TSMC founder Morris Chang
Summary
TSMC’s core moat is structural neutrality: it manufactures customers’ chips but never designs products that compete with them. Chang said he interpreted Tim Cook’s “Intel just does not know how to be a foundry” as a statement about customer trust and responsiveness, while the hosts argue that Arm enabled architecture, design, EDA and manufacturing to unbundle. At the leading edge, customer alignment compounds with scale.
NVIDIA went from a nearly bankrupt 50-to-60-person startup to a top-five TSMC customer because Chang treated small prospects as potential giants. Jensen Huang’s unsolicited 1997 letter both intrigued and “irritated” Chang because TSMC’s San Jose office had ignored it; the resulting gaming chip saved NVIDIA and generated at least the $50 million in annual business Chang considered a major account. When TSMC’s 40nm problems later damaged NVIDIA, Chang personally offered more than $100 million, valid for 48 hours, preserving a partnership that subsequently produced billions in business.
Chang’s response to the 2008 downturn was to preserve semiconductor talent and make R&D funding independent of the cycle. He offered to rehire 600-700 employees because performance ratings were subjective and because severance plus retraining erased the savings if workers were needed within a year: “If you need the people back within a year, you shouldn’t have a layoff.” Separately, he fixed R&D at 8% of revenue “regardless of whether there was a recession or not,” giving engineers confidence that multiyear node plans would survive temporary weakness.
The 28nm leadership bet combined technical conviction with a major increase in capital spending. R&D called 28nm “the sweet spot,” business development contributed market input, and Chang lifted annual capex from roughly $2 billion-$2.5 billion to nearly $6 billion in 2010, despite directors’ objections. His final answer to the board was explicit accountability: “I heard you, but I am still the person responsible for operating the company, so you need to let me go ahead with this one.”
Apple forced TSMC to balance learning-curve scale against existential concentration risk. Apple unexpectedly wanted 20nm rather than the planned progression from 28nm to 16nm, requiring billions in additional investment; after prudent financial planning and high-level advice, TSMC declined to cut its dividend or issue equity, borrowed instead and accepted only half Apple’s requested capacity. Chang rejected Jeff Williams’s suggestion to eliminate the dividend because about one-third of shareholders were seriously interested in it and the stock would “drop like hell.”
Winning Apple at 20nm carried a serious opportunity cost: TSMC lacked the R&D capacity to develop two nodes simultaneously, so 20nm delayed 16nm. Samsung skipped the detour and received Apple’s first 16nm orders, shocking Chang. Ben framed the investment as depending on converting at least 80%-90% of the 20nm equipment to 16nm; Williams promised, “As soon as you are ready with your 16 nanometers, we will buy all of our needs from you.” TSMC developed its process roughly six months later and retained most of Apple’s requirements.
The learning curve made leading-edge manufacturing look increasingly like a natural monopoly to the hosts. Chang’s refined version goes beyond “experience lowers cost”: a manufacturer forecasts mature demand, prices toward the eventual high-volume cost, fills capacity quickly and reinvests the resulting scale advantage into the next node. With new fabs costing around $20 billion and potentially $40 billion, $80 billion or $100 billion later, fewer competitors can finance the next round without TSMC’s volume.
TSMC’s Taiwan cluster turns financial scale into an operational moat that overseas fabs cannot quickly reproduce. Hsinchu places TSMC beside universities, customers and partners including Arm, Qualcomm, MediaTek, Cadence and Synopsys; older fabs remain productive for replacement parts, CMOS sensors and mature-node chips even after their original capex is depreciated. The hosts describe the semiconductor market’s growth from $26 billion when TSMC was founded in 1987 to $527 billion last year as a “ridiculous tailwind,” while distinguishing continuing demand for more computing from the technical limits of Moore’s law.
Deep dive
1. A cold letter turned NVIDIA into a top-five customer
Chang first heard from Jensen Huang through a 1997 letter sent to TSMC’s post office. Four-year-old NVIDIA had perhaps 50-60 employees, faced bankruptcy and said TSMC’s San Jose office had failed to respond.
The letter raised Chang’s curiosity and “irritated me a little bit”: TSMC already exceeded $1 billion in revenue in 1995, but he had repeatedly told salespeople never to neglect a future customer merely because it looked small.
When Chang called, Huang silenced the noisy room by shouting, “It’s quiet! Morris Chang is calling me!” Huang candidly described NVIDIA’s financial trouble but predicted its new gaming chip would save the company and make it a major TSMC customer.
Chang defined “major” as at least $50 million of annual foundry revenue. Although he disputed the hosts’ suggestion that the chip was RIVA 128, he remembered the result: within two or three years, NVIDIA became one of TSMC’s five largest customers.
