Welcome back to the stage, Patrick Collison. -[Patrick] All right. Good afternoon, folks. I hope you've enjoyed the Sessions between now and when we last saw you this morning. For this afternoon's keynote, or fireside chat, I suppose, I'm about to introduce somebody who needs little introduction. Although a fun fact that you may not know about Jensen Huang is that he's been a CEO of NVIDIA for 31 years this month, making him the longest-serving CEO in the technology industry and therefore logically-- John and I have only been doing it for a mere 14 years. So, even if
we double that, we'll still be second to him. Jensen, well, we'll talk about this on stage, attended the Oneida Baptist Institute in Kentucky. We'll definitely be asking him about it. Oregon State, worked as a waiter at Denny's. There's a Denny's close to here, actually. LSI Logic and then AMD, which is of course now run by his first cousin once removed. We'll definitely be asking about that. Before he founded NVIDIA in 1993... and NVIDIA's market cap was $8 billion when Stripe launched in 2011. And it is now, of course, more than 200 times that. So he's
been busy since. Please welcome to the stage, Jensen Huang. -[Jensen] Hey, everybody! -[Patrick] So you watched the keynote earlier? -I did. I've never seen a duet before. -[Patrick laughing] So, well-- -I've never seen a duet before. You were so synchronized. It seemed like the two of you knew each other. It's incredible. -Some acquaintance. You've been doing keynotes a long time. You are the <i>keynote goat</i>. So give us-- -Stop it. -Give us your, like... we don't have even a signature outfit yet. We're just amateurs here. So give us-- -It's because you're still young. -Well, give us
your keynote performance review. What did you think? -I thought it was A+. I thought it was A+. I was... really! You explained perfectly the purpose of the company. What inspires you guys. What keeps you guys up. What makes you work so hard. The ecosystem that you serve, the incredible platform you built, the amazing contribution you make to the world's economy. It's incredible. I thought it was great! And there was a whole bunch of technology stuff, feature stuff, money stuff. I didn't understand any of that, but... something about a CYK or something, what was that? -A
KYC. -KYC, yeah. I thought it was-- -It's a big deal in our world. -Is that right? Kentucky Fried Chicken? -We take care of KYC so that you can associate us with Kentucky Fried Chicken. -Okay. Got it. -Software-defined financial services, this idea, did that make sense to you? -Well, first of all, I think it's a giant idea. -Do you know where it came from? -You're going to tell me? -So Jensen and I were catching up... -The part that I loved was how you realized in the very beginning that financial payments was about code, not finance. I
thought that was incredible. And you explained that the first time we met. So, Jensen and I were catching up 18 months ago or so, and I guess it was a couple of years since we had last spoken. And so he was asking for the update on Stripe, and I was explaining. And you said, "Oh, so it's like software- defined networking, but for money." And that was still ricocheting around in my mind. That's where we got to this idea for software-defined financial services. So I hope we don't have to pay a licensing fee for that. -I
got zero equity for that good idea. -All right. You guys are doing okay. I was thinking about this. Tesla's earnings were reported yesterday, and Elon announced that Tesla is going to have 85,000 A100s by the end of this year. And I was just reflecting on-- it's quite a success to sort of build a business where CEOs kind of compete with each other to announce who has spent more buying your product. So I think you've done something quite impressive. But anyway, I actually want to start out talking a little bit about-- -All of my CEO friends,
they all have the most. -So I want to start out talking a little bit about, a remark you made at a Stanford event recently, at the GSB, I think. And you said, "I wish upon you ample doses of pain and suffering." Elaborate. -Well. Let's see. There is a misunderstanding. There's a phrase that said, "You should choose your career based on your passion." Usually people connect passion with happiness. And I think there's something missing in that. Nothing there is wrong, but there's something missing. And the reason for that is because if you want to do great
things, and I know this to be true about you creating Stripe-- And by the way, this is one of the world's finest CEOs, young as he may be, yep. You guys know, I've met a lot of CEOs, I've heard about a lot of companies, and this is genuinely one of the world's great visionary companies. And so anyways, I just want to say that. It's the reason why I just love what... -No more compliments allowed, it makes us terribly uncomfortable. -I know, I could tell. I could see him, he's starting to sweat. And so the thing
is, when you want to build something great, it's not easy to do. And when you're doing something that's not easy to do, you're not always enjoying it. I don't love every day of my job. I don't think every day brings me joy, nor does joy have to be the definition of a good day. And every day I'm not happy. Every year I'm not happy about the company. But I love the company every single second. And so I think that what people misunderstand is somehow the best jobs are the ones that brings you happiness all the
time. I don't think that that's right. You have to suffer. You have to struggle. You have to endeavor. You have to do those hard things and work through it in order to really appreciate what you've done. And there are no such things that are great that were easy to do. And so by definition, I would say therefore, I wish upon you greatness, which, by my way of saying it, I wish upon you plenty of pain and suffering. -Anything in your upbringing that taught you that idea, or is it just somehow innate to your makeup? -I
