WEBVTT - NVIDIA: At the Heart of the AI Boom 0:00:15.356 --> 0:00:15.796 Pushkin. 0:00:20.236 --> 0:00:21.836 Just a quick note at the start of the show. 0:00:21.916 --> 0:00:24.396 We are very eager to hear what you think of 0:00:24.396 --> 0:00:26.796 the show and how we can make it better, how 0:00:26.836 --> 0:00:29.316 we can make a show that you like more than 0:00:29.476 --> 0:00:30.236 the show we're making. 0:00:30.236 --> 0:00:30.436 Now. 0:00:30.516 --> 0:00:32.236 Tell us what we should do, more of what we 0:00:32.276 --> 0:00:35.596 should do, less of what we should do, wildly differently. 0:00:35.956 --> 0:00:41.196 Email us at problem at pushkin dot fm. Again problem 0:00:41.436 --> 0:00:45.916 at pushkin dot fm. I'm looking forward to reading every email. 0:00:47.476 --> 0:00:53.716 Artificial intelligence feels very abstract, feels very ephemeral, feels more 0:00:53.796 --> 0:00:57.556 like an idea than a thing. But there is in 0:00:57.596 --> 0:01:01.956 fact a thing there. AI is rooted in the physical world. 0:01:02.476 --> 0:01:05.996 The thing or things are these little pieces of silicon 0:01:06.156 --> 0:01:12.716 and metal called graphics processing units GPUs. These GPUs are expensive. 0:01:12.996 --> 0:01:15.836 They cost thousands of dollars each. If you want to 0:01:15.836 --> 0:01:17.876 build a state of the art AI model, you have 0:01:17.916 --> 0:01:22.116 to buy tens of thousands of these GPUs, and most 0:01:22.156 --> 0:01:26.236 of the GPUs come from a single company, in Vidia, 0:01:26.676 --> 0:01:30.236 which is why in just the past few years Nvidia 0:01:30.316 --> 0:01:34.556 has become one of the most valuable, most important companies. 0:01:34.236 --> 0:01:34.796 In the world. 0:01:40.996 --> 0:01:43.796 I'm Jacob Goldstein, and this is what's your problem. My 0:01:43.876 --> 0:01:47.236 guest today is take him. He's a staff writer at 0:01:47.276 --> 0:01:50.116 Baron's and the author of a new book about in Nvidia. 0:01:50.156 --> 0:01:54.156 The books called The Nvidia Way. The book, of course, 0:01:54.196 --> 0:01:57.076 points up being a lot about Jensen Wong, who co 0:01:57.156 --> 0:02:00.596 founded the company back in nineteen ninety three and who 0:02:00.636 --> 0:02:04.716 is still the CEO and who is really a wild 0:02:04.836 --> 0:02:08.436 kind of terrifying, brilliant figure. And Tay and I talk 0:02:08.516 --> 0:02:12.076 a lot about Jensen later in the interview, but we 0:02:12.116 --> 0:02:16.236 started with this moment in two thousand and one that 0:02:16.316 --> 0:02:19.836 led from Nvidia being a company that made graphics cards 0:02:19.876 --> 0:02:23.516 that let people play video games on computers to becoming 0:02:23.556 --> 0:02:29.556 this key company at the center of the AI revolution today. 0:02:30.716 --> 0:02:33.276 So you write about this moment in two thousand and 0:02:33.356 --> 0:02:37.476 one when a researcher, right, an academic at the University 0:02:37.476 --> 0:02:41.516 of North Carolina, realizes he can basically like hack these 0:02:41.556 --> 0:02:44.316 graphics cards, right, this hardware that's just made to make 0:02:44.356 --> 0:02:48.036 pretty graphics on the computer to help in his research, 0:02:48.076 --> 0:02:50.516 which is like not at all about graphics, right, It's 0:02:50.556 --> 0:02:54.396 like modeling the weather or something. And it seems like 0:02:54.636 --> 0:02:58.556 this moment is a big winds up being a big 0:02:58.636 --> 0:03:02.076 turning point in the history of technology. Really tell me 0:03:02.116 --> 0:03:02.836 about that moment. 0:03:03.236 --> 0:03:05.756 The key moment was with Mark Harris, where he was 0:03:05.756 --> 0:03:09.796 a researcher at the University of North Carolina, and him 0:03:09.876 --> 0:03:12.756 and a lot a bunch of other academics realized there's 0:03:12.796 --> 0:03:16.436 all this computing power that you can hack the algorithm 0:03:16.836 --> 0:03:21.156 to use that computing power to model thermodynamics of fluids 0:03:21.196 --> 0:03:23.076 inside clouds, which was his thesis. 