The Man Who Wrote the Book on NVIDIA

Yeah, yeah, absolutely. So Tay wrote a book called The Nvidia Way. It's really good. I just read it. I recommend that other people read it. And it is just like, it's a look at what is an incredibly important company that, as you mentioned at the beginning of the book, it sounds like you were kind of shocked that nobody had done an Nvidia book, and that's kind of weird. Like, we use products that are built on their products all the time, and even if we don't use them directly, like, we are all beneficiaries of the flops that Nvidia produces at great scale and surprisingly low cost. And yeah, I find it a really fun book because it feels like sometimes when you read a book about a company, you do find they had some secret, some concept, some big idea that they were able to execute on. And you look at a company like Google and it's like at every step the question is always, where is the world's unorganized information? How do we organize it? How do we make it accessible? And then how do we put ads into it? And like, that is a really powerful insight that can show up just about everywhere. Or you could look at like Microsoft's early insight of, hey, computer hardware is really expensive. You know, it was really expensive when they started. It's only going to get cheaper, and therefore you want to own the software, which is this complement to it, where every time there's a new iteration, not only is the hardware cheaper, but now there are more users. And if your software has almost no marginal cost and the hardware does have a marginal cost, like, you want to be selling the near-zero marginal cost thing and let everyone else do the annoying hardware stuff. And then Nvidia does the annoying hardware stuff. Like, they are the ones who are actually connecting wires to things and figuring out how many watts this thing needs and how much heat it throws off and all of that stuff. And so it felt like part of your thesis in the book seems to be like, there isn't really a secret. Like, Nvidia, you know, hired great people, has an amazing founder, and they just work really, really, really hard and set incredibly high standards. And it seems like they're just always on. Like, you have this anecdote about how Jensen, the way he unwinds on the weekends is that he pours himself a glass of nice scotch and then just answers work emails for hours. Yeah, so tell me a little bit about just, you know, like, I guess this somewhat rehashes some of the early parts of the book, but what was the genesis of this? And, you know, how did you think about the process as you were writing this book? Like, were you looking for the secret and, you know, didn't find it? Or was it like you pretty much had this impression that, no, they just work really hard and they're smart, and so that hard work is a lot of leverage?

So the genesis was I got a cold email from a publisher in May of last year, and a friend of mine, Matthew Ball, who wrote the Metaverse book, suggested to his publisher that Tay could write the book on Nvidia. My first thought was, there's no way no one has written a book on Nvidia. I mean, this is a little bit before the AI stuff really took off, but, you know, it was the largest chip maker by market cap. As a PC gamer, they've pretty much dominated the graphics card market for video games for decades. So I went to Amazon and took a look, and I couldn't believe it. No one has written the book on Nvidia. And, you know, there's four or five books on Facebook, there's multiple books on Amazon, Apple, and I just thought for a sec, I could do this. Like, I know the technology, you know, being a PC gaming nerd and addict for the last 30 years. I've been in media for the last 10 years, so, you know, I've been, you figure out how to become a better writer, and I know the business because I've been on Wall Street for a bunch of years as a technology analyst, business strategy. So I thought I had the three components to really do a good job. So I, you know, quickly within a few weeks, I met the publisher, got a book agent, and we're off and running. And the first thing I did was call LinkedIn. It's an amazing resource for an author, and, you know, just reaching out to Nvidia employees. And within six to seven calls, I realized the big thing that differentiates Nvidia, like you said, was the culture. It's completely different from other companies, the large technology companies. Every former Nvidia employee that goes to a large tech company like Google and Microsoft says it's extremely difficult to adjust to the culture. Nvidia is faster, but the work ethic is on a different level. Bernie, you mentioned something about the secret sauce. There's one other anecdote about when they acquired the assets and hired 100 engineers from 3dfx, and 3dfx was the number one graphics player in the mid to late 90s, and they got beat up by Nvidia and pretty much destroyed. And the 3dfx engineers came to Nvidia and were looking for the secret. What's the secret? What's the magic trick? And then they realized the only secret is you just have to work crazy. And the relentless execution and the work hours, that was just working harder. So I think when you look at the Nvidia way, I think the work ethic, you know, from the beginning and even today, you mentioned Jensen working on Sunday night. He loves to work. Like, this is, he literally says it's relaxing to him. When he goes to movies, he doesn't remember the movie because he's constantly thinking about work. So that relentless kind of hard work ethic is a key component. The second component is cultivating talent, and we could go into this later, but he's constantly paying his superstar engineers. If they're doing a great job, he is going to cultivate and retain that talent by doubling, tripling their stock comp. And he's constantly looking to hire the best people. So when he interviews people, he's finding out who the best person is at their former jobs. And the last thing is just speed and velocity in doing everything. I think large companies, when you go to large organizations, the internal politics just take over, and everyone's looking out for their own, you know, power, backstabbing, taking credit. So when there's new projects, new disruptive projects, they just get stuck in the quicksand where you have to get approvals from five different leaders for it to go forward. Nvidia is the exact opposite. Nvidia, Jensen just gets the top 10, 15 people in the room, including, you know, the lower-level employees, and they figure out what's the main problem, what's the main thing to discuss, hash it out in an intellectually honest way, and then they move and execute that vision. So that speed and velocity, I think they call it speed of light, Nvidia is a huge component of why they've been able to succeed.

