Latest / The Productivity Podcast with Fexingo: Output, Efficiency, and Long-Term Economic Growth / How Taiwan's TSMC Built the World's Most Productive Chip Foundry
Transcript
- Lucas: If I asked you to name the single most productive manufacturing operation on the planet, what would you say? Luna: I'd probably default to something like a Tesla gigafactory or maybe the Port of Shanghai. But I'm guessing you have a different answer. Lucas: I'd put Taiwan Semiconductor Manufacturing Company — TSMC — near the top. Specifically, its Fab 18 in Tainan, which cranks out the world's most advanced chips on 3-nanometer and 5-nanometer nodes. And what's wild is, it's not just one fab. TSMC runs four so-called 'giga-fabs' in Taiwan, each with a capacity of over 100,000 wafer starts per month. The company alone produces more than 90 percent of the world's most advanced logic chips. Luna: Ninety percent — that's almost a monopoly. How did a contract manufacturer get there? Lucas: The key decision was made in 1987 by Morris Chang, who founded TSMC as a pure-play foundry. Before TSMC, chip companies like Intel and AMD both designed and manufactured their own chips. Chang believed there was a business in just manufacturing other people's designs — and that by focusing exclusively on that, TSMC could out-invest anyone in process technology. Luna: So the bet was that specialization would drive higher productivity than integration. Lucas: Exactly. And it paid off in a way that even Chang probably didn't foresee. Because TSMC doesn't compete with its customers — it makes chips for Apple, NVIDIA, AMD, Qualcomm — those companies trust it not to build its own products. That trust creates a virtuous cycle: more customers mean more volume, more volume means more data, more data means faster yield improvements, and faster yields mean lower costs per chip. Luna: Yield improvement is the hidden productivity metric in chipmaking. Walk me through how that works. Lucas: A chip fab costs around $20 billion to build — Fab 18 alone was around $20 billion. But once it's built, the real competitive edge is yield: the percentage of chips on a wafer that actually work. When you're making chips with billions of transistors, even a tiny defect rate matters. TSMC's yield on a new node might start at 60 or 70 percent, but through relentless process tweaks, they push it above 90 percent within a year or two. Luna: And that bump from 80 to 90 percent effectively doubles the number of good chips per wafer. That's a massive productivity gain. Lucas: Massive — and it compounds. Because every percentage point of yield improvement drops the cost per good chip below competitors who are stuck at lower yields. That's why TSMC commands a 90 percent market share in the most advanced nodes. Intel, for all its engineering talent, has struggled to match TSMC's yield on its own 7-nanometer and 5-nanometer processes. Luna: Intel recently announced it would separate its foundry business from its design business. Is that an attempt to copy TSMC's model? Lucas: It is, but they're decades behind. TSMC has been perfecting the pure-play model for almost 40 years. And the scale difference is staggering. In 2025, TSMC spent about $36 billion on capital expenditure and R&D combined. Intel spent roughly half that on its foundry efforts. When you're investing $36 billion a year, you can afford to develop extreme ultraviolet lithography tools that cost $150 million each, and you can buy hundreds of them. Luna: EUV lithography — that's the technology that pushed Moore's Law forward past 7 nanometers, right? Lucas: Right. Each EUV machine can pattern incredibly fine features on a chip. But they're incredibly complex — one machine contains over 100,000 parts and requires a team of engineers just to keep it running. TSMC had the scale to integrate these machines into high-volume production faster than anyone else. By the time Intel was ramping EUV, TSMC had already printed millions of wafers on the same tools. Luna: So productivity here isn't just about worker output per hour. It's about capital productivity — getting the most value out of a $150 million machine. Lucas: Exactly. And that's where TSMC's operational culture comes in. The company runs its fabs 24 hours a day, 365 days a year. The manufacturing process is broken down into thousands of steps, each optimized for speed and consistency. A single wafer might travel through over a thousand processing steps over a period of three months. The coordination required is mind-boggling. TSMC's yield management software tracks every single wafer in real time. Luna: I've read that TSMC's engineers have access to decades of process data — billions of data points from previous production runs. That's a moat that's nearly impossible to replicate. Lucas: It's the ultimate learning curve. Each time TSMC runs a new product, they collect data on every defect, every variation. That data feeds back into the process, reducing defects on the next run. Competitors like Samsung have to start from scratch on each new node, while TSMC builds on a mountain of experience. Luna: And that learning curve is compounded by the sheer geography of TSMC's operations — most of its advanced fabs are within an hour's drive of each other in central Taiwan. That's an efficiency that Intel, with fabs scattered across the US, Europe, and Israel, can't match. Lucas: A huge advantage. TSMC can move engineers and equipment between fabs quickly. If one fab has a yield issue, experts from another fab can be on site within hours. That geographic concentration also means shared infrastructure — water recycling plants, power grids, even a specialized workforce trained at