didn't realize I had to lay down for this. I'm about to tell you things I've never told anyone. Not even my family. I was an immigrant. And when I came in 1973, I was nine. My older brother was almost 11. And this was a foreign country and there was nothing easy about that. And we also grew up with really, really terrific parents. But we weren't wealthy. And so they worked hard, they work hard today. And so they passed along a lot of life lessons by working hard. I had all kinds of jobs and we went
to a school that included a lot of chores. -It's in Kentucky? -Yeah, Kentucky. Oneida Baptist Institute. I don't think it's the same as 'MIT', that 'I' is not the same. It's the same word, but it's different. It's a different type of institute. But my institute required you to go to school, and it was a dormitory, and so there were a lot of chores. I was the youngest kid in school, and so all of the other kids got the hard work. They had to work in the tobacco farm, and I got the easy job. I was
nine years old, and so after they left, I had to clean all the bathrooms. I never felt that I got the easy job, because what they left behind was... You can't unsee that kind of stuff. But that was my job and so I did it delightfully. And then I had plenty of other jobs, and Denny's was one of them. And I started out as a dishwasher and became a busboy and became a waiter. And I loved every one of them. I loved every one of them. Somehow, I've always found... I want to say joy, but
that's not quite right. Just everything that I was doing, I wanted to do the best I could. And maybe that was kind of ingrained from the very beginning, But I was definitely the best bathroom cleaner the world's ever seen. I'm sure of it. Yeah. -So if we fast forward just a little bit, to the NVIDIA of today, how large is your leadership team? -How large is...? -Your leadership team. -NVIDIA's leadership team is 60+ people. -And they all report to you? -Yeah, they all report to me. -60 direct reports? -60 direct reports. Yeah. -Which is not
conventionally considered a best practice. I agree that the best practice... -I'm certain that's the best practice. It's not conventional, but I am certain it's the best practice. And by the end of this, I'm going to convince all of you to have 60 people on your direct reports. -The floor is yours. -The reason... first of all, the reason is because the layer of hierarchy in your company really matters. Information really matters. I believe that your contribution to... the work should not be based on the privileged access to information. I don't do one-on-ones and I don't... My
staff is quite large, and almost everything that I say, I say to everybody all at the same time. And the reason for that is because I don't really believe there's any information that I operate on, that somehow only one or two people should hear about. And these are the challenges of the company or this is the problem I'm trying to solve, or this is the direction we're trying to go into. These are the new endeavors. This isn't working. That's working well. And so all of this type of information everybody should be able to hear it.
I love that everybody is working off of the same song sheet. I love that there is no privileged access to information. I love that we're able to all contribute to solving a problem. And when you have 60 people in a room and oftentimes... Well my staff meetings are once every other week, and it's all based on issues, whatever issues we have. Everybody's there working on it at the same time, everybody heard the reasoning of the problem. Everybody heard the reasoning of the solution. Everybody heard everything. And so that empowers people. I believe that when you
give everybody equal access to information, it empowers people. And so that's number one, empowering. Number two, if the CEO's direct staff is 60 people, the number of layers you've removed in a company is probably something like seven, depending on how it is. -60 at every layer, or only 60? As in, if I'm one of the fortunate 60, do I also have 60 direct reports? -No. I also don't think that that's scalable downward. And the reason for that is because you need more and more supervision, depending on certain levels. And at the E-Staff level, if you're
so unfortunate to be serving on NVIDIA's E-Staff, it's very unlikely you need a lot of managerial. -And so I rarely find myself having to stand up for conventional wisdom. But if I were to kind of steel mine the other side, I'd say, well, one-on-ones are where you provide coaching, where you maybe talk through goals together, personal goals, career advancement, what have you. Where maybe you give feedback on something that you see somebody systematically not doing so well and so forth. So there's all these things that kind of one is, again, conventionally supposed to do in
the one-on-one. Do you not do those things or do you do them in a different way? -Really good question. I do it right there. I do it right there. I give you feedback right there in front of everybody. And in fact, this is really a big deal. First of all, feedback is learning. Feedback is learning. For what reason are you the only person who should learn this? Now you created the conditions Because of some mistake that you made or silliness that you brought upon yourself. We should all learn from that opportunity. So you created the
conditions, but we should all learn from it. Does that make sense? And so for me to explain to you why that doesn't make sense or how I differ from it-- half the time I'm not right-- but for me to reason through it in front of everybody helps everybody learn how to reason through it. And so the issue-- the problem I have with one-on-ones and taking feedback aside is you deprive a whole bunch of people that same learning. Learning from mistakes, other people's mistakes, Is the best way to learn. Why learn from your own mistakes? Why
learn from your own embarrassment? You got to learn from other people's embarrassment. That's why we have case studies. Isn't that right? We're trying to read from other people's disasters, other people's tragedies. Nothing makes us happier than that. -Have you succeeded in getting other leaders at NVIDIA to adopt this practice, or is that difficult? -I give people the opportunity to decide for themselves, but I really discourage one-on-ones. I really discourage one-on-ones. Nothing is worse than the idea that somebody says, "Oh, Jensen wants us to do this." Why does that have to be said to anybody? Everybody