Yeah, yeah. Well, I do want to get to the world's weirdest org chart in a bit, but I'm also like, one of the things you pointed out that was really striking me because it just hadn't occurred to me, but it is of course true, is Jensen's actually one of the longest-tenured CEOs of any publicly traded company. Like, you have Buffett at Berkshire, and then I think you mentioned one other person, but it's like pretty much he's it. And, you know, he has been running Nvidia longer than Gates ran Microsoft, and longer than Larry was CEO of Oracle, still calls a lot of the shots. And so, yeah, and, you know, there's obviously going to be a selection effect where, yeah, he has great performance because, you know, great performance and a really long tenure because of his performance. Really bad at some point, the board or an activist investor would have told him to find someone else. But, yeah, what do you make of that? Like, how does someone have that kind of endurance? And I guess the follow-up question is, if you have someone like that who has been, he was there day one, you know, naming the company, you know, they're like in the Denny's booth figuring out what the initial business plan would be, is still there today, you know, on stage and doing all these events and things. Who comes next? And does Nvidia have to have a different kind of structure and approach once there's someone else in charge?

I think part of it, like you said, is he's just, he's been so successful across the decades. So he built up a level of credibility that no one can take away from. I mean, right now they're on top of the world. It's like unprecedented, the level of revenue. I think six quarters ago they were doing seven billion in revenue, and now they're doing thirty-five billion in revenue a quarter. This is hardware and chips. This is not software. You could just copy the bits over and over again. So you have this level of success. I think he just loves, like I said, he loves to work. This is his passion. Nvidia's life. He says, I love Nvidia over and over again. So if you have that combination where he loves his job, he works harder than anyone else, he talks about how I might not be the smartest person in the world, but no one's ever going to outwork me. Like, he just has that engine inside him that is unbeatable. So you combine those two things and the level of success he had, no one's going to question and no one's going to tell Jensen what to do. There was an activist fund, Starboard, which is pretty famous, that almost did something. You know, my rational kind of speculation, talking about covering that in the book, but lucky for Jensen and Nvidia, they just did big stock buyback. The stock popped twenty percent, and Jeff decided to move on. Maybe history is a little different if the stock didn't go up and he took the gain. So there were times when Wall Street was all over him when the gross margins were coming down. This is not just one or two quarters. This is like four or five years of weak gross margins because he was investing in the CUDA cores on the GPUs. So I mean, a lot of it is kind of maybe happy accidents, but over time, like, you know, he's proven to be able to see beyond corners of all these big tech moves all along the way. So yeah, that's where you have. In terms of future, what happens after Jensen, the first thing I answer to that is he's in his early sixties. He loves his job more than anything. He's working harder and just is doing better than ever in the last thirty years. I just don't see him leaving anytime soon. I mean, someone mentioned Morris Chang, like, worked till his eighties. So there's still many, many years ahead. When he does eventually leave, the big question is who the next CEO, the competence of the next CEO. And I think history has shown that it could be good or bad. I mean, John Sculley was a disaster, almost drove Apple to bankruptcy. Steve Ballmer, you know, wasn't that great, even though the financials are better. I think, you know, he made a lot of mistakes.

