Latest / Business Strategy Talks with Fexingo: Competitive Advantage, Market Positioning, and Growth Planning / How Nvidia Built an Unbreachable Moat in AI Chips
Transcript
- Lucas: So here's a number that stopped me this week. Nvidia controls somewhere north of 90 percent of the market for AI training chips. And that's not just a stat you can shrug off—it's one of the most dominant market positions we've seen in technology since, I don't know, Microsoft in the late 90s. Luna: Ninety percent—that's absurd. But isn't that just because they got lucky with the AI boom? They were making gaming GPUs, and then suddenly everyone needed them for neural networks. Lucas: That's the common narrative, and there's truth in it. Jensen Huang himself has said the deep learning breakthrough in 2012 was a 'happy accident' for Nvidia. But what turned that accident into a durable competitive advantage isn't just hardware. It's CUDA. Luna: CUDA being their parallel computing platform, right? The software layer that lets developers write code that runs on Nvidia GPUs. Lucas: Exactly. And CUDA launched in 2006, years before AI was a mainstream business. Nvidia invested heavily in making it easy for researchers to use their chips for general-purpose computing. By the time the AI boom hit, there were already hundreds of thousands of developers trained on CUDA, and billions of lines of code optimized for it. Luna: So the moat is really the ecosystem. It's not that Nvidia makes the best chip—though they do—it's that if you're a researcher or a startup, switching to an AMD or Intel chip means rewriting all your code from scratch. Lucas: That's exactly the switching cost that creates the moat. And it's not trivial. Some of these models take months to train. If you've invested that time optimizing for CUDA, you're not going to throw it away for a 10 percent cost saving on hardware. The total cost of switching is too high. Luna: So they're raising the bar on both fronts: hardware performance and software lock-in. But who's trying to challenge them? AMD has the MI300 series, and Intel has Gaudi. Are those credible threats? Lucas: They're credible in the sense that they have big budgets and decent specs. But the challenge is software. AMD has ROCm, which is their open-source alternative to CUDA, but it's been plagued with compatibility issues and a much smaller developer community. Intel's oneAPI is similar—good in theory, but adoption is slow. Luna: I've heard that some big tech companies are designing their own custom chips—Google's TPU, Amazon's Trainium, Microsoft's Maia. Doesn't that erode Nvidia's moat from the top? Lucas: It does, but only for those specific companies. Google isn't selling TPUs to the world—they use them internally for their own AI workloads. So Nvidia still dominates the broader market for anyone who wants to buy chips off the shelf. The custom chip trend is real, but it's mostly the hyperscalers. And even they still buy Nvidia GPUs for certain tasks. Luna: So the moat is strongest for the mid-market and startups. If you're a company that wants to train a large language model, you're almost certainly going with Nvidia. Lucas: Right. And there's a network effect at play here, too. More developers using CUDA means more libraries and tools built on top of it, which makes it even more attractive for new developers. It's a classic flywheel. Luna: But isn't there a risk that the open-source community eventually builds a cuda compatible layer? I've seen projects like ZLUDA that aim to run CUDA code on AMD GPUs. Lucas: There are efforts, but Nvidia has been aggressive about protecting CUDA. They've made their EULA prohibit using translation layers for competing hardware. And even if a translation layer works for 80 percent of operations, the remaining 20 percent often causes performance issues or bugs. In AI, where you're pushing the limits of hardware, that 20 percent can be a dealbreaker. Luna: Let's zoom out. If this conversation has been useful to you in thinking about competitive advantage, it's worth mentioning that the podcast stays ad-free thanks to listeners who support us at buy me a coffee dot com slash fexingo. We don't do sponsorships—just a small way to keep the show independent. Lucas: Yeah, and those contributions really do help us keep digging into these case studies. So if today's episode moved your thinking forward in any way, that link is a simple way to say thanks. Now, back to Nvidia. Lucas: One angle I find fascinating is the geopolitical dimension. Nvidia has had to navigate export controls on advanced chips to China. They designed a lower-spec chip, the H800, to comply, but then the US tightened the rules further. Luna: That actually creates a weird dynamic. Instead of hurting Nvidia, the export controls limit competition from Chinese chipmakers who might otherwise build on cuda compatible platforms. So it reinforces their moat in the rest of the world. Lucas: Exactly. And it also reduces the addressable market for potential competitors, because they can't sell into China either. So Nvidia's dominance in the West is effectively protected by policy. Luna: But let's talk about the long-term threat. If AI workloads shift from training to inference—that is, running models rather than building them—does the moat weaken? For inference, you don't necessarily need the absolute best GPU. You might use a cheaper, lower-power chip. Lucas: That's a really good point. In fact, Nvidia has been pushing its inference capabilities with TensorRT and other software optimizations to make sure their chips remain competitive even for inference. But there's no question that the moat is strongest in training, where performance and ecosystem really matter. Luna: So what's the single biggest risk to Nvidia's moat? Is it AMD catching up on software? Custom silicon? Or something else? Lucas: I'd say the biggest risk is that AI itself evolves in a way that doesn't require the kind of parallel processing that GPUs excel at. If we see a shift to architectures like neuromorphic computing or optical computing, that could reset the playing field. But that's years away, if it happens at all. Luna: So for now, Nvidia's moat looks as strong as ever. They've got the hardware lead, the software ecosystem, the developer lock-in, and even geopolitical tailwinds. Lucas: Right. And it's a great case study for anyone studying competitive strategy. The moat isn't just one thing—it's a combination of technology, timing, and ecosystem that creates a virtuous cycle. Competitors can't attack just one piece; they'd have to dismantle the whole system. Luna: It reminds me of how Microsoft dominated the PC era with Windows—through developer lock-in, not just the OS itself. Lucas: Exactly. And we know how long that moat lasted—decades. So if Nvidia plays its cards right, we could be looking at a similar run in AI hardware. Luna: That's a compelling thought to end on. Thanks, Lucas. Lucas: Thanks, Luna. See you next time.