Latest / The Operator's Manual with Fexingo: Daily Decisions, Tactical Wins, and Business Execution / How Nvidia Broke Its Own Supply Chain and Fixed It
Transcript
- Lucas: So back in early 2024, Nvidia was in a weird spot — they had customers screaming for H100s, they had Blackwell on the horizon, and yet they were struggling to ship anywhere near what the market wanted. Luna: Right, I remember headlines about twelve-month lead times on some SKUs. That's insane for a company worth two trillion dollars. Lucas: Exactly. And the root cause wasn't demand being unpredictable — everyone knew demand was huge. The problem was execution on the supply side. Specifically, three bottlenecks that Nvidia had to solve simultaneously. Luna: Let's hear them. I'm guessing one is the CoWoS packaging at TSMC — that's been a known constraint for years. Lucas: That's number two. But the first, and maybe the most surprising, was chip-on-wafer substrates. Those are the ceramic or organic interposers that connect the GPU die to the rest of the package. In 2023, Nvidia was competing with every other chipmaker for a global supply that was maybe 300 million units a year — and they needed half of that for just two products. Luna: So they basically had to corner the market. What did they do? Lucas: They went to the three main substrate suppliers — Ibiden, Shinko, and Unimicron — and signed multi-year agreements with prepayments. I've seen estimates that Nvidia paid roughly $15 billion in non-cancellable orders across 2024 and 2025 just to lock down substrate capacity. That's not a loan, that's cash upfront that you don't get back if demand drops. Luna: That's a massive bet. It's basically saying 'we trust our forecast more than we trust the flexibility of the market.' Lucas: Right. And that's the core operational tension in this story — speed versus flexibility. Nvidia chose speed. And it worked in the short term: by Q3 2025, they were shipping 3.6 million GPUs per quarter, up from about 1.2 million in Q1 2024. Luna: But what about the second bottleneck — the CoWoS packaging? TSMC was already expanding capacity, but it takes years to build new fabs. Lucas: Right. CoWoS — chip-on-wafer-on-substrate — is a 2.5D packaging technology that Nvidia uses for both H100 and Blackwell. TSMC's capacity in 2023 was about 120,000 wafers per month. Nvidia needed maybe 80,000 of those just for themselves. So they essentially did the same thing: prepaid for capacity expansion. There's reporting that Nvidia funded a big chunk of TSMC's new CoWoS fab in Zhunan, Taiwan, which is slated to come online late 2026. Luna: So they're vertically integrating without actually owning the assets. It's like 'we'll pay for your factory if you promise to fill our orders first.' Lucas: Exactly. And the third bottleneck was memory bandwidth. H100 uses HBM3 memory from SK Hynix and Samsung. Blackwell uses HBM3e, which is even more advanced. Nvidia locked up supply by co-investing in SK Hynix's new HBM production lines in Cheongju, South Korea. That meant they got priority allocation when supply was tight in early 2025. Luna: So across substrates, packaging, and memory — all three were solved by writing big checks upfront. But what's the downside? If demand softens, Nvidia is stuck with billions in prepaid commitments. Lucas: That's the risk. And it's not theoretical — in late 2025, there were reports that some hyperscalers were slowing their H100 orders as they transitioned to Blackwell. Nvidia had to manage that transition without being caught with too much old inventory. They managed it well — they repurposed some H100 capacity for Blackwell by using the same CoWoS lines — but it required constant coordination with TSMC. Luna: So the playbook is: identify bottlenecks early, pay to own the capacity, and design your products to share as much of the supply chain as possible. That sounds like a masterclass, but it also sounds fragile. Lucas: It is fragile. And that's the lesson. Nvidia's approach works when demand is growing 200 percent year-over-year. But if growth slows to 20 percent, you've got a balance sheet full of prepaid assets that are worth less. Jensen Huang himself said at an investor conference in March that 'supply chains are not built for exponential curves — they're built for linear ones.' Luna: He's basically admitting that they're forcing a round peg into a square hole. But they're getting away with it because they have the cash and the leverage. Lucas: Right. And that's the specific operational insight I want listeners to take away: when you're in a hypergrowth phase, the conventional wisdom of 'keep your supply chain flexible' can be exactly wrong. Sometimes you need to lock down capacity with hard commitments, even if it means taking on risk. But you need to be honest about what that risk costs. Luna: Let's talk about the cost. Nvidia's gross margin actually dipped from 73 percent in 2024 to around 69 percent in 2025. Part of that was higher packaging costs from the prepaid capacity. Lucas: Exactly. The prepayments don't show up as a direct cost on the P&L — they're on the balance sheet as deposits. But the actual cost of goods sold increased because the capacity they locked in was at premium pricing. Suppliers knew Nvidia needed it, so they charged accordingly. Luna: So the margin compression was basically a tax on speed. They chose to pay more per unit in exchange for getting more units sooner. Lucas: Yes. And the interesting thing is, analysts initially freaked out when margins dropped. But by Q1 2026, margins stabilized around 70 percent because Blackwell's higher average selling price offset the cost increase. So the bet paid off — but only because they executed the product transition exactly right. Luna: And if they hadn't? If Blackwell had been delayed or had lower demand? Lucas: They'd be sitting on billions in stranded deposits and capacity they couldn't use. That's the knife edge. And honestly, that's what makes this story worth studying — it's not a simple success story. It's a case study in managing risk under extreme uncertainty. Luna: Before we wrap, I want to mention something that's related. These deep-dive episodes take a lot of time to research and produce — we dig into SEC filings, supplier contracts, earnings call transcripts. If you've gotten something useful out of them, a couple of dollars a month genuinely makes a difference. It's what keeps the show ad-free and independent. Lucas: Yeah, we've heard from listeners who've used specific examples from episodes like this in their own work. If that's been you, consider tossing a few bucks at buy me a coffee dot com slash fexingo. It's a small way to say 'keep going.' Luna: Exactly. And we really appreciate it. Okay, back to Nvidia. Lucas: So the last piece of this puzzle is what Nvidia does next. As of May 2026, the next bottleneck is already emerging: liquid cooling for data centers. Blackwell GPUs draw 700 to 1000 watts each, and you can't air-cool that many in a rack. Luna: So now Nvidia is dealing with a bottleneck that's not even in their supply chain — it's in the customer's infrastructure. That's a whole new category of operational challenge. Lucas: Right. And they're responding by partnering with cooling companies like CoolIT and Boyd Corporation, and even designing reference architectures for liquid cooling that they give to hyperscalers for free. It's the same playbook: identify the constraint, invest to remove it, but this time the investment is in someone else's business. Luna: So the lesson evolves. First, lock down your own supply. Then, help your customers build the infrastructure to use your product. It's vertical integration through influence, not ownership. Lucas: Exactly. And that might be the most durable takeaway: the best supply chain strategy is the one that makes your customers' problems your own — because then you have the incentive to solve them first. Luna: Well said. That's it for today. Thanks for listening. Lucas: See you next time.