2. The 40nm failure pulled Chang back into operating control
During Chang’s first attempt to hand the CEO role to a successor while remaining chairman, TSMC developed manufacturing-yield and quality problems at 40nm. Management accepted the quality director’s argument that TSMC was not at fault and offered NVIDIA nothing.
The yield problem affected every customer, including Qualcomm and Intel, but NVIDIA bore much of the damage as perhaps the node’s largest customer. It was also strategically urgent: “Only if we did 40 well could we do 28 well.”
After retaking the CEO position in 2009, Chang called every major customer. Huang remained friendly but firmly reminded him of the “quality, delivery, and manufacturing problem”; Chang requested a couple of weeks and made the dispute one of his immediate priorities.
Another warning light was financial: committed customer price reductions were arriving faster than manufacturing-cost reductions, pushing gross margin down. Fixing the node therefore required both technical recovery and restored cost discipline.
3. Chang treated layoffs as a breach of credibility, not a cost lever
The action that precipitated Chang’s return was the dismissal of roughly 600-700 employees during the 2008 financial crisis. The CEO avoided the word “layoff,” presenting the terminations as punishment for the lowest performance ratings despite Chang’s instruction to seek board approval first.
Chang’s objection was both cultural and economic. Ratings came from hundreds of different supervisors and were too subjective to command respect; moreover, severance cost about half a year’s pay and training a replacement took at least another half-year. “If you need the people back within a year, you shouldn’t have a layoff.”
More than 100 dismissed workers protested outside Chang’s home; later, about 25 camped overnight in a nearby park. At 6 a.m., his wife Sophie bought a Chinese-style breakfast, including fried bread, for them — a gesture after which they abandoned that day’s plan to demonstrate at the president’s palace.
Chang offered the employees their jobs back but did not humiliate his former CEO. He preserved the executive’s grade, salary and bonus while moving him to the new solar-cell and LED businesses; solar was undermined by Chinese subsidies, while concentrated patent ownership constrained LEDs. The executive later became vice chairman and CEO of MediaTek.
4. A $100 million-plus offer repaired NVIDIA without bargaining
Chang spent nearly half of his first four or five weeks back as CEO resolving NVIDIA’s claim. Manufacturing teams were already raising yields for TSMC and every customer; the remaining dispute was “a matter of money” and the downstream claims NVIDIA faced from its own customers.
After gathering sales intelligence and studying the damage, Chang emailed Huang that he would arrive at Huang’s home at 6 p.m. They preserved their familiar ritual — salad and pizza until 8 p.m. — before entering the study to discuss business.
Chang offered more than $100 million, effective for 48 hours. If Huang rejected it, TSMC would proceed to arbitration rather than revise the offer; Huang accepted within two days, and the repaired relationship went on to support many billions of dollars in subsequent business.
5. An 8% R&D rule made the 28nm bet possible
Chang’s benchmark came from Texas Instruments, where worldwide semiconductor R&D consumed 4.8% of revenue and his repeated requests to raise it to 5.5% were denied. At TSMC, he wanted to eliminate annual bargaining between the R&D director and CEO.
Because TSMC was already spending 6%-7%, Chang “literally picked a number”: 8% of revenue, regardless of whether there was a recession. The R&D leader regarded this as the best thing TSMC had done for R&D because planned people and programs were no longer vulnerable to yearly cuts.
That certainty encouraged bigger technical claims. Engineers told Chang that 28nm would be “the sweet spot,” like striking a tennis ball at the center of the racket; technical reasoning, market forecasts and strategic business-development input persuaded him to believe them.
The accompanying capex jump — from roughly $2 billion-$2.5 billion annually to nearly $6 billion in 2010 — alarmed directors. Invoking Shakespeare’s “tide in the affairs of men,” Chang accepted responsibility: “I heard you, but I am still the person responsible for operating the company, so you need to let me go ahead with this one.” The node coincided with the smartphone era.
6. Functional organization and a small business-development role shaped succession
Chang rejected organizing TSMC into divisions because advanced and mature fabs often served the same customers. Qualcomm and Apple used leading-edge processes for flagship chips but older processes for numerous supporting components, making fab-by-fab ownership unnatural.
An earlier president, Don Brooks, had proposed dividing the company by fab. McKinsey spent about a month or two and, Chang recalled, perhaps a couple million dollars confirming that a functional structure was best. Its analogy was Boeing: although Boeing had commercial and government divisions, it did not divide the company by individual aircraft models.
Chang allowed CEO Rick Tsai to split operations into advanced and mainstream groups in 2005 because “sometimes you have to let the CEO make his own mistakes and learn from them.” On returning, Chang recombined operations and created one strategic business-development organization.