0:03:23.436 --> 0:03:26.556 And he realized that worked better than the CPU than 0:03:26.596 --> 0:03:28.876 just sort of doing it the traditional way through the computes. 0:03:28.996 --> 0:03:32.716 Yeah, and a GPU has all these different cores that 0:03:32.836 --> 0:03:36.156 run in parallel, and the CPU only typically has four 0:03:36.196 --> 0:03:39.556 to eight cores, and GPUs have hundreds of thousands of cores. 0:03:39.996 --> 0:03:43.356 What's a CPU better at than a GPU? Like, why 0:03:43.396 --> 0:03:45.756 would you have four to eight when you could have thousands? 0:03:46.196 --> 0:03:50.436 Well, it could run more complicated pieces of instructions and software, 0:03:50.676 --> 0:03:54.356 and GPUs typically break down into much simpler tasks. But 0:03:54.396 --> 0:03:57.476 the difference is you could run all those tasks across 0:03:57.476 --> 0:04:02.916 a thousands So for certain workloads like AI, it's so 0:04:03.036 --> 0:04:06.396 much faster than a CPU, which has to do things 0:04:06.556 --> 0:04:09.836 kind of one after another serially. And Mark Harris and 0:04:09.836 --> 0:04:14.036 all these academics started hacking into the graphics algorithms and 0:04:14.876 --> 0:04:18.876 using that power to do all this scientific high performance computing. 0:04:19.316 --> 0:04:23.236 And Mark Harris started a website and congregated all these 0:04:23.516 --> 0:04:26.676 researchers and everyone's sharing the knowledge, sharing the tricks that 0:04:26.796 --> 0:04:30.356 actually hack into this. So Nvidia sees this, they're seeing 0:04:30.436 --> 0:04:33.836 people use this to do model stock options, the model 0:04:33.916 --> 0:04:36.876 weather and saying, wow, this is actually really interesting. 0:04:37.436 --> 0:04:40.156 Yeah, that's like your dream if you're a company. Right, 0:04:40.196 --> 0:04:42.236 It's like, oh, there's all these other things you could 0:04:42.276 --> 0:04:44.196 do with this thing we're already. 0:04:43.796 --> 0:04:47.236 Making exactly so, and Jensen kind of had the foresight 0:04:47.356 --> 0:04:49.236 to see this, Wow, this is a really big deal. 0:04:49.476 --> 0:04:51.956 We should invest in this. So they wind up hiring 0:04:51.996 --> 0:04:56.156 Mark Harris and look at all the things that all 0:04:56.196 --> 0:04:59.196 these researchers are doing. And this is the beginning of 0:04:59.276 --> 0:05:03.756 development Kuda, which is a programming platform for general purpose 0:05:04.076 --> 0:05:05.036 GPU computing. 0:05:05.476 --> 0:05:07.436 And just to be clear, Kuda is like, it's a 0:05:07.436 --> 0:05:11.436 programming language, but in this case they're programming what used 0:05:11.476 --> 0:05:13.276 to be just the thing for graphics, but it turns 0:05:13.276 --> 0:05:15.396 out to be good for these other things. And that 0:05:15.956 --> 0:05:18.956 language it's in video's own language. Right, This is I 0:05:18.956 --> 0:05:20.676 feel like, going to be important in the sort of 0:05:20.716 --> 0:05:23.876 business story for why and video is so dominant today, right, 0:05:23.956 --> 0:05:27.716 Like they write this language, they own the language, and 0:05:27.756 --> 0:05:30.116 it's written just to work with their chips, is that right? 0:05:30.276 --> 0:05:33.876 Yes, no one else can use this. It's extensions on 0:05:33.876 --> 0:05:37.636 a programming language that makes parallel computing easier for programmers 0:05:37.676 --> 0:05:43.156 to make. So Jensen just invests in this, and he 0:05:43.676 --> 0:05:46.916 actually thinks longer term than the average CEO. So most 0:05:46.956 --> 0:05:49.716 CEOs looking out maybe the next quarter, maybe a year 0:05:49.836 --> 0:05:53.996 or two. Jensen thinks in five, ten, fifteen year increments. 