I do think Ballmer's underrated, but it is definitely true that Gates was just a very special CEO.

I think, yeah, when you compare Gates to Ballmer, I mean, Gates is on a different level in terms of knowledge of technology. I mean, you could argue, you know, Ballmer did a good job, sort of like Tim Cook, like managing the financial aspect in the business. But yeah, that's the big question. Are you going to get like a triple, quadruple-A CEO again, or are you going to get someone that's, you know, not so good? And corporate America is full of CEOs that are pretty, I'll just say it, they're pretty incompetent. And as, like I said, as of now, I don't see Jensen leaving anytime soon.

Is there an obvious successor at Nvidia?

No, I mean, I don't see anyone that's obviously up to this level, but yeah, but we'll see. You know, you never know.

And, you know, I think one of the other things that makes that long tenure really impressive is that Nvidia is in a cyclical business. And it feels like a lot of maybe the marginal buyer of Nvidia stock right now is maybe less aware of that, someone who's been following the company for a long time. But like, it is the case that sometimes there are a lot of GPUs just gathering dust somewhere and not selling as fast as expected. In fact, it hasn't been that long since Nvidia had to take a big write-down. So, you know, how do you think about how they navigated that cycle, especially because it does feel like the vibe in the book is that Jensen is always saying, you know, we won last time, so let's bet everything on this next thing. And that seems like a tough, you know, definitely if you get that kind of call right a couple cycles in a row, it does really well. But if there is a long, an extended difficult cycle, not so good. So yeah, how do you think about them, how they navigated that cyclical, like the combination of like, it's a secular growth story that, yeah, we will need more GPUs over time, but any given year you don't know if we need, you know, fifty percent more or like twenty percent less than the year before.

So this is a nuanced situation and answer. I mean, I agree with you. In the past, they've been very cyclical. Sometimes the product, I think Turing and NV330, they have a couple really disastrous, not so great products where a competitor does design a better chip than them. So there's always that risk. And, you know, there may be one every three, four GPUs, they just don't make a great product. There was this, the crypto stuff, which I actually didn't have time to cover in the book. And I actually, in my day job at Barron's, I was very out in front, kind of predicting the problems of crypto and Ethereum when they changed the way Ethereum worked. I do think the AI stuff is a little different because this is the early stages of a massive computing shift. When you have the interim CEO of Intel saying this is one of the biggest computing shifts in decades from CPUs to GPUs. I mean, Intel is the dominant player in CPUs, and the interim co-CEO of Intel is saying this is a really big shift, the biggest in decades. That to me tells me like this is going to be as big as mobile when Apple did the iPhone, as big as Wintel when Microsoft and Intel, you know, just rode the PC revolution for decades. So I do think the cycle here is less of an issue. It's really a massive computing shift where we're going from the CPU stuff, which, you know, is very static, retrieval-based on files, right? Windows Server, you try to look for one file, to this GPU computing paradigm where it's a completely different paradigm. You're basically ingesting all the information you have, distilling it into an AI model, and being able for that AI model to generate a business intelligence insight or generate a new piece of content. It's completely different. So I do think we're still in the early innings of this wave, computing shift wave that is going to last for several years. But I mean, if you look at the chart of Nvidia, the data center revenue didn't really take off until like early, early last year. Like the vast majority of its revenue was gaming GPUs, like fifty to eighty percent for the vast majority of its life. So it's only in the last six, seven quarters that the data center part of their business has really taken off.

