SPEAKER_1: Alright, so last episode we established the core idea: find the constraint, follow the capital. Now I want to apply that framework to the first and most visible case—Nvidia. Because this is where the whole chain started. SPEAKER_2: Right, and it's worth being precise about what the market believed before the rerating. Most investors saw Nvidia as a gaming GPU company that had a decent data-center business on the side. Not a platform. Not infrastructure. A chip vendor with a niche. SPEAKER_1: So what broke that belief? What was the actual demand shock? SPEAKER_2: Generative AI adoption at scale. When frontier model training became a serious commercial priority for hyperscalers, the compute requirement was enormous—and it turned out Nvidia's GPUs were essentially the only viable option at that scale. By 2023, Nvidia held an estimated 98% of data-center GPU shipments. That's not a market share number. That's a monopoly on a critical input. SPEAKER_1: Mm-hmm. And the H100 specifically became the symbol of that scarcity. SPEAKER_2: Exactly. The H100, from Nvidia's Hopper generation, became the scarce engine of the generative-AI boom. By August 2023, Nvidia had effectively reached its sales capacity limit for H100s for the entire year. Orders fully booked. That's not a demand story anymore—that's a supply story. And supply stories reprice assets differently. SPEAKER_1: Wait—so the rerating wasn't just about AI being exciting. It was about Nvidia being the only door into AI infrastructure. SPEAKER_2: [emphasis] That's the key idea. Control over a scarce enabling layer triggers a valuation reset before downstream profits are even clear. The application companies building on top of those chips hadn't proven their business models yet. But Nvidia didn't need them to. It was selling shovels in a gold rush, and there was only one shovel factory. SPEAKER_1: That actually explains something counterintuitive—why downstream AI application companies attracted less immediate investor enthusiasm than the chip supplier enabling them. SPEAKER_2: Precisely. The application layer has competition, uncertain monetization, and long sales cycles. The chip layer had none of that ambiguity. Demand was structural, supply was constrained, and pricing power was absolute. Think of it this way: if every AI lab needs H100s and there aren't enough H100s, Nvidia sets the terms. SPEAKER_1: So what did the revenue numbers actually look like when the rerating confirmed itself? SPEAKER_2: Dramatic. Data-center revenue alone was $22.6 billion, up 427% year-over-year. And for full fiscal year 2025, data-center revenue reached approximately $115.2 billion, more than doubling from the prior year. Those aren't incremental beats. That's a business transformation in real time. SPEAKER_1: Right—but how? Like, what specifically made Nvidia so hard to displace? Because 98% market share invites competition. SPEAKER_2: Three layers. First, the GPU hardware itself—H100s and then Blackwell chips purpose-built for AI training. Second, the networking stack that connects GPUs inside a data center. Third, and this is the durable moat: CUDA. It's the software platform developers have built on for years. Switching away from Nvidia means rewriting enormous amounts of code. That's not a six-month project. SPEAKER_1: So the software lock-in is actually doing as much work as the hardware scarcity. SPEAKER_2: [short pause] Maybe more, long-term. The hardware scarcity is what triggered the rerating. The software ecosystem is what sustains the multiple. By mid-2026, independent analysis put Nvidia's AI accelerator market share at roughly 80 to 88%. That's after years of competitors trying to close the gap. SPEAKER_1: Now, the bottleneck didn't stop at the chip itself. There's a packaging constraint that most people missed. SPEAKER_2: CoWoS—chip-on-wafer-on-substrate packaging at TSMC. TSMC's chair stated publicly in 2023 that the binding constraint wasn't a shortage of AI chips per se, but a shortage of CoWoS capacity. That constraint was expected to persist for around a year and a half. Delivery times for high-end Nvidia GPUs stretched to around 40 weeks at the peak of shortages. SPEAKER_1: Forty weeks. That's almost a full year just to get the hardware. SPEAKER_2: And that's what makes this a real bottleneck by our framework from last episode. Supply takes years to build, demand is structural, pricing power is clear. TrendForce projected TSMC's total CoWoS capacity would grow 150% in 2024 and more than 70% in 2025—with Nvidia occupying nearly half of that capacity. The constraint was real and durable. SPEAKER_1: So what are the risks that could weaken the thesis from here? Because the obvious winner is already priced. SPEAKER_2: Four pressure points. Custom silicon—hyperscalers like Google and Amazon building their own AI chips to reduce Nvidia dependence. Export controls limiting Nvidia's addressable market in certain geographies. Supply eventually catching up, which eases pricing power. And valuation multiple compression if growth decelerates. The Omdia forecast puts AI data-center chip shipments at around $207 billion in 2025 and growing—but custom ASICs are gaining ground. SPEAKER_1: So the takeaway for someone tracking this chain: Nvidia was the first rerating, and it pointed directly to two next questions—who manufactures the chips, and what feeds them memory fast enough. SPEAKER_2: That's exactly where the capital moved next. TSMC for foundry