0:05:54.636 --> 0:05:57.916 So he's always looking out, what's the next big computing shift, 0:05:57.956 --> 0:06:01.636 what's the next big technology phase. So he saw KUDA 0:06:01.756 --> 0:06:04.476 as the main thing where eventually his GPUs, you have 0:06:04.556 --> 0:06:07.676 all this latent, enormous computing power will be able to 0:06:08.076 --> 0:06:13.636 be used for science research, all these other simulation things. 0:06:14.196 --> 0:06:16.636 And he didn't give up even when Wall Street wents 0:06:16.676 --> 0:06:18.796 down on his next day, Why you wasting all this 0:06:18.916 --> 0:06:22.636 die space is crushing your gross profit margins? He said, no, 0:06:22.796 --> 0:06:25.196 this is the future computing. I'm going to invest in 0:06:25.236 --> 0:06:28.196 it and it will work out someday. 0:06:28.596 --> 0:06:30.596 And when you say wasting all this die space, like 0:06:30.676 --> 0:06:33.036 de space is like space on the chip, right, Like 0:06:33.076 --> 0:06:37.356 he's making a physical thing and to commit to Kuda 0:06:37.436 --> 0:06:40.196 to have this dedicated programming language, you actually have to 0:06:40.316 --> 0:06:43.196 give up space on the chip. So that's it's costly, right, 0:06:43.236 --> 0:06:47.636 Their margins go down. They're not optimizing for profits at 0:06:47.636 --> 0:06:48.156 that moment. 0:06:48.276 --> 0:06:51.716 It's literal circuits called Kuda cores. Their hardware circuits are 0:06:51.756 --> 0:06:56.996 optimized to run the Kuda language. So even internally executives 0:06:56.996 --> 0:07:00.996 are like, why are we spending all this money allocating 0:07:01.036 --> 0:07:03.876 dye space for something that people really aren't using that much, 0:07:03.996 --> 0:07:07.516 that isn't generating revenue. But Jensen saw this as a future. 0:07:07.716 --> 0:07:10.676 I actually kind of bring up the analogy of read 0:07:10.716 --> 0:07:14.676 Hastings and Netflix. So when he started Netflix, you know, 0:07:14.676 --> 0:07:18.196 the technology wasn't ready, consumers didn't really have broadband, but 0:07:18.276 --> 0:07:22.156 he had this intuitive sense that someday video will be 0:07:22.156 --> 0:07:24.476 streamed over the Internet and that was the future, and 0:07:24.556 --> 0:07:27.116 it makes it's so obvious now, but back then it wasn't. 0:07:27.796 --> 0:07:32.396 So Hastings positioned Netflix. He made money with DVD rentals 0:07:32.396 --> 0:07:35.436 for a while and just stayed on top of the technology, 0:07:35.516 --> 0:07:38.196 kept on investing in investing, and when the technology was 0:07:38.236 --> 0:07:42.836 good enough, that's when he really pivoted Netflix to dominate 0:07:42.836 --> 0:07:49.196 Internet stream Jensen did this multiple times with three D graphics, 0:07:49.236 --> 0:07:52.116 video game graphics. He knew that someday PC video games 0:07:52.156 --> 0:07:56.756 would be a big market programmable GPUs, Kuda and later 0:07:56.876 --> 0:08:00.676 with AI on these full stack data center AI servers. 0:08:00.916 --> 0:08:03.036 He just sees the future and is willing to keep 0:08:03.076 --> 0:08:06.756 investing even if it's five ten years out. And Kuda 0:08:06.956 --> 0:08:10.036 started in two thousand and six. I mean, things got 0:08:10.036 --> 0:08:12.836 incrementally better for the next ten years, but it didn't 0:08:12.876 --> 0:08:15.156 really take off till twenty twenty two. 0:08:16.236 --> 0:08:19.436 Till chat GPT. Basically twenty twenty two is chatchipt That's 0:08:19.476 --> 0:08:21.436 when in video goes totally bonker. 0:08:21.556 --> 0:08:23.836 And that was the power of the large language model 0:08:23.876 --> 0:08:27.036 of AI and all that stuff. And he's investing throughout 0:08:27.036 --> 0:08:28.956 this whole thing for ten fifteen years. 