should know. Or somebody said, "That E-Staff said that." Nothing drives me nuttier than that. -You once told me that you really didn't like firing people and very seldom did it. Can you elaborate on that? -Well, I'd rather improve you than give up on you. When you fire somebody, you're kind of saying-- Well, a lot of people say, "It wasn't your fault" or "I made the wrong choice" or-- There are very few jobs-- Look, I used to clean bathrooms, And now I'm a CEO of a company, I think you could learn it. I'm pretty certain you
can learn this. And there are a lot of things in life that I believe you can learn. And you just have to be given the opportunity to learn it. I had the benefit of watching a lot of smart people do a lot of things. I'm surrounded by 60 people. They're doing smart things all the time. And they probably don't realize it, but I'm learning constantly from every single one of them. And so, I don't like giving up on people because I think they could improve. And so there's a-- it's kind of tongue-in-cheek-- but people know
that I'd rather torture them into greatness. -That was the phrase that I was hoping to uncover. Yeah, I remember you mentioned that. -Yeah. So I'd rather torture you into greatness because I believe in you. And I think coaches that really believe in their team torture them into greatness. And oftentimes they're so close. Don't give up, they're so close. Greatness kind of comes all of a sudden, one day he's like, "I got it". Do you know what I'm saying? That feeling that you didn't get it yesterday and all of a sudden one day something clicked. "Oh,
I got it." Could you imagine you gave up just that moment right before you got it? I don't want you to give up on that, so I'll just keep torturing you. -How's your work-life balance? -Well, it depends on who you ask. I think my work-life balance is really great. It's really great. I work as much as I can. I feel like he's judging me. I'm older than you. I have more wisdom than you. So what I... -These are all the highlights from our conversations that I think more people should get to hear, so... -Well, I
work from the moment I wake up to the moment I go to bed, and I work seven days a week. When I'm not working, I'm thinking about working. And when I'm working, I'm working, and so... And I sit through movies, but I don't remember them because I'm thinking about work. And so that's-- But my work is not, as you know, it's not working as in, there's this problem and you're trying to solve this problem. You're thinking about what the company can be, and are there things that we could do even better? Or sometimes it's just
trying to solve a problem, you know? But sometimes you're imagining the future. And, boy, if we did this and that. It's working, you're fantasizing, you're dreaming, right. I mean, that's incredible. -Well, so yeah, to concretize this a little bit and we will get to talking about AI, which I hear is a thing these days. -It's a thing. -Yeah, officially a thing. But to concrete concretize this a bit like, what does a day in Jensen's life look like? -Well, I used to wake up at five. These days, I wake up at six because of my dogs.
And the reason why six, is somehow we decided that six o'clock is when they should wake up. And I don't know what it is. I don't mind waking anybody up but I feel guilty when I wake the puppies up. And it actually burdens me. So, I don't want to move. They pick up on any vibration in the house and it wakes them up and so we kind of stay in bed. And I just read in bed until six o'clock and it's time-- -But you're thinking about GPUs? -Oh yeah. Yeah, yeah, sure. I'm obsessed about GPUs.
I mean, what can you do? I'm constantly... No. I'm just... -And then the day is all, I guess, group meetings. Because it can't be one-on-one meetings. -Yeah. I get my work done before I go to work and then when I get to work... -And how many meetings in a typical day? -Pretty much all day long. And so, I select the meetings that are really important to me. I try not to have regular meetings, regular operational meetings. Because I've got amazing people in the company who are doing regular operational meetings. And so we're pinch hitters, CEOs
are pinch hitters. We should be working on the things that nobody else can or nobody else is. So you're jumping in to projects that are stuck or off track. -That's right. -Or new ideas. -Wherever we can move the needle. No reporting meetings. I hate reporting meetings. They don't have to report to me. I just have problem meetings. And so problem meetings, or idea meetings, or brainstorming meetings, or creation meetings, or whatever it is, those are the meetings I go to. And so usually I call them, I try really hard not to have Outlook manage my
life. And so we purposefully decide what kind of things we want to do, we want to work on. And so I try to live a life of purpose, and I manage my time accordingly. Yeah. -You used a phrase, once, $0 billion markets, that $0 billion markets are your favorite markets. -Yeah. -What do you mean? -If you take a step back, our purpose, almost all of our purposes should be to go and do something that has never been done before. That is insanely hard to do. That if you achieve it, could make a real contribution. I
know your company does that. I try to do that. And if that's the case, it hasn't been done before, it's incredibly hard to do, It's probably-- and it's never been done before-- that market is probably $0 billion in size. Because it has never been done before. I'd rather be a market maker, market creator, than a market taker. To create something new that never existed before versus thinking share I don't love thinking about share. I don't like the concept of share. And the reason for that is because if you think about it in the big picture,
Stripe existed out of thin air, you vaporized. You created something out of vapor. It wasn't as if there was another-- something else. And so I'd like to think that we can come up with something that is $0 billion. A $0 billion market is a good way to cause the company to think about how to go create something for the first time. -So our mission is to grow the GDP of the internet, and the GDP of the internet-- the clause in that usually gets most of the attention. But I think the most important part is just