Mark Liu declined that roughly 60-to-70-person group because he already led about 10,000 advanced-technology employees. C.C. Wei accepted the role quite delightedly; Chang viewed the smaller position as essential marketing and business-development training. Wei is now chairman and CEO of TSMC.
7. Apple arrived uninvited — and requested a 20nm detour
Chang knew Apple was closed-mouthed: soliciting it would produce rejection, and “they will come to see you when they are ready.” The opening came when Foxconn founder Terry Gou, Sophie Chang’s second cousin, unexpectedly brought Apple COO Jeff Williams to dinner.
Williams skipped most pleasantries and delivered perhaps 80% of the conversation. He said Apple would “let us have a 40% gross margin,” apparently intending generosity; Chang did not respond because TSMC already earned 45% and was trying to reach 50%.
Chang expected Apple to want the newly developed 28nm process. Williams instead answered “20,” surprising and disappointing him because the planned Moore’s-law progression was directly from 28nm to 16nm.
TSMC then lacked the resources to develop two nodes concurrently. Chang’s immediate framing was that 20nm was a “half step” and “a detour”: it would contribute knowledge toward 16nm, but would consume time and capital outside the existing roadmap.
8. TSMC funded only half Apple’s demand and protected the dividend
Chang had cultivated Goldman Sachs well before TSMC needed high-level financing advice, including through its New York ADR listing. Apple’s request created precisely that need: the company had already expanded for 28nm and now required billions more for an unplanned node.
Management considered cutting the dividend, issuing shares in Taiwan or the United States, borrowing and accepting less demand. After prudent financial planning, it chose debt, no equity issuance, no dividend cut and capacity for only half Apple’s requested volume.
Apple’s lower-level purchasing people initially told C.C. Wei, “You must be crazy.” Chang then presented Williams with the result of “prudent financial planning”; Williams suggested eliminating the dividend, but Chang refused because roughly one-third of shareholders were seriously interested in it and the stock would “drop like hell.”
TSMC had learned to demand customer deposits when capacity was scarce because “it’s our money and it’s only their word.” Salespeople delighted in warning that TSMC might “confiscate” deposits, though it never did; after the internet recession around 2000-2001, customers delayed orders and eventually every one of them used its deposit.
9. Intel lost Apple on foundry behavior, not just process technology
In February 2011, Williams paused TSMC discussions for two months because Intel CEO Paul Otellini had approached Tim Cook. Intel already supplied processors for all Macs, giving it an existing relationship with Apple.
Chang was less frightened than he would have been in Intel’s 1990s prime. His assessment put TSMC approximately level on technology, ahead in manufacturing and, crucially, ahead in customer trust.
One month into the pause, Chang requested an update. Williams agreed but said he would be away and had asked Cook to receive him — an unusual escalation because executives normally delegate visitors downward, not upward.
Over cafeteria lunch, Cook gave the concise verdict: “Intel simply did not know how to be a foundry.” Chang interpreted this as a service-culture issue: TSMC responded courteously even to “crazy” or irrational requests, whereas Taiwanese PC makers dealing with Intel all wished for another supplier and none liked or trusted Intel.
10. Apple’s 20nm win delayed 16nm — and briefly handed Samsung the lead
Pricing still required negotiation. Chang said cost was the primary input, followed by whether the desired price would be accepted by the customer.
The serious trade-off emerged later: because TSMC could not develop two nodes simultaneously, 20nm delayed its 16nm process. Samsung, having lost the 20nm business, skipped that work and moved ahead at 16nm.
Chang was “really shocked” when Apple placed its first 16nm orders with Samsung. Ben framed TSMC’s investment as depending on converting at least 80%-90% of the 20nm equipment to 16nm; losing Apple would have threatened the logic behind that investment.
Williams immediately promised to visit and explained, “As soon as you are ready with your 16 nanometers, we will buy all of our needs from you.” TSMC developed its process roughly six months later, and most of Apple’s 16nm requirements remained with TSMC.
11. Qualcomm’s move exposed IBM’s weakness before IBM admitted it
TSMC considered Qualcomm a prime fabless prospect throughout the 1990s, but Qualcomm’s operations vice president remained polite while awarding little business. Chang’s intelligence indicated that IBM was its main foundry.
Around 1997-98, Qualcomm began shifting substantial volume to TSMC. Chang inferred that IBM Semiconductor was in trouble because its main business was supplying Qualcomm and a few other small fabless companies.
IBM’s next move therefore caused no surprise: in 1999 it asked TSMC to co-develop the 0.13-micron, or 130nm, generation. Chang refused immediately; he expected TSMC would have to send engineers to IBM and risk becoming dependent on a joint process rather than preserving its own development capability.
