The Dinner That Built the Modern World
In the autumn of 2010, Jeff Williams — Apple's operations chief, a man whose job was to make the impossible logistically routine — flew to Taipei and drove south to the home of a seventy-nine-year-old semiconductor executive. They had dinner. Morris Chang and his wife Sophie hosted. Apple and TSMC were not doing business together at the time. The conversation was exploratory, the stakes almost comically understated for what would become perhaps the most consequential supplier relationship in the history of consumer electronics. What Williams wanted was simple to articulate and nearly impossible to execute: leading-edge process technology at established-technology volumes. What Chang wanted was the thing he had spent his entire career constructing the conditions for — a customer whose ambitions were large enough to justify bets no rational financial model would sanction. Within four years, TSMC would invest $9 billion to bring up a single fab in Tainan in a record eleven months, staffing 6,000 engineers around the clock to deliver Apple's A8 processor. There was no backup plan. Apple committed 100% of its application processor volume. TSMC committed to the kind of capital expenditure that most countries cannot manage. The bet worked — flawlessly, as Williams would later describe it at TSMC's thirtieth anniversary in October 2017. By then, the two companies had shipped over half a billion chips together, and TSMC had invested $25 billion in the relationship. But the dinner matters less for the deal it produced than for what it revealed about the architecture of the company Chang had spent twenty-three years building before that evening. TSMC did not win Apple because it had the best technology — Intel arguably did, at that moment. It won because it had designed an entire business model around a single, radical premise: the most valuable thing a semiconductor manufacturer can offer is the promise that it will never compete with its customers. That promise, simple enough to fit on a napkin, required an organization of extraordinary discipline to keep. It required declining to design chips when the margins on design were intoxicating. It required investing tens of billions of dollars on behalf of customers who could, in theory, walk away. It required a culture so fanatically oriented toward manufacturing excellence that the company would operate "Nightingale Armies" — three-shift R&D teams working through the night, earning 30% salary premiums and 50% dividend bonuses — to hold a two-year technology lead over Samsung. And it required a founder whose biography was, in some essential way, the biography of the semiconductor industry itself.
By the Numbers
TSMC at a Glance (FY2025)
$90B+Estimated annual revenue (FY2024 ~$87.1B; FY2025 growth ~30% guided)
56%+Gross margin (sustained at leading-edge nodes)
~95%Market share in sub-3nm logic manufacturing
534Customers served in 2025
12,682Products manufactured across end markets
17M+12-inch equivalent wafer annual capacity
$52–56BPlanned capital expenditure for 2026
~$900B+Market capitalization (as of early 2026)
A Career Measured in Nanometers and Decades
Morris Chang was born in 1931 in Ningbo, China, into a world that was, as he would later write in his memoir, defined by "war, poverty, and injustice." His childhood was a series of dislocations — fleeing the Sino-Japanese War, watching his country tear itself apart, finally leaving for the United States at eighteen, arriving at Harvard in 1949 with the particular determination of someone who has seen civilization's infrastructure collapse and decided to spend his life building things that cannot easily be destroyed.
He transferred to MIT, studied mechanical engineering, and then — in a turn that reads like a parable about the semiconductor industry's own accidental origins — failed to negotiate a satisfactory salary with Ford Motor Company. "Frustrated and a bit embarrassed," as he put it, he chose a different path. He joined Sylvania Semiconductor in 1955. The frustration with Ford was, in retrospect, the most consequential salary negotiation in the history of technology.
Chang spent the next quarter-century at Texas Instruments, rising to senior vice president — the highest-ranking non-founder in the company, running a semiconductor business generating over $1 billion annually. He was brilliant at the technology, better at the business. He understood, with the clarity of someone who had watched an entire industry emerge from the physics of silicon, that the economics of semiconductors were fundamentally different from the economics of anything else humans manufactured. The cost structure was dominated by the fixed investment in fabrication; the marginal cost of each additional chip was negligible. This meant that scale was not just an advantage — it was the only advantage that mattered over time. And it meant that any company forced to spread its fabrication investment across both design and manufacturing was structurally disadvantaged against a company that could aggregate the manufacturing demand of an entire industry.
He carried this insight to Taiwan in 1985, recruited by the government to run the Industrial Technology Research Institute (ITRI). K.T. Li, the minister without portfolio who was Taiwan's de facto industrial strategist, had a vision for a semiconductor industry. Chang had something more specific: a business model.
I had the idea for a pure-play foundry in 1985. Nobody else in the world had the same idea. The key idea was: we would be everyone's foundry. We would compete with nobody.