0:08:29.036 --> 0:08:31.316 Yeah, so there is a moment. There's a moment in 0:08:31.356 --> 0:08:33.116 the middle there. I do obviously want to get to 0:08:33.156 --> 0:08:35.156 the chat GPT moment and the present, but there is 0:08:35.196 --> 0:08:38.116 one more moment I feel like in the in the middle, 0:08:38.156 --> 0:08:40.116 that is where in video is in the center of it. 0:08:40.156 --> 0:08:43.076 And that's twenty twelve, right, So you have early you know, 0:08:43.076 --> 0:08:46.276 two thousand and one, people realize, oh, you can pack 0:08:46.316 --> 0:08:48.996 the GPU to do other things. Andybody's like, oh, that's interesting, 0:08:49.076 --> 0:08:51.276 let's build a whole programming language. So people didn't do that, 0:08:51.716 --> 0:08:56.356 And then in twenty twelve, you have this moment that 0:08:56.516 --> 0:09:01.476 really seems like the birth of modern AI. Right, this 0:09:01.516 --> 0:09:06.516 moment when when this AI model called alex net sort 0:09:06.556 --> 0:09:09.836 of emerges into the world and everybody in the AI 0:09:09.916 --> 0:09:14.116 world is like, oh my god, AI is here, right, 0:09:14.556 --> 0:09:15.676 tell me about that moment. 0:09:16.236 --> 0:09:19.836 So alex Nett was a program that was created by 0:09:19.876 --> 0:09:24.516 two researchers at the University of Toronto and they competed 0:09:24.516 --> 0:09:28.796 in this competition called image net, which basically fed images 0:09:28.836 --> 0:09:31.396 into the model, and they were able to recognize and 0:09:31.436 --> 0:09:35.436 category rise image much more effectively than any other model 0:09:35.436 --> 0:09:38.876 in the past. And the breakthrough was they used GPUs 0:09:38.916 --> 0:09:40.116 for the first time and. 0:09:40.116 --> 0:09:42.036 It was like just a few, right, Like they were 0:09:42.076 --> 0:09:44.676 grad students and they got their hands on like a 0:09:44.716 --> 0:09:47.396 few in video GPUs and had them like running on 0:09:47.476 --> 0:09:48.836 servers in the hallway or something. 0:09:48.916 --> 0:09:52.396 Yes, right, and these are video game GPUs. To be clear, 0:09:52.636 --> 0:09:56.076 this isn't some enterprise complicated. They literally went to a 0:09:56.116 --> 0:09:59.116 store and bought a bunch of video game GPUs and 0:09:59.236 --> 0:10:02.036 it turned out to be very effective where they could 0:10:02.076 --> 0:10:06.676 do the power and effectiveness of two thousand CPUs on 0:10:06.676 --> 0:10:07.916 only twelve GPUs. 0:10:08.276 --> 0:10:11.116 And to be clear or that kind of vision AI, 0:10:11.996 --> 0:10:16.436 it's machine learning. It's the same basic technology as is 0:10:16.516 --> 0:10:19.036 used in large language models. Right, it's it's modern AI. 0:10:19.316 --> 0:10:23.436 And so that moment is this moment when it's like, oh, 0:10:23.516 --> 0:10:29.756 holy shit, GPUs are one hundred x better than CPUs 0:10:29.956 --> 0:10:33.196 for AI. Yes, that's interesting, that's useful to know. 0:10:33.956 --> 0:10:38.196 So it was a combination of GPUs data and algorithms 0:10:38.276 --> 0:10:41.796 that kind of combusted at that perfect moment. And then 0:10:41.916 --> 0:10:44.556 Jensen saw this and it's like this is a big 0:10:44.596 --> 0:10:48.076 deal and someday AI is going to be huge and 0:10:48.156 --> 0:10:50.476 we need to invest big in this. So on top 0:10:50.516 --> 0:10:54.316 of KUDA, he invests in all these AI libraries that 0:10:54.836 --> 0:10:59.476 effectively managed the GPUs to do AI workloads the most 0:10:59.516 --> 0:11:02.636 in the most effective manner. And he invested in libraries, 0:11:02.796 --> 0:11:07.116 invested in software, he invested in researchers, and allocated a 0:11:07.116 --> 0:11:10.876 lot of employees to this project starting in around twenty thirteen. 0:11:11.436 --> 0:11:14.676 And so that's that's more of the like it's not 0:11:15.196 --> 0:11:18.636 just the chip, right, it's like the chip that is 0:11:18.836 --> 0:11:23.396 optimized to be not only efficient on the hardware level. 