holocron
中/en
← Business
Business

Learning Curve

One line: every time cumulative production doubles, unit cost falls by a predictable, fixed percentage — Morris Chang absorbed this theory working alongside BCG at Texas Instruments around 1970 and used it for the rest of his career: "I absorbed a lot of learning curve stuff, which I used up to now. I found that highly fruitful as a thinking tool." It became the operating system behind TI's and TSMC's pricing, capacity expansion, and the bet-the-company decision to take Apple's full order — and, in Berkshire 之一：Buffett 合伙基金年代（1930-1970）, it's also the defining feature of exactly the kind of business Warren Buffett ruled out of his circle of competence, at the cost of missing Intel.

The Mechanism

  • Origin: starting from refrigerators and cars. Around 1970, BCG founder Bruce Henderson called TI CEO Mark Shepherd directly, arguing that experience-curve theory would benefit the semiconductor industry; Bill Bain then spent roughly three days a week at TI for close to two years, working alongside Morris Chang — who'd been assigned as TI's counterpart — to refine the theory into something a semiconductor company could actually use (Morris Chang 访谈：张忠谋亲述 TSMC). Morris's own explanation, given to a "novice": "The simple explanation of learning curve is that as you make more of one thing, anything — actually started with refrigerators and cars — if a company makes more cars, then its cost per unit car goes down. That's why it's also called experience curve. You gain more experience, you become more efficient." But he immediately added the qualifier that runs through every case in this page: "one will be foolish if one just takes the simple explanation and thinks that that's all it is."
  • On the manufacturing side: yield is learned. A new process starts with low yield by necessity — "there's a learning curve to getting the yields right and learning how to manufacture a new process" — a pattern the TSMC episode traces back to Morris's first win at TI in 1958: taking the IBM 7090 production line's yield from 0% to 20% in about four months, twice IBM's own line (10%) (TSMC：纯代工模式发明者).
  • On the pricing side: price is a lever for accelerating learning, not a way to recoup cost. Industry convention charged a premium on new product lines to recoup fab investment up front. Morris and BCG inverted it: "Start low and then continually automatically reduce the price every quarter even when the market did not demand it. … A lot of people thought we were being foolish. … But we did it because we believed in it, and indeed our market share just kept expanding." The mechanism is a loop: price high → demand stays low → the fab never runs at capacity → learning is slow, yield climbs slowly → unit cost never comes down — a vicious cycle. Price low → volume grows → the fab runs at capacity → learning is fast, yield climbs fast → unit cost falls fast → share keeps expanding — a virtuous one (TSMC：纯代工模式发明者).
  • Working backwards from the endgame. The hosts give the fullest reconstruction of the logic in the Morris Chang interview: the goal is to be the highest-volume player once the game is over, so you cut prices toward that endgame level today — even pricing below cost at the start — to crowd out competitors and aggregate all the demand; scale then funds the cost reduction and the next generation's R&D. This single insight drives pricing, strategic finance, and every debt decision — "almost Costco-like in the ballet" (Morris Chang 访谈：张忠谋亲述 TSMC).
  • The downside is just as extreme as the upside. The manufacturer has to bet on its customers' end markets — how many next-generation iPhones will sell, how big AI will get — because "if you're off by 5%-10%, that's going to tank your entire profitability for that node generation, which is going to tank your free cash flow, which is going to mean you can't play the game in the next turn." (Morris Chang 访谈：张忠谋亲述 TSMC)
  • The curve's endgame shape: stacked with Moore's Law and Rock's Law, it converges to natural monopoly. The TSMC episode sets the learning curve inside a larger industry dynamic: Moore's Law (transistor count doubling, the OpEx side) compounding with Rock's Law (fab cost doubling roughly every four years, the CapEx side) drives the cost threshold up exponentially while the number of players who can keep pace shrinks exponentially — whoever runs fastest down the learning curve naturally converges into the monopolist (TSMC：纯代工模式发明者; see 7 Powers 护城河框架 for the Scale Economies / Process Power verdict this feeds).