the verb "grow". Because, to your point, we shouldn't be thinking about, well, which are the transactions that are already happening or which are the businesses that already exist, we should be thinking about, which are the transactions that don't exist and which are the businesses that don't exist. The GDP of the world is around $100 trillion, but it doesn't have to be $100 trillion. It could be $200 trillion or $1000 trillion. -That's exactly right. And most of the value we're going to create over the next several decades are likely not limited by physical things. And so
this is a pretty extraordinary time. -And so with this concept of $0 billion markets, if I'm at NVIDIA, am I coming to you with some proposal for some project and maybe there's several billion dollars of CapEx involved or it's a many-year pursuit or something. And there are no customers for it today, there's no demand that I can demonstrate for it. And you guys are just making a gut call to say that, "Yes nobody is doing this today. We think they could, we think they should. And therefore we're going to pursue it." -Really close. Yeah. It's
kind of like that. And it's a gut call in the sense that your intuition says something as a starting thesis, But then you have to reason through it. And the reasoning of it is much, much more important to me than a spreadsheet. I hate spreadsheets because you can make spreadsheets do whatever you want. You can make any chart you want out of a spreadsheet. You just got to type in some numbers. And so I don't love spreadsheets for that reason. I love words for that reason, words are reasoning. Tell me, how did you reason through
this? What's our intuition? Why do we believe that matters? Why do we think it's hard? I like hard things because it takes a long time to do, and if it takes a long time to do... a lot of people who are less committed probably won't do it. If it's really, really hard to do, it takes a long time to do, it takes a really resilient and a really dedicated, really committed person to go after it. And if it also takes a long time to do, you can kind of flounder around for a couple of years
and nobody notices. And so I could be incompetent for several years and everybody goes, "Well, who saw it?" -And where did Cuda come from? -Cuda came originally from two ideas. One is called... I hate to get technical, but we created, we pioneered, this idea called accelerated computing. Accelerated computing is like an IO device, Something that you sit on PCI Express, if anybody's in the computer business, an IO device that allows the application to interact with that IO device in such a way as to accelerate parts of the application. And UDA was an invention in 1993,
and it's really a profound invention, allows the software programmer to directly program an IO device, write an application directly to the IO device, because the IO device is virtualized and... it's architecturally compatible across multiple generations. It's, anyways, we invented this idea called accelerated computing, and that was-- we call it unified driver architecture for whatever reason. And then several years later, We thought we could make our GPUs more programable to high level programing languages. And we invented this idea called CG, C for graphics, okay. C for graphics processors. And that opened up some really exciting opportunities.
And we thought, you know what this is going to work. But CG, the programing model, wasn't exactly right. And so we invented Cuda which is compute, with you know... So anyways that's how. It's a horrible story frankly. Anyways, we invented this idea called accelerated computing. We pioneered this approach. I guess the real question is was it a smash hit overnight? -No, it was an incredible disaster overnight. And it kind of went like this. -This is one of your $0 billion markets you went after. -Yeah. -And it was a disaster. -Yeah. Because it was a $0
billion we went after. But it cost so much to go after that $0 billion market, it actually crushed the $1 billion market we were enjoying. And the reason for that is because Cuda added a ton of cost into our chips, but there were no applications. And if there are no applications, customers don't value the product and they won't pay you a premium for it. And if people aren't willing to pay you for it, but your cost went up, Then your gross margins get crushed. Our market cap was low and it went down to really low.
I think our market cap went down to $1 billion or something like that. I wish I had bought it, but anyways. -Okay. And so therefore you immediately canceled Cuda and went back to the old strategy. -No, no, I believed in Cuda, because you reasoned about it. You reason about it. Look, we really believed that accelerated computing was going to be able to solve problems that normal computers couldn't. And if we wanted to extend the architecture to be much more general-purpose, we had to make that sacrifice. And so I deeply believed in the mission of our
company, I deeply believed in its opportunities. -And so were analysts-- -I deeply, deeply believe that people were wrong. They just didn't appreciate what we built. I deeply believed it. -And so weren't analysts and the board and employees-- you've torpedoed this existing revenue stream. You have this hyped thing that you're selling a lofty dream around, that nobody seems to actually want. The business is really suffering, Talk us through that. You believed. -Well, it goes something like this, "Oh, gosh, they're so dumb." Something like that-- denial. No, I'm just kidding. You go back to what you believe.
And if you believe something-- -Did the board put pressure on you during this? -They... I start every conversation with what I deeply believed and they believed it because they saw me deeply believe it. And I reasoned about it. It wasn't like it was a spreadsheet, and therefore you've got to believe the spreadsheet. They had to believe the reasoning, the words. -How long did it take it to start working? -Probably ten years, yeah. It wasn't that long. Ten years. It comes and goes. -Ten years. -Less than a third of your tenure. -Yeah, it comes and goes.