IBM reacted angrily and approached UMC, which accepted and, Chang said, “seriously regretted” doing so. His broader observation was that sitting at the foundry junction let TSMC see customer migrations and competitive deterioration before their consequences became obvious elsewhere.
12. Learning-curve discipline turned volume into strategy
Around 1970, Boston Consulting Group founder Bruce Henderson presented the experience curve to TI chief Mark Shepherd. Shepherd assigned Chang to work with Bill Bain, who spent roughly three days a week in an office near Chang’s in Dallas for about two years gathering semiconductor cost, price and product-family data.
Chang did not claim to invent the learning curve, but said he helped refine it until a semiconductor company could use it effectively. Bain later told him he was leaving BCG to create his own firm because there was a “world imperative”; Chang said that apparently meant a personal imperative.
The simple version is that producing more units creates experience and lowers unit cost. Chang warned that anyone who stops there “really hasn’t learned anything”; the useful version connects cumulative volume, pricing, capacity and competitive strategy.
The hosts’ formulation works backward from mature scale: forecast the node’s total demand, price early toward its eventual low cost, move through low volume quickly, aggregate customers and reinvest the advantage. The danger is forecasting Apple’s iPhone or NVIDIA’s AI demand incorrectly by even 5%-10%, impairing profitability and the next node’s funding.
13. Chang saw fabless demand before TSMC’s customers existed
In Chang’s final months at General Instrument, Gordon Campbell requested $50 million to start a semiconductor company but had no written business plan. Three weeks later Campbell said he needed only $5 million: “I’m not going to build a fab.” Chang called that his first glimpse of the fabless future.
Another early entrepreneur had already started a company called Atmel and sought access to General Instrument’s idle fabs without owning manufacturing. The resulting argument was the foundry model in miniature: every designer wanted the fab run its way, while the owner could operate it only one way that would satisfy many customers “more or less.”
Asked whether TSMC was an unlikely success, Chang conceded that its size and importance exceeded his expectations, “but not by an order of magnitude.” Nor did he seriously plan to stop after two fabs; as “a serious student of the learning curve,” he knew scale required continued expansion.
The hosts emphasize that early TSMC nevertheless survived on integrated manufacturers’ excess, less strategic orders while waiting for fabless demand to emerge. That unattractive business accumulated real factories, operating skill, volume and capacity, leaving TSMC ready when independent chip designers arrived.
14. Pure-play neutrality worked because the value chain unbundled
The hosts’ central postgame conclusion is “not competing with your customers.” TSMC neither designs chips nor sells end products, so NVIDIA, Apple and Qualcomm can give it their most important designs without financing a direct rival elsewhere in the value chain.
They argue that Arm was indispensable to this outcome. Had x86 and fully integrated Intel remained dominant, there might have been little space for independent designers and a neutral leading-edge manufacturer; Arm let CPU architecture separate from chip design and fabrication.
That unbundling produced a mutually dependent group of standalone architecture, manufacturing, EDA and chip-design companies. The hosts call Arm and TSMC “coupled at the hip of history.”
Apple is their largest counterfactual: without a viable Arm platform and Apple’s standardization on Arm-based Apple Silicon, Intel might have remained positioned to manufacture the leading chips in Apple devices. Instead, the value chain reorganized around specialized firms whose economics reward collaboration.
15. Hsinchu’s cluster and relentless reinvestment deepen the lead
Visiting Hsinchu Science Park made the ecosystem physical for the hosts: TSMC sits beside Arm, Qualcomm, MediaTek, Cadence, Synopsys and two universities producing PhDs who are absorbed into the industry. Flights into Taiwan looked like “technology buses” filled with people heading to the semiconductor ecosystem.
They doubt overseas fabs can recreate that density quickly. Arizona capacity may serve customer and government objectives, but reproducing Taiwan’s network of partners, research, talent and communication would take decades. The hosts also saw the enormous phased facility where TSMC was preparing its 2nm process for small production runs and a planned ramp in the second half of the year.
TSMC preserves older fabs rather than continually converting everything to the newest process, as Intel historically did. Mature facilities still make replacement parts, CMOS image sensors and chips that do not need leading-edge geometry; once initial capex is depreciated, they can generate very high-margin revenue after maintenance costs.
The hosts describe advanced fabrication as having “natural monopoly characteristics”: a current fab costs roughly $20 billion, and future rounds might demand $40 billion, $80 billion or $100 billion. They connect that reinvestment to the semiconductor market’s growth from $26 billion in 1987 to $527 billion last year and to the continuing demand for more computing.