Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How Nvidia Redesigned Its Supply Chain for AI Dominance
Transcript
- Lucas: A couple of years ago, if you wanted to buy an Nvidia H100 GPU, you waited twelve weeks. Now, for their latest Blackwell chip, that wait is down to about six weeks — for the most in-demand product in the history of computing. Luna: Six weeks still sounds long, but cutting it in half while demand has gone parabolic is genuinely impressive. What changed? Lucas: The core operational shift was moving from a push-based supply chain to a build-to-order model. Historically, Nvidia made GPUs for gaming and sold them through distributors — they'd forecast demand, build inventory, and push it out. That worked when annual growth was maybe twenty percent. Lucas: But when AI demand exploded, the old model broke. You can't forecast a market that's doubling every year. So around 2023, Nvidia started treating its data center business more like a custom fabrication shop. Luna: How does a build-to-order model work for a chip company? It's not like they can just flip a switch on the fab line. Lucas: Right, the actual silicon wafer fabrication still runs in big batches at TSMC. That part hasn't changed. But Nvidia changed what happens after the wafers are cut. They used to take finished chips and put them into standard configurations — a fixed number of memory modules, a standard cooler, a standard board. Lucas: Now they hold partially configured chips in a 'work in progress' state at their own facilities. Once a customer — say, a cloud provider like Azure or Google Cloud — places an order, Nvidia configures the chip with the exact memory, the exact networking interface, and the exact cooling solution that customer needs. That's the build-to-order part. Luna: So the bottleneck shifted from the fab to the configuration step. And they invested heavily in automating that. Lucas: Exactly. They built a new operations center in Santa Clara, essentially a mini-factory that does final assembly and test for data center GPUs. It's highly automated, with robots handling the delicate thermal interface material application and the high-bandwidth memory stacking. That facility alone cut the final assembly time from three weeks to one. Luna: One week for final assembly is remarkable. But I imagine the real operational challenge is demand forecasting for raw materials — especially the advanced packaging capacity at TSMC. Lucas: That's the real headache. Nvidia's chips use what's called CoWoS — Chip-on-Wafer-on-Substrate — which is a complex packaging technology. TSMC has been scrambling to add capacity. Nvidia now gives TSMC a twenty-four-month rolling forecast, updated monthly. They also committed to multi-year prepayments — billions of dollars — to lock in capacity. Lucas: In 2022, TSMC had about 30,000 CoWoS wafers per month. By early 2026, that number is reportedly over 100,000. Nvidia is the biggest customer driving that expansion. Luna: And they've diversified their supply base too. I know they started working with Samsung for some advanced packaging, and even with Intel Foundry Services for certain components. Lucas: Yeah, they've gone from single-sourced to multi-sourced for several critical materials. The high-bandwidth memory — HBM — which is stacked right next to the GPU, was almost entirely supplied by SK Hynix. Now Samsung and Micron are also qualified suppliers. That gives Nvidia leverage and resilience. Luna: And the results show up in the financials. Their gross margin went from around 60 percent in 2022 to over 70 percent in 2024 and has stayed there. Part of that is pricing power, but part of it is operational efficiency. Lucas: Exactly. A big piece is yield improvement. Nvidia's chips are enormous — the Blackwell die is nearly as large as a reticle limit, around 800 square millimeters. Defects are inevitable. But they worked with TSMC to improve defect density, and they also added redundant circuit blocks that can be laser-trimmed if a defect is found. Lucas: That's a classic operations tactic: design for manufacturability. They spent engineering time upfront to make the chip more tolerant of defects, which boosted usable yield from around 60 percent to over 80 percent in two years. Luna: So every chip that comes off the line is more likely to be sellable. That directly expands effective capacity without building a new fab. Lucas: Precisely. And they've applied the same thinking to the supply chain software. Nvidia actually uses its own AI platforms to model demand. They have a system internally called 'Sawtooth' — I think that's the code name — that ingests signals from cloud provider buildouts, enterprise IT budgets, and even government AI spending plans. It generates a probabilistic demand forecast every week. Luna: So they're using their own chips to predict demand for their own chips. That's almost circular. Lucas: It's a flywheel. Their GPU clusters run the forecasting models, and the models tell them how many GPUs to build. They say the forecast accuracy improved by about 30 percent year-over-year in 2025. That matters because overproduction is expensive — a single H100 board costs thousands in raw materials. Luna: And underproduction is even worse when you're leaving money on the table in a seller's market. Lucas: Right. So the whole operations philosophy boils down to flexibility. They've built a supply chain that can handle wild swings in demand — which is exactly what they need when governments, cloud giants, and startups are all racing to buy GPUs. Lucas: But there's one more aspect that I think is underappreciated: their inventory strategy. Nvidia now carries more finished goods inventory than they used to — about 60 days' worth versus maybe 30 days pre ai boom. Luna: That seems counterintuitive. With demand soaring, why hold more inventory? Lucas: Because supply chains are fragile. If there's a disruption — an earthquake in Taiwan, a fire at a chemical plant — they want to be able to keep shipping. That extra inventory acts as a buffer. And given that their gross margin is north of 70 percent, the carrying cost is trivial compared to the risk of losing a sale. Luna: It's a calculated trade-off: slightly higher working capital costs for dramatically higher revenue certainty. That's a mature operational mindset. Lucas: Exactly. And it's a reminder that good operations isn't just about cutting costs — it's about enabling growth. If you can deliver a Blackwell chip in six weeks while your competitor takes twelve, you win the deal. Luna: This kind of conversation is exactly why we keep doing the show — digging into how real companies solve real operational puzzles. And it's listener-supported, which means no ads, no sponsors. If these conversations have moved your work forward or just satisfied your curiosity, you can support us at buy me a coffee dot com slash fexingo. Lucas: Yeah, a small contribution goes directly to keeping the podcast independent and ad-free. Same URL — buy me a coffee dot com slash fexingo. Luna: Alright, back to the supply chain. Lucas, you mentioned Nvidia's partnership with TSMC. How deep does that operational integration go? Lucas: It's incredibly deep. Nvidia has embedded its own engineers at TSMC's fabs in Hsinchu. They work side by side with TSMC's process engineers to optimize the manufacturing recipes. That kind of co-location is rare — most customers just hand over a design and wait for wafers. Lucas: Nvidia also shares its production schedule with TSMC's internal planning system. So TSMC's own capacity allocation is partially driven by Nvidia's demand signals. That alignment helps both companies avoid bottlenecks. Luna: So it's not just a contractual relationship; it's a deeply integrated operational partnership. That's hard to replicate. Lucas: Exactly. And that's the takeaway: in a world where your product is scarce, the company that can reliably deliver at scale wins. Nvidia's operations team turned a potential crisis — overwhelming demand — into a competitive moat.