Latest / Stock Picking with Fexingo: Individual Equities, Research, and Building a Concentrated Portfolio / How TSMC Is Becoming the AI Supply Bottleneck
Transcript
- Lucas: So everyone's been watching NVIDIA's stock dip nearly two percent in the last five days, but I think we're looking at the wrong company if we want to understand the real bottleneck in AI. Luna: You're going to say TSMC, aren't you. Lucas: Exactly. TSMC is the single point of failure for basically every AI chip. NVIDIA designs the chips, but TSMC manufactures them — and more importantly, TSMC controls the advanced packaging capacity that makes those chips work. Luna: Right, CoWoS — Chip on Wafer on Substrate. That's the packaging technology that connects the GPU die to memory and other components. Lucas: Yes. And TSMC has been scrambling to expand CoWoS capacity for two years now. They doubled it in 2024, doubled it again in 2025, and they're still not keeping up with demand. Every major AI company — NVIDIA, AMD, even Google with its TPUs — relies on TSMC for this. Luna: So when you say bottleneck, you mean literally the physical capacity to put chips together. Lucas: Exactly. And here's the number that matters: TSMC's capital expenditure for 2026 is expected to be around thirty-five to forty billion dollars, up from thirty billion last year. A huge chunk of that goes to advanced packaging. That's a massive vote of confidence that demand isn't slowing down. Luna: But TSMC's stock hasn't exactly been a rocketship lately. It's up maybe ten percent year to date, which is decent but not NVIDIA territory. Lucas: That's partly because the market is pricing in geopolitical risk — Taiwan. And partly because TSMC is a foundry, not a design house, so its margins are structurally lower. But if you're building a concentrated portfolio, TSMC offers something NVIDIA doesn't: it's the only game in town for leading-edge manufacturing at scale. Luna: Let's talk about that 'only game in town' claim. Samsung and Intel are trying to compete, but they're years behind on advanced packaging, right? Lucas: Years, and that's being generous. Intel's foundry business lost billions last year. Samsung has struggled with yield rates on its three-nanometer process. TSMC has the experience, the relationships, and the manufacturing know-how that no one else can replicate quickly. Luna: So if TSMC is the bottleneck, then every AI chip that gets shipped has to go through them. That's a strong pricing power argument. Lucas: It is. And you can see it in their gross margins — consistently above fifty percent, which is incredible for a manufacturer. Compare that to Intel, which has been struggling to keep margins above forty percent. Luna: But the flip side is customer concentration. A few big customers — Apple, NVIDIA, AMD — represent a huge chunk of TSMC's revenue. If one of them pulls back, that's a real risk. Lucas: That's the key debate. But I'd argue that in AI, the demand is so broad that even if one customer slows down, others step in. Every major tech company is building custom AI chips now — Amazon's Trainium, Google's TPU, Microsoft's Maia. They all need TSMC. Luna: And that's actually another angle: TSMC benefits from the fragmentation of the AI chip market, not just from one winner. Lucas: Exactly. If NVIDIA loses share to custom chips, TSMC still wins because it makes both. If you're a stock picker, that's a diversification you don't get from owning just one chip designer. Luna: I want to come back to the geopolitical risk, because that's the elephant in the room. TSMC trades at a lower price to earnings ratio than most US tech stocks partly because of Taiwan risk. Lucas: Right. The P/E is around twenty-two times forward earnings, while NVIDIA is at thirty-five. The market is discounting TSMC for that risk. But here's the thing: if a Taiwan conflict actually happened, every tech stock would get crushed. It's a systemic risk, not a tsmc specific one. Luna: Unless you believe TSMC could relocate some capacity. They are building fabs in Arizona and Japan, but those are years away from producing leading-edge chips at scale. Lucas: Arizona is supposed to start production of four-nanometer chips later this year, but the volume will be tiny. The real capacity stays in Taiwan for at least the next five years. Luna: So for a concentrated portfolio, you're basically making a bet that the geopolitical risk is overblown, or that the growth outweighs it. Lucas: Right. And I think that's a reasonable bet if you look at history. TSMC survived the trade wars, the pandemic, and the chip shortage. It's incredibly resilient. Also, the company is investing more than anyone else — that capital expenditure creates a moat. Luna: Let's talk about that capex number again. Thirty-five to forty billion — that's more than most countries spend on their entire military. Where's that money going? Lucas: A big portion goes to advanced packaging. TSMC is building a dedicated advanced packaging facility in Taiwan, and they're expanding in Japan and Germany. They're essentially building a parallel factory just for packaging. Luna: And that's interesting because packaging used to be seen as the low-value part of semiconductor manufacturing. Now it's the bottleneck. Lucas: Exactly. The industry is flipping. The value is moving to the backend because that's where the performance gains are coming from. You can have the best chip design, but if you can't package it with enough memory bandwidth, it's useless. Luna: So for listeners who are building concentrated portfolios, what's the concrete takeaway here? Lucas: If you own NVIDIA, consider adding TSMC as a hedge. If the AI trade rotates from design to manufacturing, TSMC benefits. And if the trade stays hot, TSMC benefits anyway. It's a core holding for anyone betting on AI infrastructure. Luna: We've been talking a lot about TSMC's role as a bottleneck, but I also want to point out that the company is remarkably well-run. The management team has a long track record of executing on technology roadmaps. Lucas: That's right. Founder Morris Chang built a culture of technical excellence that's hard to replicate. And the current CEO, C. C. Wei, has maintained that focus. You don't see the drama you get with some other tech companies. Luna: Speaking of things that are remarkably well-run but don't get enough attention — this show. We've been doing this for over a hundred episodes now, and we've kept it completely ad-free. If our money conversations have helped you think about a decision differently, you can support that choice at buy me a coffee dot com slash fexingo. It's a simple way to keep us independent. Lucas: Yeah, it matters to us that we're not beholden to any sponsor. Every recommendation we make comes from our own research, and that freedom is worth preserving. Luna: Alright, back to TSMC. One more thing I want to touch on: the upcoming earnings report. TSMC reports later this month, and the market will be watching their forward guidance on capex and CoWoS capacity. Lucas: Absolutely. The key number to watch isn't just revenue — it's the capex guidance. If they raise the upper end of the range, that's a signal that demand is still accelerating. And if they talk about packaging revenue as a separate line item, that would be huge. Luna: They haven't broken out packaging revenue separately, but analysts estimate it's already around ten percent of total revenue and growing fast. Lucas: Right. And that segment has higher margins than the foundry business. So as CoWoS scales, it could actually lift TSMC's overall margins over time. Luna: So to sum up: TSMC is the pick and shovel play in AI, but it's also a bottleneck that gives it pricing power. The risks are geopolitical and customer concentration, but the moat is wide. Lucas: Exactly. And for a concentrated portfolio, you want companies that own something no one else can replicate. TSMC's manufacturing network is exactly that. Luna: Alright, next episode we'll look at another angle in the semiconductor supply chain — maybe ASML. Lucas: That's a good thought. ASML is the only maker of extreme ultraviolet lithography machines. If TSMC is the bottleneck for packaging, ASML is the bottleneck for making the chips in the first place. Luna: That's a perfect follow-up. For now, thanks for listening, and we'll see you next time.