The Business Model as Competitive Weapon
The idea was, in 1987, genuinely radical. To appreciate how radical, consider what the semiconductor industry looked like at that moment. Integrated device manufacturers — Intel, Texas Instruments, Motorola, NEC, Toshiba — designed their own chips and manufactured them in their own fabs. The concept of a "fabless" chip company barely existed as a category; it was more of an aspiration held by a few venture-backed startups in Silicon Valley who couldn't afford fabs and had to beg IDMs to sell them excess manufacturing capacity. The IDMs, naturally, treated this overflow business as an afterthought. They would fill fabless orders when they had spare capacity and bump them the moment their own products needed the line. Delivery times were unpredictable.
Quality was inconsistent. The message was clear: if you don't own a fab, you're a second-class citizen in semiconductors.
Chang proposed to build a company that manufactured chips for other companies and nothing else. No proprietary chip designs. No competing products. Just manufacturing, performed at the highest possible quality, for anyone who would pay. The pure-play foundry.
The initial reaction ranged from skepticism to derision. Who would build a multibillion-dollar fabrication facility to serve companies that didn't yet exist at meaningful scale? Intel certainly wasn't going to outsource manufacturing. The Japanese giants had their own fabs. The few fabless startups that existed were tiny. Chang was proposing to build the world's most capital-intensive hotel and hoping guests would show up.
What made the bet rational — barely — was a structural reading of where the industry was heading. Chang understood that the cost of building a leading-edge fab was escalating on a curve that would eventually exceed the ability of any single company's product revenue to justify the investment. In 1987, a state-of-the-art fab cost perhaps $100 million. Within a decade it would cost $1 billion. Within two decades, $5 billion. By 2026, TSMC would be planning to spend $52 to $56 billion in a single year on capital expenditure — more than the
GDP of roughly half the world's nations. The only way to justify that spending was to spread it across the demand of the entire industry. The pure-play foundry, in other words, was not a niche play. It was the inevitable organizational form of semiconductor manufacturing, and Chang had seen it twenty years before anyone else.
Taiwan's government provided the initial capital — roughly $220 million — and Philips, the Dutch electronics conglomerate, contributed both money and technology, taking a 27.5% stake and licensing its process technology. Chang personally recruited engineers from ITRI, the research institute he had been running, effectively spinning the company out of the government's technology development apparatus. The founding team included people like Y.P. Chyn, who joined in 1987 and would still be at the company nearly four decades later as Executive Vice President and Co-Chief Operating Officer — a tenure that itself tells you something about TSMC's culture of institutional loyalty.
Key milestones in TSMC's founding era
1985Morris Chang arrives in Taiwan to lead ITRI; conceives the pure-play foundry model.
1987TSMC founded on February 21 in Hsinchu Science Park with backing from the Taiwan government and Philips.
1988Fab 1 begins production using 3-micron process technology — already two generations behind the leading edge.
1993TSMC lists on the Taiwan Stock Exchange (TWSE: 2330).
1994Revenue reaches $1 billion for the first time; the fabless ecosystem begins to emerge in earnest.
1997TSMC lists ADRs on the New York Stock Exchange (NYSE: TSM).
The early years were unglamorous. TSMC's first processes were not cutting-edge. The company survived on a diet of small orders from fabless startups that had nowhere else to go, supplemented by overflow work from IDMs during cyclical capacity crunches. The technology gap with Intel and the Japanese giants was real. But Chang was patient in a way that the quarterly earnings cycle does not reward — he was building a platform, not a product, and platforms take time to develop the gravitational pull that makes them indispensable.
The Ecosystem Grows Its Own Gravity
The flywheel that would make TSMC the most important company in the semiconductor industry — arguably, by the 2020s, the most important company in the world — took roughly a decade to begin spinning visibly. Its logic was elegant and, once understood, almost frighteningly difficult to reverse.
Each new fabless design company that outsourced to TSMC added volume to TSMC's fabs. More volume meant more revenue to reinvest in next-generation process technology. Better process technology attracted more design companies, because a foundry with the best manufacturing processes could offer chips with superior performance and power efficiency. More design companies meant more volume, and the cycle repeated. But there was a second, subtler loop running underneath the obvious one: as more companies designed chips on TSMC's processes, the design ecosystem — EDA tools, IP libraries, design kits — became increasingly optimized for TSMC's specific manufacturing capabilities. Synopsys, Cadence, and the other EDA companies invested disproportionately in making their tools work perfectly with TSMC's process design kits. This created enormous switching costs. A chip designer who had validated an entire design flow on TSMC's 7nm process couldn't simply move to Samsung's 7nm process without months of redesign and revalidation. The manufacturing platform had become, in software terms, an operating system.