0:11:23.396 --> 0:11:25.356 But like there's all this software. So if you are 0:11:25.876 --> 0:11:28.516 if you are building AI, if you're if you're writing 0:11:28.556 --> 0:11:31.716 code for AI, like they make it really easy for 0:11:31.756 --> 0:11:34.196 you to do it on in video GPUs. 0:11:34.236 --> 0:11:37.156 So it's all the math libraries, it's all the AI 0:11:37.236 --> 0:11:41.436 frameworks work the best on in video GPUs. And they 0:11:41.476 --> 0:11:44.756 added these things called tensor cores similar to krudercres. They 0:11:44.756 --> 0:11:49.236 actually added hardware circuitry inside the GPUs that are optimized 0:11:49.276 --> 0:11:53.076 to run AI training and workloads and all the all 0:11:53.116 --> 0:11:55.916 the software that you're running. And this is over again, 0:11:56.116 --> 0:12:00.156 this is over another nine eight years where they're building 0:12:00.156 --> 0:12:03.236 all the software and hardware circuitry to run deep learning 0:12:03.276 --> 0:12:06.076 AI the best. 0:12:06.556 --> 0:12:11.236 Okay, it's time. It's twot twenty two. Twenty twenty two, 0:12:12.396 --> 0:12:16.796 chat GPT comes out. What's that mean for in video? 0:12:17.316 --> 0:12:21.676 So actually check chip takes off and it's it takes 0:12:21.676 --> 0:12:25.116 about two quarters for the whole world to realize this 0:12:25.196 --> 0:12:29.356 is a big deal. This, this model of natural language 0:12:29.356 --> 0:12:34.956 processing where the computer actually understands what you're asking, and 0:12:35.076 --> 0:12:40.436 that ability to draw insights and be effective with all 0:12:40.476 --> 0:12:44.996 that natural language stuff just just blows people away, and 0:12:45.196 --> 0:12:48.636 companies and startups realize we have a new AI boom here, 0:12:48.756 --> 0:12:54.476 because it really unleashes a wave of capabilities that wasn't 0:12:55.076 --> 0:12:58.516 wasn't doable before. So about six months later, that's when 0:12:59.196 --> 0:13:01.916 the big bang I call it for Nvidia happens when 0:13:01.916 --> 0:13:06.436 they say, oh my gosh, we're gonna we're gonna beat 0:13:06.516 --> 0:13:09.116 our numbers that the street expects by four billion up. 0:13:10.036 --> 0:13:12.276 And the stock went up like one hundred and seventy 0:13:12.316 --> 0:13:15.396 billion dollars in value because the world realized. 0:13:15.036 --> 0:13:17.556 In like a day. Yes, right, there was one day, 0:13:17.636 --> 0:13:21.516 and it was spring of twenty twenty three. Way talking 0:13:21.556 --> 0:13:21.956 about and. 0:13:21.916 --> 0:13:25.316 This is three, this is three weeks after I had 0:13:25.396 --> 0:13:28.436 the meeting with the book publisher who asked me if 0:13:28.476 --> 0:13:30.996 I could do a book on video. So literally I started. 0:13:30.756 --> 0:13:33.436 Also good timing for you, good timing for a video, 0:13:33.516 --> 0:13:37.356 Good timing for you, And and right, no, I remember 0:13:37.396 --> 0:13:40.876 that day. And and there was a question then of like, 0:13:40.956 --> 0:13:42.956 oh my god, is this a one off? Are they 0:13:42.996 --> 0:13:44.156 going to keep growing this way? 0:13:44.236 --> 0:13:44.396 Right? 0:13:44.436 --> 0:13:48.316 Basically what they said was we sold way way more 0:13:48.436 --> 0:13:51.996 GPUs than we thought we were gonna. Right, that's that's 0:13:52.036 --> 0:13:54.516 what That's the basic thing they reported in their earnings 0:13:54.596 --> 0:13:58.156 report and the subtext is this is the world and 0:13:58.196 --> 0:14:02.036 in particular, like big tech companies with lots of money realizing, oh, 0:14:02.196 --> 0:14:05.596 we've got to get into this AI game more than 0:14:05.636 --> 0:14:07.316 we have been and to do that, we got to 0:14:07.356 --> 0:14:11.436 buy a lot of really expensive GPUs from video. 0:14:11.796 --> 0:14:16.876 And Jensen is really smart at selling this stuff. Like, 0:14:17.396 --> 0:14:22.356 he's very smart because every company, every company CEO, every startup, 0:14:22.956 --> 0:14:26.156 they saw this risk as the extential for them because 0:14:26.476 --> 0:14:30.996 if your competitor comes out with the AI featured offering 0:14:31.236 --> 0:14:34.156 that's much better than what you have today, that that 0:14:34.276 --> 0:14:37.996 isn't advantaged by the AI models, Yeah, that competitor could 0:14:38.876 --> 0:14:41.996 could drive you out of business. So Jensen was very 0:14:41.996 --> 0:14:46.196 smart as at selling to people that you need to 0:14:46.316 --> 0:14:47.316 get on board. 