I barely remembered it. The suffering, I barely remembered it. -Could NVIDIA be as successful in AI without Cuda? -No, impossible. It is potentially one of the most important inventions in modern computing. We invented this idea called accelerated computing, and the idea is so simple but deeply profound. It says the vast majority... a small percentage of the code of programs occupies, consumes, 99.999% of the runtime. And this is true for a lot of very important applications. And that small little kernel, or some several kernels, can be accelerated. It's not all just parallel processing. It's not as
simple as that. But the idea is that we can take that kernel, that piece of software, that part of the software, and accelerate the living daylights out of it. And today, when Moore's Law has run its course and CPU scaling is basically stopped, and if we don't accelerate every software, you're going to see extraordinary computation inflation. Because the amount of computation the world does is doubling every year still, And yet if CPUs and general-purpose computers are not increasing in performance because it stopped, then what's your alternative? Your cost of computing is going to keep going
up exponentially. And so the time has come for us to do that. -So everyone here runs a business and... -Accelerate everything! -And you heard it here first. And probably everyone has some version of Cuda or a thing that they think really makes sense for the sector or makes sense for their technology or what have you, but where the market doesn't see it yet. Do you think it's possible to extract any kind of generalizable principles around when you should really doggedly trust that vision, and when perhaps it's worth reconsidering in a fashion that we could extrapolate
from, in the case of Cuda and other "Cudas" that have existed over the course of NVIDIA's history? -Yeah. The question is determination and commitment versus stubbornness and that line is fuzzy. I gut-checked against my core beliefs every day, I still do. And you gut-check against it. The first principles by which you reason about your strategies, the first principles by which you reason about your strategies, Those first principles are easy to remember, and it's not a long list. And now the question is, are those principles, did they change in some fundamental way? Are external conditions such
that they no longer matter as much as before? Did somebody else solve the problem, and therefore that problem has now disappeared? Is it that there will never be any need? You gut-check it, right, constantly. And to the extent that-- that's number one, gut check. You have to first of all be really careful to distill down the first principle, instead of, "I want something", that's stubbornness. You can't reason about it. I just want it. We're not five-year-olds, right? And so you've got to reason about it, number one. Number two, you have to be clever. The fact
of the matter is there are a lot of new companies being created here. It's amazing how many great companies are in the audience and young companies in the audience. You have to be clever. And so we found ways to monetize, even in a small way, Cuda. And so we found-- we looked everywhere for applications. We found an application with CT reconstruction. We found an application with seismic processing. We found another application with molecular dynamics, and so we're constantly looking for applications. They didn't make it a home run but it sustained us just enough. And bought
us time for it to really happen. -Okay. So let's talk about about AI. I'm going to just do some math to ground things here. Let's just say that the total, sort of compute capacity of all GPUs in the world today is X. What do you think... What multiple of X will we be at in five years? -First of all, you know that I'm going to regret saying this. And this is be... I'm a public company, you crazy person. How nice is it to be private? -Safe to say considerably more. -Well, let's reason about it, shall
we? Okay. So let's let's reason about it. Let's reason our way through okay. So first of all, it goes like this. The world has installed about a trillion dollars worth of data centers. Those trillion dollars worth of data centers use general-purpose computing. General-purpose computing has run its course. We cannot continue to process that way. And so the world is going to accelerate everything, data processing, you name it. And so we're going to accelerate everything. When we accelerate everything, every single data center, every single computer will be an accelerated server. Well, there's about a trillion dollars
worth of computers if we don't grow at all over the next four years that we have to replace. Four years, six years, pick your number of years. But if the computer industry continues to grow at some 20% or so, we'll probably have to replace, over the course of next, pick your number of years, about $2 trillion worth of computers with accelerated computing. So just make that GPUs, okay? That's number one. And this is the second part. This is the reason why, all of you, Stripe, you're onto something just absolutely monumental. This idea called, and you've
heard me say an industrial revolution, Let me tell you why. We are producing something for the very first time that has never been produced before. And we're producing it in extremely high volume. And the production of this "thing" requires a new instrument that never existed before. It's a GPU. And the thing that we're producing for the very first time, for the mathematicians and all the computer scientists in the room, for all of you, you know that we're producing tokens. We're producing floating point numbers at high volume for the first time in history And the floating
point numbers have value. The reason why they have value is because it's intelligence. It's artificial intelligence. You can take these floating point numbers. You reformulate it in such a way that it turns into English, French, proteins, chemicals, graphics, images, videos, robotic articulation, steering wheel articulation. We're producing tokens at extraordinary scale. Now we've discovered a way through all of the work that we do with artificial intelligence, to produce tokens of almost any kind. So now, the world is going to produce an enormous amount of tokens. Now these tokens are going to be produced in new types
of data centers. We call them AI factories. Back in the last industrial revolution, water comes into a machine, you light the water on fire, right? Turn it into steam and then it turns into electrons. Atoms come in, electrons go out. In this new industrial revolution, electrons come in and floating point numbers come out. And just like the last industrial revolution, nobody understood why electricity is so valuable and is now sold, marketed kilowatt hours per dollar. And so now we have a million tokens per dollar. And so that same logic Is as incomprehensible to a lot
of people as the last industrial revolution, but it's going to be completely normal in the next ten years. These tokens are going to create new products, new services, enhance productivity on a whole slew of industries. $100 trillion worth of industries on top of us. And so this industry is going to be gigantic. And in order to monetize that, transact that, you're going to need Stripe. I have to tell you, this is one of my favorite companies. The first time I met Patrick, he had to explain Stripe to me. First of all, it was so complicated.
