SYLVIE DOUGLIS, BYLINE: NPR.
(SOUNDBITE OF DROP ELECTRIC SONG, "WAKING UP TO THE FIRE")
DARIAN WOODS, HOST:
All right, we have a riddle for you. See if you can guess what we're talking about today.
WAILIN WONG, HOST:
I was born at a Denny's restaurant.
WOODS: My name is Latin for envy.
WONG: I played a lot of video games in my younger years. Today, I'm more into poetry and translation.
WOODS: I'm 31 years old, and I've been down and out so many times that my unofficial motto is that I'm 30 days from going out of business.
WONG: Still don't have it? What if we said, last week, for a moment, I was the most valuable company on the planet?
WOODS: Yeah, that's right. The answer is the computer chip designer Nvidia. And Nvidia's chips are the leading choice for artificial intelligence developers. In the last couple of years, Nvidia's sales have shot up like a jet plane - its stock price more like a rocket.
WONG: But there are several chip companies out there, like AMD and others. So why is Nvidia the one dominating the AI scene? This is THE INDICATOR FROM PLANET MONEY. I'm Wailin Wong.
WOODS: And I'm Darian Woods. Today on the show, Nvidia - we learn about how this company went from selling niche graphics cards for gaming to becoming the biggest company in the world. And we ask whether Nvidia has the moat to defend its very valuable castle.
(SOUNDBITE OF MUSIC)
WOODS: Nvidia is often called a chipmaker, and that's a bit of an oversimplification. First of all, Nvidia doesn't actually manufacture the semiconductor chips. That's done by companies like TSMC and Samsung. Nvidia does the designing, like where the circuits all go on the chip. It also makes the software so that developers can work with those chips.
WONG: This combination has its roots in 1992. Jensen Huang and two other engineer friends met at a Denny's restaurant in San Jose. Their idea was to improve video games by building specialized chips for rendering 3-D graphics. This decision would become incredibly lucrative.
WOODS: One person who's spent a lot of time unearthing in Nvidia's history is David Rosenthal. David is one half of the tech podcast "Acquired," which made this epic series of podcasts on the company's history.
DAVID ROSENTHAL: Nvidia corporate communications got in touch and said, who were your sources? You know, who told you all this in the company? And we were like, well, nobody. We just - you know, we watched a lot of YouTube videos with Jensen, and we read a lot of papers.
WONG: This led to an in-person interview with Jensen Huang towards the end of last year.
(SOUNDBITE OF PODCAST, "ACQUIRED")
ROSENTHAL: First question - when did you realize that only eight were going to work?
JENSEN HUANG: We should have - I realized I didn't learn about it until it was too late. We should have...
WOODS: By all accounts, working for Jensen is super intense and not for everybody. Failures are shared publicly, for example. But David says those who stay are loyal.
ROSENTHAL: Jensen is truly singular in Silicon Valley, and I think - I mean, the man is 60 years old. Nobody from his generation is still running their company in the same way that he is. He is probably more engaged in every detail at Nvidia than any of those other founders ever were at their companies.
WONG: David says the way that Jensen is involved in the details is one key to why Nvidia rode the AI wave so astutely.
ROSENTHAL: He started moving the Nvidia ship in the direction that it is today with AI and machine learning starting in, like, 20 years ago.
WOODS: That's incredible foresight. You know, it's obviously way before ChatGPT, but it's even before driverless cars or even voice recognition. How did he see this coming?
ROSENTHAL: Well, I think this is the key to Jensen. If you were to say that sentence to him, I think he would respond and say, no, it was not before all those things. All of those things were happening. They were just happening in the deep scientific computer science research community. He was so deeply plugged into all of this that he knew the principal actors who now are some of the, you know, true leaders in AI research at OpenAI, at Anthropic and elsewhere these days. He knew them personally. He was reading the white papers, he was visiting them at universities - all back in the mid-2000s.