By the mid-1990s, the fabless model was no longer an experiment. Companies like Qualcomm, Broadcom, Marvell, and MediaTek were designing increasingly complex chips and outsourcing all manufacturing to TSMC and its smaller rival UMC. Nvidia, founded in 1993 by
Jensen Huang, would become one of TSMC's most important customers and, three decades later, one of the most valuable companies on earth — a company with 30,000 employees and a market capitalization above $3 trillion, achieving what one analyst calculated as over $100 million in value per employee, a figure possible only because it owned zero manufacturing capacity and relied entirely on TSMC's fabs.
The symbiosis was — and is — total. Without TSMC, Nvidia's GPU designs are blueprints. Without Nvidia and Apple and Qualcomm and AMD and Broadcom and the hyperscalers, TSMC's fabs are empty clean rooms burning cash. But the asymmetry in the relationship is real and growing: there are many fabless chip designers, and there is one company that can manufacture their most advanced designs at commercial yields. TSMC's customers need TSMC more than TSMC needs any individual customer, and every year the technology gap widens, this asymmetry deepens.
The [Trust](/mental-models/trust) Premium
The foundry model contained a paradox that Chang recognized from the beginning and spent decades resolving: how do you convince the world's most secretive technology companies to hand over their most valuable intellectual property — their chip designs, the crown jewels of their competitive advantage — to a third party?
The answer was trust, engineered as a structural feature of the business model rather than a personal quality of the leadership. TSMC's commitment to never design its own chips was not merely a strategy; it was an existential covenant with its customers. Every semiconductor company that outsourced to TSMC needed the absolute certainty that TSMC would never use the knowledge gained from manufacturing their chips to compete against them. Intel's foundry efforts, by contrast, would be permanently handicapped by the fact that Intel was simultaneously a chip designer — a fox offering to guard the henhouse.
The 40nm crisis of 2009 demonstrated what this trust was worth in practice. TSMC's 40nm process had severe yield problems — chamber matching issues were producing wafers with yields reportedly as low as 20%, devastating for Nvidia and AMD, whose product roadmaps depended entirely on that node. Morris Chang, who had come out of retirement to retake the CEO role during the global financial crisis, flew to Silicon Valley, sat down with Jensen Huang, and offered a compensation package worth, by some accounts, hundreds of millions of dollars. Huang had 48 hours to accept. He did.
The way Apple does business is we put all our energy into products, and then we launch them, and if we were to bet heavily on TSMC, there would be no backup plan. You cannot double plan the kind of volumes that we do.
The message to the industry was crystalline: TSMC would absorb the cost of its own failures rather than leave a customer exposed. This was not generosity. It was the most coldly rational business decision imaginable — the trust that enabled Apple to commit 100% of its processor volume, the trust that enabled Nvidia to design chips it could not manufacture anywhere else, the trust that enabled the entire fabless ecosystem to exist. The compensation to Nvidia was an investment in the infrastructure of credibility. It paid for itself a thousandfold.
The Apple Catalyst and the Race to the Leading Edge
The Apple relationship, seeded at that 2010 dinner and consummated with the A8 chip in 2014, was the catalytic event that transformed TSMC from the world's best foundry into something closer to a natural monopoly at the leading edge.
Before Apple, TSMC had been a step behind Intel in process technology — a perfectly good foundry, but not the place you went if you needed the absolute smallest transistors on earth. Intel had held the manufacturing crown for decades, riding the tick-tock cadence of
Moore's Law with a discipline that bordered on religious devotion. But Intel had a structural problem: its manufacturing and design divisions were coupled. When a process node stumbled — as the 10nm node catastrophically did — the entire company's product roadmap collapsed. There was no firewall between the factory and the products.
Apple changed TSMC's trajectory because Apple's demands were simultaneously enormous in volume and unforgiving in specification. The iPhone was the highest-volume advanced-logic chip product on earth. Apple needed leading-edge transistor density for performance and power efficiency, and it needed it at scales measured in hundreds of millions of units per quarter. Meeting Apple's requirements forced TSMC to invest at a pace that would have been irrational for any other foundry customer base. The $9 billion Tainan fab — Fab 14 Phase 5 — was essentially Apple's private factory, built to Apple's timeline, ramped to Apple's volume.
The investment fundamentally altered the economics of leading-edge process development. TSMC could now amortize the staggering cost of developing each new node — $3 billion, $5 billion, eventually north of $10 billion — across Apple's volume plus the volume of every other fabless company. Intel, by contrast, had only its own product revenue to justify equivalent investments. The math was inexorable. By 2018, TSMC had caught Intel at 7nm. By 2020, it had surpassed Intel. By 2024, TSMC was manufacturing 3nm chips at scale while Intel was still struggling to stabilize its own competing processes.