-We tried to refine the descriptions over time. -First of all, you're in a complicated business no matter what. But nonetheless, I was so inspired by it. Incredible what you guys have built. -We're going to get you migrated to Stripe Billing now that we have a usage-based billing? -I wish I had a business that required billing. -I think your public filing suggests you're doing a lot of billing. We'll follow up on it. All right. -It's only ten transactions. Just so you know. Your economics serving us is like nothing, it's like ten transactions. -We'd happily take the
2.9% . But anyway, we can discuss that separately. -Done! -You can't say that. You're a public company. Thinking about these token factories. I feel like a big question right now is whether the models saturate in the sense that, we demoed the Sigma assistant on stage earlier and you can write some natural language, and we convert that to SQL. And going from maybe a 7 billion parameter model to a 70 billion parameter model or something like that, there might be a significant kind of, consequential improvement in query accuracy for the user for the typical kinds of
queries that people tend to construct. But maybe going to a model that's ten x larger than that is sort of unnecessary. Like at some point you get to good enough, You can reliably convert the natural language to SQL. I think there's a question of, for the use cases for which LLM's are being deployed, what does that saturation curve look like, and for how many use cases does one need a trillion parameter model or a 10 trillion parameter model? Or do we simply reach a point where, some number that is, say, less than 100 billion, is
sufficient? Do you have any point of view on that, or is that even a reasonable way to look at the question in the first place? -Okay, let's break it down. Let's reason it out. -In public. -Almost everything, every question I get, okay, let's break it down, let's reason it out. So, let's start with an example. In 2012, AlexNet was a computer vision. ImageNet, image recognition 82%, or something like that, accuracy. Over the next... almost not quite ten years, I think it was like seven years, every single year, the accuracy error reduced in half. Every year
the error reduced in half or otherwise known as Moore's law. So you double the performance, you double the accuracy, and you double its believability every single year. Over the course of seven years, it's now superhuman. Same thing with speech recognition, same thing with natural language understanding. We want to know, we want to believe, not know, we want to believe that the answer that's being predicted to us is accurate. We want to believe that. And so the industry is going to chase that believability or that accuracy, and double its accuracy 2x every year. I believe that's
going to be the same thing with natural language understanding. And of course the problem space is a lot more complicated, but I have every certainty that we're going to double its accuracy every single year To the point where it is so accurate. And we've we've largely tested across many of your examples, when you interact with it that you go, "You know what? This is really, really good." I believe the answer that it's producing for me, that condition is very important. The second thing is this. Today's language models, today's AI and everything that we've shown are
one shot. And yet you and I both know that there are many things that we think about that are not one shot. You have to iterate. And so how do you come up-- How do you reason about a plan? How do you come up with a strategy to solve a problem? Maybe you need to use tools. Maybe you have to look up some proprietary data. Maybe you have to do some research, in fact. Maybe you have to ask another agent. Maybe you have another ask another AI. Maybe you have to be human in a loop.
Ask a human. Triggering events, send an email to somebody or text to somebody. Get a response before you can move on to the next step of that of that plan. And so a large language model has to iterate and think of a plan. That's not a one shot thing. And once it comes up with a plan as it traverses that graph, there's a whole bunch of language models that are going to get instantiated and initiated. And so I think your future models are going to iterate. And so instead of a one shot model, it's going
to be a planning model with a whole bunch of other models around it that are particularly good at particular skills. And so I think we have a long ways to go. -Meta garnered a lot of attention last week for the release of Llama 3, which seems to be the most impressive open-source model thus far. Any thoughts on open-source models? -If you ask me, what are the top most important events in the last couple of years? I would tell you, of course, ChatGPT, reinforcement learning, human feedback, Grounding it to human values and having the technology necessary
to do that. Obviously a breakthrough, and democratized computing. It made it possible for everybody to be a programmer. Everybody is now doing amazing things with it. ChatGPT, the work that OpenAI did, Greg and Sam and the team, really proud of them. The second thing that I would say that is just as important, I would say, is Llama, not Llama 1, but Llama 2. Llama 2 activated just about every industry to jump into working on generative AI. And it opened the floodgates of every industry being able to access this technology, Health care, financial services, you name
it, manufacturing, you name it, customer service, retail, all kinds. I think Llama 2 and Llama 3, because it's open-sourced, it engaged research, it engaged startups, engaged industry. It made generative AI accessible. I think that's a very big deal. And so I think ChatGPT democratized computing. I think Llama democratized generative AI. Does that make sense? And I think without it, it's very hard to have activated all of the research on safety and all of the different ways of chains of thoughts. And all the reasoning technology that's now being developed And all the reinforcement learning stuff. And
that stuff would have been very hard to have activated without Llama. -Dario Amodei was on Ezra Klein's podcast two weeks ago. And he, as many others have, many others in particular who are involved with Frontier Labs, was predicting AGI in the relatively near term. Conceivably the next couple of years, years like 2027, and so on, are frequently thrown around. Thoughts? -Depending on how you define AGI. Now, first of all, as an engineer, You know that we can only solve a problem, ultimately, if you can measure it. And so you have to express the problem statement,