Yeah, and you know, I really liked the bits that you wrote about this transition from Nvidia as a graphics card company where their core market is gamers who want to play the latest games, maximum resolution, full detail. You know, they want the fog effects and the lighting and the explosions needs to look good, etc. Like somehow there's a transition from that to like, we're not modeling this fog in order to make Silent Hill look even creepier. We're modeling fog because we work at a major lab and we're doing some kind of really elaborate weather modeling thing. And I like that very, very early on, it was basically people tricking the GPU into thinking of something as like, pretend you're drawing a triangle and then solve this scientific computing problem for us. And that almost felt a lot like just the way that sometimes human cognition works, where you look at a string, a series of numbers, and you don't see a pattern, and then you make a graph and you realize, oh, it's trending up or trending down, or you know, it's getting more or less volatile, or why is there this weird flip here? Yeah. So talk to us a little bit about how that transition worked. And I'm especially, and this kind of shades into the other question of like, how aggressive Nvidia is as a business. How did Nvidia navigate this, especially when they had products that were meant for these data center type scientific computing workloads, and then they also have just off-the-shelf, much cheaper GPUs that were designed for graphics but could be used for other things that a lot of researchers just prefer to buy because they're cheaper?

So you touched up on a great point that a lot of the AI stuff, all the high-performance scientific research, it's matrix multiplications, the same kind of math and all that calculations and math is the same math that was in video games. Like I think I mentioned in the book how for video games, you have to simulate the physics of a bullet, like it comes out of the gun, where it lands, and all that kind of high-performance, intense computing simulation is the same stuff you do with science, you know, figuring out how water moves inside of fluid dynamics of a cloud, you know, trying to predict how biological molecules would work for potential drugs. So all this is kind of simulating real-life physics, and that's how Nvidia was able to move from video games to CUDA and all the AI stuff later on. It's basically the same kind of math. It's the same kind of physics. So that really helped Nvidia. The funny thing is everything Nvidia has learned in the past in terms of business, it's the same playbook that they won with video games. They're doing to win in data center AI. I touch upon how in the late 1990s, Jensen figured out how all these graphics companies would dominate for one cycle and then lose their perch. And he learned that the problem was that a typical chip took eighteen months to develop and ship, and then a different graphics company would come out with a better chip because their eighteen-month cycle was up. How about I just change that to six months?

Yeah, I always like that because it feels like maybe, you know, what if someone at 3dfx has the same observation of, hey, our product cycle is eighteen months, PC product cycle is six months. And, you know, I think for a lot of people, they would say, therefore, our business is lumpy and we're not going to be number one in two consecutive years. And then the Nvidia answer is, hey, all we have to do is work three times as hard and then we'll always have something new whenever there's a new PC cycle. And it seems like that worked out well. Like, do you think that was literally it? Was that the competing companies just didn't either didn't have the capital to do that or just did not have the will to work that hard?

I think there are two components. I mean, I'm being a little simplistic when I say, oh, just go from eighteen months to six months. I mean, it was the eighteen-month architecture, but they created two derivatives that were faster. It wasn't like a completely different architecture every six months. But I think there are two components of why Nvidia was able to do it. One is just technical expertise. One of the co-founders, Curtis Priem, was able to do this mini kind of virtualized operating system that allowed them to, if they couldn't make a new great feature in hardware fast enough, you could virtualize that into software and vice versa. So that gave them a lot more flexibility than the competitors that didn't have this virtualized operating system where you could put features in software or hardware. So, and then it's just the relentless drive of Jensen. Like, as I mentioned in the book, he is very, very tough on his employees, especially if they don't come through on the schedules that he wants. So he will dress that person down. The one anecdote, one best anecdote I have is during one company meeting in front of the entire company, he dressed down this manager of the RivaTNT chip. He was late, and he just kept on getting the camera in his face like, RivaTNT, you need to finish this chip. You're not doing a good job. And just humiliate him in front of the entire company. That might not be great. That might not feel great, but it definitely gets people to work really hard. Like, you don't want to be that guy in front of the whole company being dressed down. So it's a combination of technical expertise and just that hard-driving culture that's driven by Jensen. So that's why they're able to do that six months instead of eighteen months. And that same playbook is being done today with AI. In last October, over a year ago, he said the data center AI GPUs, we're going to accelerate the product cadence from two years to every year. So if you're AMD, that's really hard. You're behind already, but now you have to match Nvidia's new product cadence. And the other part, not just the hardware, is Nvidia from the beginning has invested tremendous amounts of money on software. They're software drivers, and from the beginning, they sent their engineers to all the video game publishers to make sure, you know, their software is working correctly. It's a strategy they call devs before revs, developers before revenue. So Jensen always knew from the beginning it's really important to make sure your software is the most reliable, the best, the most feature-rich, and he's willing to invest tons of money even though it directly hurt their P&L statement. But he knew the secret sauce for their competitive advantage was going to be this, was going to be the software. That's exactly the same thing that's happening with AI right now.