Cases in This Library

The Morris Chang interview: a lifelong thinking tool, and the calculator behind an endgame bet

Throughout the interview Morris insists he wasn't the theory's originator, just the person who "had a role in refining it to the point where a semiconductor company can use it effectively" — and that framing directly drove his read on TSMC's expansion pace. Asked whether the original plan was to stop after Fab 2, he pushed back: "We were just talking about learning curve. How could we plan to? If I didn't know anything about learning curve, I would say, yeah, maybe we'll stop after two. But I was a serious student of learning curve and I would never stop at just two, perhaps." The hosts summarize the whole logic as a fully backwards-planned operating system (see the Mechanism section above), and use it directly to explain the risk calculus of taking Apple's entire order: "The example from Apple. We are about to go get the absolute whale customer and we have to balance taking on all of their order, which the learning curve would tell you, you want to get the deepest down the learning curve possible, we should go take all their order. That exposes you to existential risk in your business when you're not within spitting distance of doing that volume on your own. So is it really worth betting the entire company?" Morris is unsentimental about the stakes: "Yeah, I know, bet the company. But I didn't think I would lose." (all Morris Chang 访谈：张忠谋亲述 TSMC)

The TSMC company history: pricing as a weapon, and an endgame for the whole industry

The company-history episode traces the same logic back from personal thinking tool to corporate weapon. After Morris was made a TI division general manager in 1967, he noticed the industry's habit of charging premium prices on new fab-dependent product lines to recoup investment — which suppressed demand and slowed the yield climb. He and the then-tiny BCG did the opposite: the same "learning curve pricing" quoted above, priced low from the start with automatic quarterly cuts. It made TI's IC business the largest and most profitable in the world, ahead of Fairchild, National, and the newly founded Intel. Three decades later the episode reapplies the same logic to explain the endgame of the foundry industry itself: under the compounding exponentials of Moore's Law and Rock's Law, "There is no amount of money you could spend to catch up next year. You can't because engineering is so hard and the learning curve takes decades to get to this point." The episode uses this to explain why money alone can't let SMIC catch up, and why the field of leading-edge players collapsed from 22 down to two (TSMC：纯代工模式发明者).

Berkshire Part I (the investor's view): the exact kind of business Buffett ruled out of his circle of competence

Neither "learning curve" nor "experience curve" appears anywhere in this episode itself — but it supplies a reliable mirror case. In his 1967 letter to partners, Buffett laid down a new rule: "We will not go into businesses where technology, which is way over my head, is crucial to the investment decision. I know about as much about semiconductors or integrated circuits as I do about the mating habits of this chrząszcz [beetle]. This is very unfortunate." On the strength of that rule, he approved Grinnell College's endowment committee investing $100,000 in Intel's seed round — while keeping his own partnership, and his own money, entirely out of it, missing what became a company worth hundreds of billions. The rule held for 45-plus years, broken only for Apple, and even then on Todd Combs's initiative rather than Buffett's own. Semiconductors are close to the archetypal learning-curve battlefield: they require betting simultaneously on whether the technology will pan out and on whether demand can be forecast precisely — precisely the kind of wager Buffett's entire method (margin of safety, only invest in what you can understand, avoid technology judgment calls) is built to rule out structurally (Berkshire 之一：Buffett 合伙基金年代（1930-1970）).

Discussion and Boundaries

  • The simple explanation isn't the understanding. Morris makes this point twice in the interview: "cost per unit falls with cumulative volume" is the entry-level version; the real use of it is turning that fact into a complete operating system for pricing, capacity, and endgame judgment — anyone who stops at the simple explanation "really hasn't learned anything" (Morris Chang 访谈：张忠谋亲述 TSMC).
  • The curve doesn't turn itself — it needs active pricing to drive it. Ben's supplementary framing in the TSMC episode: the industry originally passed its own cost structure straight through to customers without realizing the operating leverage sitting on the table; the right move was to eat the fixed cost up front and earn it back on the margin of every chip that came off the line at volume (TSMC：纯代工模式发明者).
  • The upside and downside are two faces of the same coin. The positive case in this material (TSMC's share gains, its eventual monopoly) and the negative frame (Buffett's deliberate avoidance) share the same root cause: the strategy requires betting on an unverifiable future — end-market demand, whether the technology pans out. That's the source of Morris's confidence ("bet the company, but I didn't think I would lose") and equally the reason Buffett stayed out entirely ("technology... way over my head"). The two episodes reach opposite verdicts on the same underlying business model; this page records both without adjudicating which is right.
  • Honest cross-episode note. The "investor's view" link to Berkshire is this page's own inference drawn from two separate episodes, not a connection the show itself makes on the record — that layer is editorial synthesis, not an Acquired quote.