the mission, somehow, in some some measurable way. If you told me that AGI is the list of benchmarks we currently use, there are math tests, and English comprehension tests, and reasoning tests, and you got medical exams, and bars, and you make your list of all of the tests that you want. It doesn't matter what it is, just make your list. If you make your list, I am certain we will achieve excellent results in a very nominal amount of time. And if that's the definition of AGI, I'll make a guess, It's probably definitely within the next
five years. And so all of the tests that we currently measure these models with, they're improving their accuracy or their error rate is reducing in half every six months. And so there's no reason why we shouldn't expect it all to be superhuman pretty soon. -So again, everyone in this audience-- -That doesn't-- That doesn't meet the standard. Just be clear. That doesn't meet the standard of a normal person thinking it's AGI. Does that make sense? An on-the-street-person, "Hey AGI." That's probably not what they're thinking, what I defined it as. The way I defined it is simply
an engineering way of defining it So that you can answer that question. The second way of answering the question is when can you achieve AGI in an undefined way? If it's undefinable, then how long would you know? How long would it take? Undefinable. -And so everyone in this audience again runs a business. And a practical question they/we all face is how do you know if you're, in the face of the kinds of changes you just depicted, how does one know, how can one know whether one is responding appropriately, sufficiently, in the right ways, etc.? Any
advice? -If you're not engaging AI actively and aggressively, you're doing it wrong. You're not going to lose your job to AI. You're going to lose your job to somebody who uses AI. Your company is not going to go out of business because of AI. Your company is going to go out of business because another company used AI. There's no question about that. And so you have to engage AI as quickly as possible. You have to engage AI as quickly as possible, so that you could do things that you think cost too much to do. For
example, if the marginal cost of intelligence was practically zero. There are a lot of things that you would do now that you wouldn't have done otherwise. And so notice how often we do search And these days notice how often we ask questions. Any random question, I'll be asking Perplexity. Right. And so why not. -Or, just gave a talk here at Sessions. -Okay, I love using it. And even if I know the answer, I'll just ask it anyways, just to see what it comes up with. And so, I think we want that to happen. We want
the marginal cost of these type of activities to be as low as possible so that you use it in abundance. Second, if you could use AI to be productive-- You know that productive companies leads to higher earnings. Higher earnings leads to more employment. More employment leads to more social growth. And so, there's a lot of reasons to want to drive productivity into companies. -And apart from just changing your manufacturing plans and your CapEx plans, how has AI changed how NVIDIA works internally? -We were one of the first technology companies to invest in our own AI
supercomputers. We can't design a chip anymore without AI. At night, our AIs are exploring design spaces, vast and wide that we would never do ourselves because it costs too much money to explore it. And so our chips are so much better. Because of an AI, we could reduce the amount of energy used for our chips as higher performance. Our software. We can't write software without AI anymore. We have to explore all the-- the design space of optimizing compilers is too large. We use AIs to file bugs. So our bugs database actually tells you what's wrong
with the code, who's likely involved and activates that person to go fix it. So, I think I want everybody, every organization in our company to use AI very aggressively. I want to turn NVIDIA into a one giant AI. How great would that be? And then I'll have work life balance. Are there any favorite examples you've heard of businesses in maybe some kind of unexpected sector, some unexpected use case where you feel that they can serve as a poster child for some of the dynamics you're describing, where they've really realized some of this opportunity? -The biggest
surprise of AI that shouldn't be a surprise for a lot of people, is that when we say it's a large language model, the word language doesn't mean human language only, and it doesn't mean English only, or French only, or Irish only, or that's a whole different language. But... Is there a large language model for Irish? -I've tried it. -It works? -Yeah, it works well. John and I spent most of our education in Ireland, being taught in Irish. And so, these models are some of the first people I've had the chance to have a dialog with
<i>as Gaeilge</i> [in Irish]. -Very surprising. -In many years. Yeah. They do well, and actually I've been-- Have you played with Suno? -Suno? -Suno is an app for creating music, kind of synthetic music. -Okay. -And I've been enjoying creating... -Irish music. -I of course tested it on that. And Celtic dubstep is a thing that it can do. -Fantastic. Okay. Makes sense. Like if it could do that, then of course it could learn the language of life. Of course they can learn the-- And if a language model can understand sound which is a sequence time series, it's
a sequence. Why can't it learn robotics articulation which is a sequence. You just have to figure out how to tokenize it. And so the idea that all of a sudden, "Oh hey, look, listen, I could also learn SQL, I could learn ABAP, I can learn Lightning, I can learn all these proprietary languages, I can learn Verilog, I can learn, right. So all of a sudden you realized, hang on a second, I can put a copilot on top of every tool on the planet. -And to this point, NVIDIA being one big AI, is the future one
of 100,000 models or 100 million models, or is the future one of one model? And they're just a model that does all the things. -I think that it would be great to have super models that help you reason about things, in general. But, for us, for all companies that have very specific domain-specific expertise, we're going to have to train our own models. And the reason for that is because we have a proprietary language. That difference between 99% and 99.3% is the difference between life and death for us. And so it's too valuable to us. No
different than fraud detection for you, it's too important to you. -That's been exactly our experience. -Yeah. It's too important to you. However good the general model is, you're going to want to take that and fine-tune it, right, improve it into perfection because it's just too important to you. Yeah. -And so we're going to shortly run out of time here. And there's a whole bunch of questions I haven't gotten to yet. I've exercised poor discipline on the time management front. So there's a bunch that I think I was told I definitely had to ask you, but