0:14:47.156 --> 0:14:52.196 At making everybody scared of losing. Yes, well, and in particular, right, 0:14:52.276 --> 0:14:55.116 I mean there is a small number of very large, 0:14:55.276 --> 0:15:00.156 very rich tech companies that are a very big part 0:15:00.156 --> 0:15:03.036 of in videos revenue. Right, it's like whatever the ones 0:15:03.116 --> 0:15:09.756 you quld think of, Google, Meta, Microsoft, Amazon, Yes, I mean, 0:15:09.756 --> 0:15:12.356 I guess open ai sort of is kind of connected 0:15:12.396 --> 0:15:15.676 with Like those companies are buying a huge share of 0:15:15.796 --> 0:15:19.556 these GPUs, right, They account for a big, big chunk 0:15:19.596 --> 0:15:22.276 of in videos revenue. Yes, and those are good companies 0:15:22.276 --> 0:15:24.436 they have as customers because they're super rich, right, they 0:15:24.436 --> 0:15:26.356 can afford to pay you tens of billions of dollars 0:15:26.436 --> 0:15:27.156 for your GPUs. 0:15:27.516 --> 0:15:31.636 But that's being said, they're reselling that GPU power to 0:15:32.036 --> 0:15:36.316 companies and startups, right, so startups are buying that GPU 0:15:36.356 --> 0:15:38.636 competing capacity from these. 0:15:38.596 --> 0:15:41.116 GNT tech companies. Yeah, that's a good point, right. So 0:15:41.116 --> 0:15:43.276 they're in a way not the end customer. They're the 0:15:43.316 --> 0:15:45.636 sort of intermediate kind of service providers. 0:15:45.676 --> 0:15:49.516 But they also use that for their own internal systems too, 0:15:49.596 --> 0:15:52.476 like Meta uses a ton of GP power to make 0:15:52.516 --> 0:15:57.516 their advertising algorithms more effective and to pick the videos 0:15:57.516 --> 0:15:58.396 like tech talk does. 0:15:58.796 --> 0:16:01.996 And obviously Google Search is now becoming more and more 0:16:02.076 --> 0:16:04.716 AI driven, and there's Gemini, which is right, I mean, 0:16:04.756 --> 0:16:07.436 there's there's more direct use of AI by all of 0:16:07.476 --> 0:16:12.076 the big tech companies as well. So there's another piece 0:16:12.156 --> 0:16:14.396 of this which is interesting. I mean, because there's one 0:16:14.476 --> 0:16:16.636 universe wherever it's like, oh, we got to get it 0:16:16.676 --> 0:16:19.636 on AI, and they all buy the GPUs and then 0:16:19.676 --> 0:16:21.996 that's kind of it. It's like a step function where 0:16:22.036 --> 0:16:25.476 like there's this momentary rush where everybody buys the GPUs 0:16:25.516 --> 0:16:27.756 and then whatever. They just upgrade every couple of years, 0:16:27.796 --> 0:16:31.796 and it's more like a regular tech hardware business, which 0:16:31.836 --> 0:16:34.436 is like good but not amazing. But there's another piece 0:16:34.476 --> 0:16:37.316 of it that has been a huge deal for end video, 0:16:37.356 --> 0:16:40.476 and that's the scaling law or the scaling hypothesis. Let's 0:16:40.476 --> 0:16:42.796 talk about that. What's the what's the scaling hypothesis. 0:16:42.836 --> 0:16:45.276 So the scaling hypothesis, similar to what I said before, 0:16:45.556 --> 0:16:48.876 is the combination of computing power. The more computing power 0:16:48.916 --> 0:16:53.436 you add, the more data you increase, and the better 0:16:53.916 --> 0:16:57.076 ways you figure out software algorithms for each of these 0:16:57.116 --> 0:17:00.516 three buckets. If you increase it, the AI model becomes 0:17:00.556 --> 0:17:02.716 more capable and more effective. 0:17:02.636 --> 0:17:05.516 And like, just to be clear, like even if the 0:17:05.556 --> 0:17:08.676 algorithms aren't getting that much better, right, Like my understanding is, 0:17:08.716 --> 0:17:11.956 and this is kind of a new idea. It's like, oh, 0:17:12.036 --> 0:17:16.196 if we can just have more data and more computing 0:17:16.276 --> 0:17:18.276 power will get better results. 