nearby universities. Luna: But there's a vulnerability too: Taiwan's geopolitical risk. If China were to blockade the island, the global chip supply would grind to a halt. Lucas: That's the biggest risk investors and policymakers worry about. TSMC has started building fabs in Japan, Arizona, and Germany, but those are years away from producing leading-edge chips. And even when they do, they'll operate at higher costs and lower yields initially. The productivity advantage is deeply tied to Taiwan itself. Luna: So the productivity miracle of TSMC is fragile. It relies on peace, talent density, and decades of accumulated know-how. Lucas: And that's what makes the company such a fascinating case study. Productivity isn't just a matter of better machines or smarter algorithms. It's about organizational choices — the pure-play model, the obsessive focus on yield, the clustering of operations, the long-term investment horizon. TSMC spent years developing a single node before seeing a dollar of revenue. Most companies can't stomach that. Luna: Before we go further, I want to pause on something you just said — that the learning curve is built on decades of data. It reminds me of something I've been thinking about lately, which is how much we take for granted that this kind of deep knowledge is sustained by people who've stayed at the same company for decades. TSMC's average engineer tenure is something like 10 years. That's rare in tech. Lucas: Yeah, and that institutional memory is a huge part of their productivity edge. And honestly, if walking through the economy with us has made something click — maybe about TSMC, maybe about how capital and knowledge compound — the best way to keep this podcast ad-free is listener support. You can find it at buy me a coffee dot com slash fexingo. Just the smallest token, if today was worth a cup to you. Luna: Absolutely. No pressure, just a link for those who find value here. And Lucas, I want to go back to that yield curve — how does TSMC actually measure productivity at the individual fab level? Lucas: They use a metric called 'die yield per wafer' and also 'overall equipment effectiveness.' OEE measures how much of the time a machine is actually producing good chips versus idling or producing defective ones. TSMC's OEE is notoriously high — over 90 percent for most critical tools. In comparison, many other fabs hover around 70 to 80 percent. That difference directly translates to lower cost per chip. Luna: And these machines run continuously. So a 10 percent improvement in OEE is worth hundreds of millions of dollars annually. Lucas: Easily. And TSMC's culture is obsessive about eliminating downtime. They have 'just-in-time' maintenance schedules, predictive analytics on tool wear, and even dedicated teams that analyze every minute a tool is not producing. It's as much a logistics operation as a manufacturing one. Luna: It sounds almost like Toyota Production System principles applied to chipmaking. Lucas: Exactly the analogy. Continuous improvement, waste elimination, respect for the operator — TSMC's founder Morris Chang was heavily influenced by Japanese manufacturing practices. But he adapted them to a context where the 'product' is constantly changing. Every new chip design is a different product, so the process has to be incredibly flexible. Luna: That flexibility is another productivity driver. TSMC can run hundreds of different designs through the same fab simultaneously because its equipment is highly standardized and software-controlled. Lucas: Right. And they've pushed that flexibility to the extreme. The same fab can produce chips for an iPhone, a GPU for AI training, and an automotive microcontroller all in the same week. That mix is only possible because of sophisticated scheduling algorithms that optimize for overall throughput, not just volume of a single product. Luna: It's a different kind of productivity: not just getting more of one thing, but getting more total value from a fixed set of capital assets. Lucas: Bingo. That's the core insight. TSMC's productivity isn't just about making chips faster. It's about making the most profitable mix of chips, with the highest yields, using the most capital-efficient tools. And they've been perfecting that for nearly four decades. Luna: So what's the next frontier? 2-nanometer? 1.4-nanometer? Lucas: TSMC is already in development on 1.4-nanometer, and they're exploring beyond that. But each new node costs more to develop and requires new materials, new tooling. The productivity gains are getting harder to extract. At some point, the cost per transistor stops falling. Then productivity will have to come from new architectures — like chiplets or 3D stacking — where you stitch together multiple smaller chips into a single package. Luna: And TSMC is investing heavily in advanced packaging, which is effectively another manufacturing process entirely. Lucas: They're building a dedicated advanced packaging fab in Taiwan. That's a bet that the next wave of productivity won't come from shrinking transistors alone, but from smarter integration. If they're right, they'll extend their lead. If they're wrong, someone else might leapfrog. But given their track record, I wouldn't bet against them. Luna: It's remarkable how a single strategic bet — pure-play foundry — created a productivity machine that now underpins the entire global tech industry. Lucas: And that's the lesson for productivity in any industry: focus on what you can be best at, invest in learning, and build a culture that treats every defect as a data point. TSMC didn't become the world's most productive manufacturer by accident. They engineered it, one wafer at a time.