Yeah, like I, that was actually, that was going to be my next question. I noticed there was this thread throughout the book, I guess starting maybe a third of the way in with that point about how they started just releasing their own, releasing drivers directly to end users on a monthly cadence rather than just hoping that nobody would ever ship any software that would break something. And so you have that, and then you have the virtualization thing where if there's a function that the GPU can't handle in hardware, you can emulate the hardware, and then it's, you know, slower, but it's still doing, it's still full-featured. And then there's of course CUDA. So it does feel like Nvidia has just made this gradual transition towards being more of like, they're still selling hardware, but more of their advantage, where their lock-in comes from software and the fact that people have learned these software libraries and are used to being able to trust Nvidia software. So it kind of feels like that has to be a tough transition for a hardware company because they feel like their competitive advantage is hardware, and then they're not really monetizing the software directly, and they are making this investment that pays off over a really long period. So yeah, did you get the sense that that was actually a difficult transition for them to make, or was it just like, if you have high enough standards for your product, then you'll solve all of these adjacent problems because that's the only way to ship something that's actually satisfactory?

It wasn't really a transition. It was from the beginning, Jensen and the co-founders knew the importance of software from the prior experiences at Sun Microsystems for the two other co-founders. So it was there from the beginning. They knew over time that was so important for customers. One of the coolest things after this, during this book process, was having John Carmack tweet about my book. He's the creator of the two major PC gaming kind of revolutions in the early nineties. Everyone that age knows of Doom and Quake. I mean, those are the two most important PC games in history. And he said, he gave a historical context as he was reading my book, saying that the importance of unified drivers by Nvidia, like you said, was so important because typically before that, you had to download the drivers from the PC graphics card website, and then you would have to download one driver for your graphics card model. It was a convoluted, complicated process. And Carmack said most of these graphics drivers were terrible. Like they were buggy, had performance problems. So when Nvidia went with the unified graphics model where you just downloaded one file, it worked with all the Nvidia graphics cards, and Nvidia wrote the reliable drivers instead of, you know, five different PC makers and, you know, forty different graphics card makers writing one, creating one file for one particular GPU model, it just made things work. Like it was a huge revolution in how it all worked. And John said, John Carmack said this was a huge deal for the development of the PC gaming industry. It was a huge advantage for Nvidia for making these drivers. And the same thing, like I said, the same thing is with CUDA. The reason why AI startups and companies go with Nvidia's AI CUDA platform is that the software works. It's reliable. You don't want to save, you know, thirty percent on the chips, but then you put the fate of your business on a platform that has tons of bugs and has reliability issues. You don't want to do that. SemiAnalysis just literally wrote a piece talking about how they had no problems trying to build their AI stuff and test on Nvidia CUDA, and then they tried to do the same thing on AMD software, and it was, there was nothing but problems. So if you're a company that the fate of your business depends on you keeping up with AI, when your competitor's working on an AI model or a feature that's using Nvidia, you're going to use Nvidia too because you can't risk the future of your business on something not working.

Yeah, and one of the interesting things...

That you mention in this is in the book is Nvidia's strategy of selling into academic environments where they would do a grant and it would take the form, you know, partly cash, partly just hardware, and then people would learn to build their—they learn to do their simulations on Nvidia hardware. And I, through the newsletter, I do some recruiting and I talk to people who are getting PhDs in STEM topics, and it feels like half of the PhDs—it's basically—it's mostly MLOps, like it's mostly how do you get enough time on a limited set of GPUs and how you negotiate that and how do you optimize to minimize the training cost of whatever you're training. So yeah, tell us a little bit about how that strategy worked and where you think the payoff showed up.