0:17:18.396 --> 0:17:21.436 Yes, so it's great if you have all three doing 0:17:21.556 --> 0:17:25.756 you know, basically showing up. But for this time period, 0:17:26.196 --> 0:17:28.716 the last few years, the scaling law has really taken 0:17:28.796 --> 0:17:34.956 off and companies are literally ten xing their compute power, 0:17:34.996 --> 0:17:40.316 and these AI clusters going from sixteen thousand GPUs now 0:17:40.356 --> 0:17:42.676 to one hundred thousand gps and now people are saying 0:17:42.676 --> 0:17:46.516 they're going to build one million GPU clusters in the 0:17:46.556 --> 0:17:47.836 next two three years. 0:17:47.556 --> 0:17:51.396 And in videos selling most of those GPUs. 0:17:51.476 --> 0:17:53.676 Yes, So if you go from sixteen thousand to one 0:17:53.716 --> 0:17:55.436 hundred thousand GPUs in a couple of. 0:17:55.476 --> 0:17:57.116 Years, that's a good business to do. 0:17:57.556 --> 0:18:00.836 If you're increasing the scale of the hardware by ten 0:18:01.156 --> 0:18:04.196 ten x every couple of years. And the exciting thing 0:18:04.236 --> 0:18:07.516 for Nvidia is that, like what I argue is that 0:18:08.076 --> 0:18:09.996 this thing is going to this trend is going to 0:18:09.996 --> 0:18:13.196 continue in the next few years because you have the 0:18:13.236 --> 0:18:17.156 scaling laws. Then you have this thing called multimodal where 0:18:17.236 --> 0:18:19.796 they're using this GPUs not just for texts like in 0:18:19.876 --> 0:18:24.676 chat TPT, you're using it for video and images generating 0:18:24.716 --> 0:18:28.956 those things. And now there's these two other two other 0:18:29.076 --> 0:18:31.716 waves of demand that are happening. There's this thing called 0:18:31.796 --> 0:18:35.876 AI agents where these AI models can do multi step 0:18:35.956 --> 0:18:36.756 tasks for you. 0:18:37.156 --> 0:18:40.356 Right, you could be like book me a ticket to 0:18:40.436 --> 0:18:44.676 San Diego some weekend in June whenever it's the best deal, yes, 0:18:44.716 --> 0:18:45.756 and then it does that. 0:18:45.876 --> 0:18:49.156 So that's going to happen in the next twelve months. Yes, 0:18:49.596 --> 0:18:52.516 these AI agents will eventually, in the next year or 0:18:52.516 --> 0:18:56.796 two be able to do all that tedious work automatically 0:18:56.916 --> 0:18:59.716 and probably with less errors than an actual human being. 0:19:00.236 --> 0:19:03.636 So AI agents multimodal. And now there's this thing called 0:19:03.676 --> 0:19:07.516 test time compute, where the AI models, instead of just 0:19:07.756 --> 0:19:11.556 spinning back an instant apply, can actually think about what's 0:19:11.596 --> 0:19:14.556 the best way to respond to your question and spend 0:19:14.596 --> 0:19:16.716 more time thinking about it and then give you a 0:19:16.796 --> 0:19:20.556 higher quality answer. So there's all these things that all 0:19:20.596 --> 0:19:25.636 of these test time compute, AI agents, multi modal. These 0:19:25.636 --> 0:19:29.036 are all things that need more computing power, they need 0:19:29.116 --> 0:19:31.436 more GPUs, and these are all things that are going 0:19:31.516 --> 0:19:33.756 to drive in video's revenue in the next couple of years. 0:19:37.076 --> 0:19:38.756 We'll be back in a minute and we will talk, 0:19:38.796 --> 0:19:42.756 among other things, about Jensen Wong, one of the most 0:19:42.836 --> 0:19:46.556 successful entrepreneurs of the twenty first century and a man 0:19:46.596 --> 0:19:49.476 who once told the colleague that he wakes up every morning, 0:19:50.036 --> 0:19:53.396 looks in the mirror and says to himself, you suck. 0:19:59.236 --> 0:20:01.316 I want to talk about Jensen a little bit. I 0:20:01.356 --> 0:20:04.076 mean your book. You know, he is obviously the main 0:20:04.316 --> 0:20:06.516 character in the history of Nvidio is the main character 0:20:06.516 --> 0:20:09.836 in your book. He's been the CEO of n Video 0:20:09.876 --> 0:20:12.276 for more than thirty years, which is like longer than 0:20:12.276 --> 0:20:14.516 Bill Gates with the CEO of Microsoft, longer than any 0:20:14.876 --> 0:20:17.476 you know, almost anybody in the s and p. Five 0:20:17.516 --> 0:20:20.556 hundred has been a CEO at this point. His childhood 0:20:20.596 --> 0:20:22.756 is quite interesting, right, Like, tell me just a little 0:20:22.796 --> 0:20:23.876 bit about his childhood. 