there's a couple that I really wanted to ask and it's only us up here. So, Lisa Sue is your first cousin once removed? -Yes. Yeah. She's terrific. She's amazing. -And AMD is now-- -She's the CEO of AMD, by the way. -Yeah, and AMD is now one of your competitors in the GPU space. -No. We're family. -Okay. -We're all in the industry. -One of your partners in the industry. -Yeah, yeah. We buy from AMD. -What's going on in the water? How did we end up with two of the, arguably the two most important, GPU companies being
run by close relatives? What's going on? -You got to keep it close to the family. No, I have no idea how it happened. We didn't grow up together. -And that makes it even more interesting, right? -Yeah, yeah. We didn't even know each other until she was at IBM. And her career is incredible. And she's really quite extraordinary. -I think this question requires further study. So you've been operating in Silicon Valley since the early 90s. -Yes. -How has Silicon Valley culture changed in that time? -Oh, wow. I haven't thought about this in a long time. I
guess in a lot of ways probably-- Okay, here's one. When I first started NVIDIA, I was 29 years old, and... I was 29 years old with acne and, you go talk to your talk to your-- go recruit law firms and VCs and I got a big zit on my forehead and, I don't have one today, so I feel comfortable talking about it, but it could it could happen. And so anyways, you feel rather insecure because most of CEOs back then wore suits and they're quite accomplished. And they sound like adults and they use big words
and they talk about business and things like that. And, so, when you're young, you feel rather intimidated. You're surrounded by a bunch of adults. Well, now if you don't have acne, I don't think you deserve to start a company. That's one big difference. Acne. The takeaway from Jensen's speech. What it means is really we've enabled younger people to be extraordinary. I think the young generation of CEOs, the type of things that you guys know at such a such a young age, is really quite extraordinary. I mean, it took me decades to learn it. -Last question.
-That was a compliment. See how quickly he changed the... I wasn't saying you have acne. I was just saying you were smart. -NVIDIA has a market cap of roughly $2 trillion. And you're now within spitting distance of Apple and Microsoft. And I just checked and they have 220,000 and 160,000 employees, respectively. NVIDIA has 28,000 employees, so, less than a fifth of the smaller of the two there. And then you just said when we were chatting backstage, and I jotted this down, "You can achieve operational excellence through process, But craft can only be achieved with tenure."
And so NVIDIA is considerably smaller than any of the other giants. And you seem to think that tenure really matters. And I guess that craft really matters. Do you want to say a little bit more there? -I think extraordinary things-- I think a lot of good things could be made-- good things are made with operational excellence. But you can't make extraordinary things through just operational excellence. And the reason for that is because a lot of the great things in your body of work and the products that you make, the company you created, the organizations you've
nurtured, it takes loving care. You can't even put it in words. How do you put love and care in an email? And for people to go, "Oh, I know exactly what to do." You can't put that in a business process, love and care. -Is Love and Care an NVIDIA catchphrase? -Well I use Love fairly abundantly, and Care I use abundantly. -At Stripe we talk a lot about Craft and Beauty. -Yeah, right. You have to use these words because in a lot of ways there are no other words to describe it. You can't put in numbers,
you can't write it in the product specification. The product specification says, "I want you to build something great that's incredibly beautiful. That's in great, great craft." You can't specify these things. -But I'm sure there's people at Stripe who think Patrick's always yammering on about Craft and Beauty. -And it's this kind of-- -I never yammer. I just want to let you know that. I don't even know what that sounds like. -Okay. Well, yeah. -Yammering on. Go ahead, go ahead. Yeah. -You're more lucid than I am. I just babble, but, hey, "So Patrick is always going on
about this Craft and Beauty stuff, and wants things to have this particular ineffable character, but it doesn't directly serve some customer need and so forth." Like customers aren't coming to us and saying, "I want the product to be more beautiful." They're saying, "I want it to feature X or feature Y." And yet we believe that the Craft and Beauty really matters. Sounds like you're getting at something similar. Why do you think it matters? -Actually your customers, even though they didn't say it, they might not have the words to say it, but when they experience it,
they know it. There's no question. Look... Stripe's work has beauty, has elegance, has simplicity. Simplicity is not simple, as you guys know. Simplicity and simple are not... not the same thing. And it has elegance and it solves the problem, but just enough, it burdens you, but not too much, And so that balance is hard to find. And you can't specify that, you just feel your way there. And when you have a team that is with you, that feels the way they're together, in a lot of ways we've codified, we've encoded, the magic of the company
in a way that no words can describe. And you don't want to lose that. You don't want to lose that. You want to take that and take it to the next level next time. And so I don't want to reset. I don't like working with new people for that reason. Because I've encoded, I've embodied, I've deposited so much pain, suffering, joy, knowledge. Right. All that experience, life experience. You've encoded it in all the people that you've worked with. You want to carry it on. You want to take it to the next level. And that's really
the reason why I really, deeply believe in tenure. And because of that, small teams could do great things. And NVIDIA is kind of a small team with 28,000 people, people think we punch well above our weight because of that reason. And so it's amazing what you guys have done and how incredibly small you are, 7,000 people supporting $1 trillion worth of ecosystem and industry and economy and who knows how far you guys can go. So I'm very proud of you. -Jensen. Thank you.