capacity, SK Hynix and the HBM suppliers for memory. SK Hynix reported its HBM products were almost fully booked through 2025. And there's a third follow-on that's easy to miss: even when the GPUs ship, some customer sites lack sufficient power capacity to deploy them. Power availability is emerging as the next binding constraint. That's the episode after memory. SPEAKER_1: And that power constraint is worth flagging now, because it's easy to assume the bottleneck ends at the chip. But the chip has to actually run somewhere. SPEAKER_2: Right. And that's the handoff we'll get to. But first—for someone tracking this story in real time—the scale of what happened to Nvidia's financials is still striking even when you know the headline. The AI data-center chip market grew more than 250% between 2022 and 2024. That's not a sector expanding. That's a sector being invented. SPEAKER_1: And Nvidia captured most of that. By Q4 2024, GPUs generated about 87.6% of the record $32.6 billion in AI data-center chipset revenue that quarter. Nvidia held roughly 85% of the accelerator market inside that. SPEAKER_2: [short pause] Those numbers confirm something important about operating leverage. When you control a scarce input and demand is structural, revenue doesn't grow linearly—it compounds. Gartner estimated that data-center semiconductor spending nearly doubled from $64.8 billion in 2023 to $112 billion in 2024. GPUs were the primary driver. SPEAKER_1: So the rerating wasn't just about Nvidia's earnings beating estimates. It was about the market repricing what kind of company Nvidia actually was. SPEAKER_2: a gaming GPU vendor trades at one multiple. An infrastructure gatekeeper with pricing power and a software moat trades at a completely different one. The earnings confirmed the thesis, but the multiple expanded before the earnings arrived. That's what a rerating looks like. SPEAKER_1: Mm-hmm. And the CUDA point is worth pressing on. Because hardware scarcity fades eventually—supply does catch up. But software lock-in is stickier. SPEAKER_2: Much stickier. CUDA is the reason switching costs are so high. Developers have built years of tooling, libraries, and workflows on top of it. Moving to a competitor's accelerator isn't just a procurement decision—it's a rewrite. That's why, even by mid-2026, independent analysis put Nvidia's AI accelerator share at roughly 80 to 88%. Years of competition, and the gap barely closed. SPEAKER_1: Now, the risks. Because the obvious winner is priced, and someone listening to this series needs to know what could break the thesis. SPEAKER_2: Four pressure points worth naming. First, custom silicon—hyperscalers building their own AI chips to reduce Nvidia dependence. Second, export controls limiting Nvidia's addressable market in certain geographies. Third, supply eventually catching up, which compresses pricing power. And fourth, valuation multiple compression if growth decelerates. Omdia forecasts AI data-center chip shipments reaching around $207 billion in 2025—but notes custom ASICs are gaining ground. SPEAKER_1: So not a broken thesis—but a maturing one. The first-mover advantage is real, the moat is real, and the risks are also real. SPEAKER_2: [emphasis] That's the right framing. And it's also why the more interesting investment question, at this point in the chain, isn't Nvidia itself. It's who Nvidia depends on—and who depends on Nvidia. SPEAKER_1: Which brings us to the handoff. Because the bottleneck didn't stay at the chip design layer. SPEAKER_2: It couldn't. Nvidia designs the chips, but TSMC manufactures them. And the CoWoS packaging constraint meant that even Nvidia's own demand couldn't be fully met. TSMC's chair said publicly in 2023 that the binding limit wasn't chip design—it was CoWoS capacity. That constraint was expected to last around a year and a half. TrendForce projected CoWoS capacity growing 150% in 2024 and more than 70% in 2025, with Nvidia occupying nearly half of that total. SPEAKER_1: So foundry capacity becomes the next question. And then memory—because HBM has to feed those GPUs fast enough to matter. SPEAKER_2: SK Hynix reported its HBM products were almost fully booked through 2025. TrendForce projected HBM exceeding 20% of total DRAM market value in 2024, potentially 30% by 2025. That's a parallel bottleneck running alongside the GPU shortage—not downstream from it, but concurrent with it. SPEAKER_1: And then the third follow-on—power. Even when the GPUs ship and the memory is there, some sites can't actually deploy them because they lack sufficient electrical capacity. SPEAKER_2: Which is where the rerating chain moves next. Vertiv, Eaton, GE Vernova—those names start appearing in the same conversation as Nvidia because the constraint migrated. The key idea for our listener tracking this series: Nvidia was the first rerating because it was the most visible bottleneck. But visibility is temporary. Capital follows scarcity, and scarcity moves. SPEAKER_1: Nvidia's rerating wasn't just a stock story. It was the opening move in a longer chain. Control over a scarce enabling layer—chips, then packaging, then memory, then power—triggers repricing at each link. SPEAKER_2: And next episode, we follow that chain to TSMC and the foundry layer. Because once the market understood that Nvidia was the bottleneck, the next question was immediate: who actually makes the chips? That's where the capital moved, and that's where the next rerating began.