0:20:25.276 --> 0:20:28.956 He was born in Taiwan, and they moved around a 0:20:28.956 --> 0:20:33.316 little bit to Thailand, and his father came to training 0:20:33.356 --> 0:20:36.556 in New York and fell in love with America. This 0:20:36.636 --> 0:20:41.596 is like the great American dream story. And the mother 0:20:41.676 --> 0:20:46.636 and father started teaching Jensen's brother English ten words a day, 0:20:47.276 --> 0:20:49.236 and they sent him to his aunt and uncle when 0:20:49.236 --> 0:20:52.076 he was about age eight or nine, and the aunt 0:20:52.196 --> 0:20:55.716 and uncle and the families. The funny thing is they 0:20:55.756 --> 0:20:58.916 sent him to a reform school in Kentucky by mistake, 0:21:00.036 --> 0:21:02.916 thinking that he will get a great education at this 0:21:02.996 --> 0:21:04.676 boarding school in Kentucky, which. 0:21:04.516 --> 0:21:06.676 He just thought it was a boarding school, yes. 0:21:06.676 --> 0:21:09.196 But it turned out to be a school for troubled. 0:21:08.876 --> 0:21:12.956 Kids, huh, which Jensen was not. He was just like 0:21:13.036 --> 0:21:15.636 a smart kid with ambitious for him parents. 0:21:15.756 --> 0:21:18.916 Yes, But he talks about how his time at Oneita 0:21:18.956 --> 0:21:21.836 Baptist and student Kentucky was formational for him. 0:21:21.996 --> 0:21:23.076 What was it like for him? 0:21:23.836 --> 0:21:26.956 It was hard at the beginning, but he started befriending people. 0:21:27.996 --> 0:21:32.196 He started playing chess with the janitor, and he just 0:21:32.276 --> 0:21:35.876 learned how to deal with other people much better. And 0:21:36.436 --> 0:21:39.036 he talks about how he learned his street fighter. 0:21:38.756 --> 0:21:41.356 Mentality because he was literally fighting. 0:21:42.636 --> 0:21:45.716 Yes, I mean there were bullies and all that stuff, 0:21:45.436 --> 0:21:47.556 but he just learned how to deal with people and 0:21:48.436 --> 0:21:53.076 the rough and tumble of kids back then. Eventually his 0:21:53.196 --> 0:21:58.476 parents came from abroad and they settled in Oregon, and 0:21:58.676 --> 0:22:02.996 he learned his work ethic from working at Denny's. He 0:22:03.076 --> 0:22:06.956 talks about how he washed more dishes and cleaned more 0:22:06.996 --> 0:22:11.596 bathrooms than any CEO in the history of CEOs. He 0:22:11.636 --> 0:22:14.236 says he Denny's helped them in terms of social skills 0:22:14.636 --> 0:22:18.156 and dealing with time pressure and dealing with customers. But 0:22:19.836 --> 0:22:22.796 it's the work ethic that really sets him apart. So 0:22:22.916 --> 0:22:25.916 even at the beginning, he was working from nine am 0:22:26.036 --> 0:22:29.756 to midnight, and he just set a culture where Nvidia 0:22:29.756 --> 0:22:31.956 employees work really hard. 0:22:32.356 --> 0:22:34.956 And I feel like at the beginning that's common, right, 0:22:35.076 --> 0:22:38.276 Like that is the classic startup story. But the classic 0:22:38.276 --> 0:22:41.636 startup stories you do that for five years and then 0:22:42.116 --> 0:22:44.596 either you get giant and you hire a grown up 0:22:44.676 --> 0:22:47.476 CEO and you you know, go start your blimp company 0:22:47.556 --> 0:22:51.596 or whatever, or you sell or whatever. He's still doing that, 0:22:51.876 --> 0:22:55.436 right He's sixty years old and like wildly rich and 0:22:55.836 --> 0:22:57.516 and he's still working that much. 0:22:57.556 --> 0:23:02.196