Latest / The CEO Diary with Fexingo: Leadership Lessons, Executive Decisions, and Corner Office Stories / How Jensen Huang Built NVIDIA from Graphics Chips to AI Infrastructure
Transcript
- Lucas: So June 2026. NVIDIA is now worth more than the entire stock markets of most countries. But back in 2006, Jensen Huang made a decision that, at the time, looked borderline reckless to a lot of people inside his own company. Luna: And you're talking about the decision to build CUDA — the parallel computing platform — before there was any market for it? Lucas: Exactly. Because in 2006, NVIDIA was a graphics card company. Its core business was selling gaming GPUs. Revenue was about three billion dollars. And Huang decided to pour billions into developing a software platform that would let programmers use those GPUs for general-purpose computing, not just rendering polygons. Luna: And at the time, the main use case for that was basically... nothing? I mean, researchers were barely using it. Lucas: Right. There was no AI boom. There was no data center demand. Huang was essentially building a platform for a market that didn't exist yet. And the cost was enormous. Over the next decade, NVIDIA would invest something like ten billion dollars cumulative into CUDA development. That's a huge chunk of a company that, for years, had operating margins of fifteen to twenty percent. Luna: So how did he sell that internally? I mean, ten billion dollars is not a rounding error, even for a successful company. Lucas: From what I've read, he didn't really 'sell' it in the traditional sense. He simply made it non-negotiable. He told the engineering teams: 'This is the future. We are building a computer company, not a graphics company.' And he was willing to take the short-term hit to margins and stock price. Luna: That's a very different style from, say, a consensus-driven CEO who needs buy-in from all stakeholders before moving. Lucas: Huang is famously hands-on and technically deep. He still interviews every single engineering candidate at NVIDIA — and that's a company with over thirty thousand employees. He can make these bets because he genuinely understands the architecture better than almost anyone. But it also creates a culture where you either get on board or you leave. Luna: And a lot of people did leave, right? I remember reading that there was significant turnover in the late 2000s because engineers felt they were wasting time on a platform no one was using. Lucas: That's right. The attrition was real. But Huang kept going. And then, around 2012, something shifted. Alex Krizhevsky, a graduate student, used two NVIDIA GPUs with CUDA to train a deep learning model called AlexNet that won the ImageNet competition by a huge margin. That moment is often cited as the beginning of the modern AI revolution. Luna: So the bet started paying off. But it was still a few years before it showed up in the financials, right? Lucas: Absolutely. NVIDIA's data center revenue was essentially zero in 2014. By 2020, it was about seven billion dollars. By 2024, it was over forty billion. The compound annual growth rate from 2020 to 2025 was something like sixty-five percent per year. That is not just growth — that is a structural shift in the industry, and it is entirely built on that ten billion dollar CUDA bet. Luna: But here's the question I keep coming back to: is that sustainable? Because right now, everyone from AMD to Google to a dozen startups is building their own AI chips and their own software stacks. NVIDIA's monopoly on AI training is real, but it's not guaranteed. Lucas: And that is exactly the tension Huang is managing right now. The moat is not just the hardware — it's the software ecosystem. CUDA has hundreds of thousands of developers who have built tools, libraries, and entire workflows that are deeply optimized for NVIDIA hardware. Switching costs are enormous. Luna: But history shows that software moats can erode. Microsoft had a huge moat with Windows, and then came the internet and mobile. So what does Huang do to stay ahead? Lucas: He's doing two things. First, he's accelerating the hardware refresh cycle. NVIDIA now releases a new architecture roughly every two years — Hopper, then Blackwell, then Rubin — each one delivering massive performance gains that competitors struggle to match. Second, he's investing in CUDA compatibility with new workloads: robotics, drug discovery, climate modeling. He wants NVIDIA to be the computing platform for every form of scientific computing, not just AI. Luna: So it's not just about defending the castle — it's about expanding the kingdom. Lucas: Exactly. And that requires the same kind of long-term thinking that CUDA required back in 2006. The difference is, now the market is real, and the stakes are higher. Huang himself has said that the company is only at the beginning of what he calls 'the next industrial revolution.' Luna: I guess the lesson for anyone running a company is that some bets are so big you can't prove them in advance. You just have to have conviction and the stomach for short-term pain. Lucas: And that's the core of what we try to do on this show — look at how leaders make those kinds of decisions. If these conversations have moved your work forward in some small way, we're glad. And just a reminder — we deliberately don't run ads on these episodes. If you want to support that choice, the link is buy me a coffee dot com slash fexingo. No pressure, just an option for people who find value in it. Luna: Yeah, it's a small way to keep this going without any interruptions. Appreciate everyone who does. Lucas: Alright, back to NVIDIA. One more thing I find fascinating: Huang's insistence on a flat organizational structure. He has something like sixty direct reports. That is insane by traditional management theory. Luna: How does that work? I've read that it creates chaos if not done right. Lucas: In NVIDIA's case, it forces an incredibly high-bandwidth communication culture. There are no layers to filter information. Huang expects every leader to be technical enough to understand the full picture and to speak openly. It's exhausting for some people, but it keeps the company moving at a pace that competitors can't match. Luna: So the culture is a feature, not a bug. It's designed to select for people who thrive on that intensity. Lucas: Exactly. And that's the final piece of the puzzle. The architecture bet, the software ecosystem, and the organizational design all reinforce each other. Jensen Huang built a machine that is optimized for making a specific kind of long-term bet and then executing on it at a speed that's hard to replicate. Luna: It will be fascinating to see how that machine handles the next decade, when the competitive landscape is very different from the one he dominated. Lucas: No doubt. And that's a story we'll keep following.