WONG: At this time, Nvidia was mainly known by 15-year-olds upgrading their gaming computers. But in the early 2000s, Jensen essentially made a huge bet on selling supercomputing power to a wider range of people. And as part of this, the company launched a software development framework called CUDA. CUDA acts kind of like a middleman between the software developer and the chip, and the system would be crucial for AI after a turning point in 2012.
ROSENTHAL: This was the big-bang moment that kicked off research and investment at a commercial scale into artificial intelligence.
WOODS: It was at this annual competition for researchers to submit AI systems that could recognize images from a massive database. The database was called ImageNet, and progress had been gradual. Entries could correctly identify what a majority of images were but would typically have an error rate of around 25 or 30%. That year, a team from the University of Toronto submits this entry that they call AlexNet.
ROSENTHAL: It blows away the rest of the field. So the percentage of images it got wrong was around 15% - 15. This was a quantum leap.
WONG: The team used an existing method of AI called a neural network, which was powerful but had always been limited by how much computing power it required to train. To overcome this, AlexNet did something different. The standard approach had been to train these models using central processing units - or CPUs - the generalist brains in your computer. But these calculate their instructions sequentially. You can think of it like one calculation, then another. So instead, this team ran these training calculations using graphics processing units - GPUs - what you would play video games with.
ROSENTHAL: A graphics card handles tens of thousands of instructions at a time. I've really widened the pipe of the amount of compute that I can stuff through this thing at any given point in time.
WOODS: And who makes graphics cards? Nvidia, of course.
ROSENTHAL: People at Google and people at Facebook said, holy crap. You could use, they realized, these image classification systems to build way better social media feed recommenders.
WONG: So for image recognition, programmers were training the software to notice patterns in, say, pictures of cats, whiskers, fur, four legs. And similarly, the big social media companies realized they could train software to notice patterns in what kinds of pictures and posts and movies people like on the internet.
ROSENTHAL: That was, you know, billions and billions and billions of dollars of profits. So there was a good 10-year run where, like, nothing else mattered.
WOODS: And that meant that places like YouTube and Facebook and Instagram were hiring a lot of AI talent. And David says this was actually part of the motivation for Elon Musk and Sam Altman to found OpenAI.
ROSENTHAL: They were really worried that Google and Facebook had just become a duopoly of all the AI development and research talent because it was the only, you know, economically viable use case.
WONG: OpenAI's launch of ChatGPT in late 2022 was the next hinge moment for AI and for Nvidia. It awoke investors and everyday people to advancements in AI that got them dreaming about the future. It's where Nvidia suddenly started its journey becoming one of the most valuable companies on the planet.
WOODS: In the parlance of tech investing, Nvidia has this giant moat. David reckons it's protected from competition in the foreseeable future because of CUDA - that development system.
ROSENTHAL: These systems build on top of themselves over time and get more and more complex and powerful. So it would be like somebody going and starting a new phone operating system from scratch and saying, OK, what are all the list of things we need to build to make this phone operating system viable to compete with Android and iOS? That's a tall order.
WONG: David says network effects are also important. Now, millions of other developers use CUDA. So if you're a college student, that's the language you're going to learn. It's a self-reinforcing cycle. This moat is something that has gotten people speculating about government action. News outlets like The New York Times have reported that the Department of Justice has cleared the way for possible antitrust action against Nvidia.
WOODS: Jensen, though, has said he doesn't like the word moat and prefers to think of the company as working alongside the entire AI ecosystem. Whichever words you use to describe it, Nvidia is standing on top of the world right now, though, even with all its chip sales, its stock price is very highly priced by conventional measures. One misstep, and there is a long way to fall.
(SOUNDBITE OF MUSIC)
WONG: All the way into the moat filled with barra-CUDAs (ph)?
WOODS: Barra-CUDAs - that's quite funny.
This episode was produced by Julia Ritchey with engineering by Cena Loffredo. It was fact-checked by Cooper Katz McKim. Kate Concannon is the show's editor, and THE INDICATOR is a production of NPR.
Copyright © 2024 NPR. All rights reserved. Visit our website terms of use and permissions pages at www.npr.org for further information.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.