The gap widened because the physics of semiconductor manufacturing create a vicious cycle for anyone who falls behind. Each new node requires not just new equipment — the $200 million-plus EUV lithography machines from ASML, of which there are a finite number on earth — but new process recipes, new yield engineering techniques, and thousands of incremental optimizations that can only be learned through high-volume production. A company that ships ten billion transistors at a given node learns more, faster, than a company that ships one billion. TSMC's volume advantage was, and is, simultaneously a learning advantage.
The Nightingale Army
Catching Samsung at the leading edge required something beyond capital investment and customer relationships. It required a cultural institution that has no precise analogue in Western corporate life.
In the early 2010s, Samsung was aggressively investing in its foundry business, leveraging its experience as a memory chip manufacturer and its position as Apple's secondary source for A-series processors. TSMC faced the genuine possibility of losing its technology lead. The company's response was the Nightingale Army — an initiative that borrowed, with characteristic Taiwanese pragmatism, from the operational logic of Foxconn's assembly lines.
TSMC assembled an unprecedented three-shift R&D department that ran twenty-four hours a day. The night-shift researchers — the Nightingales — earned a 30% base salary premium and a 50% dividend bonus above standard R&D compensation. The program was expensive and, by the norms of Silicon Valley work culture, borderline unthinkable. But it worked. The round-the-clock R&D cadence compressed development timelines, enabling TSMC to stay one to two years ahead of Samsung at each successive node.
This was not, it should be said, a happy story for the individuals involved. The semiconductor industry in Taiwan extracts an enormous human toll. The work is technically demanding, the hours are brutal, and the culture prizes dedication to the point of self-sacrifice. Morris Chang himself has noted that "Taiwanese people are more dedicated to work and willing to work overtime" — a description that flattens a complex social reality into a competitive advantage. The Nightingale Army was effective precisely because it exploited a cultural willingness to subordinate personal life to institutional mission that would be difficult to replicate in Phoenix, Arizona, or Dresden, Germany. This is not incidental to the TSMC story. It is central to the question of whether the company can successfully operate outside Taiwan — the most consequential strategic question it faces.
The Sacred Mountain
Taiwanese people call TSMC the hùguó shénshān — the "Sacred Guardian Mountain." The term did not originate in Taiwan; it came from foreign journalists groping for metaphors adequate to the scale of what TSMC means to the island's security. But Taiwanese people have embraced it because it captures a truth that economic statistics only approximate.
Integrated circuit companies in Hsinchu Science Park reported NT$11.3 trillion ($363 billion) in revenue in 2022, over 75% of the total generated by the park's 500-plus companies. Taiwan's total GDP that year was approximately $720 billion. TSMC alone accounted for roughly half of that IC revenue. A single company, in a science park that locals once called monga-bo — "graveyard," in Taiwanese slang — had become the load-bearing pillar of the national economy and, by extension, the island's primary deterrent against invasion.
The "Silicon Shield" thesis — that China cannot risk attacking Taiwan because it would destroy the fabs that produce the chips China's own economy depends on — is elegant, widely cited, and not entirely convincing. It assumes that strategic calculus will prevail over ideology, that Xi Jinping values access to 3nm logic chips more than reunification. History is not kind to such assumptions. But the thesis does not need to be perfectly correct to be strategically useful. It needs only to add a marginal increment of deterrence, to make the cost-benefit calculation of invasion slightly less favorable. TSMC's fabs serve this function whether or not anyone in Beijing or Washington explicitly acknowledges it.
The geopolitical dimension also explains the pressure TSMC faces to build fabs outside Taiwan — and the company's reluctant, complicated acquiescence. The U.S. CHIPS and Science Act, signed in August 2022, allocated $52.7 billion in subsidies and incentives for domestic semiconductor manufacturing. TSMC's Arizona project — announced in 2020, expanded repeatedly, now encompassing multiple fabs and a total investment reportedly approaching $65 billion — is the centerpiece of this effort. In April 2024, TSMC struck an $11.6 billion deal with the Biden administration for CHIPS Act subsidies to support the Arizona buildout.
The unit cost of making chips in the United States is 50% higher than in Taiwan. If you want to re-establish a complete semiconductor supply chain in the U.S., you will not find it a different cost advantage; you will find it very, very expensive.
Chang's bluntness on this point — delivered publicly, repeatedly, with the authority of someone who had spent five decades in the semiconductor industry — was uncomfortable for Washington policymakers. But the math was real. The cost differential was not merely about labor rates; it reflected the density of the supply chain ecosystem in Taiwan, the proximity of chemical suppliers and equipment vendors, the institutional knowledge embedded in a workforce that had been building chips for thirty years, and yes, the cultural willingness to work shifts and schedules that American labor law and social norms do not easily accommodate.
The Arizona Question
TSMC's first Arizona fab, producing chips on the N4 process (a variant of 4nm), achieved a remarkable milestone that surprised even skeptics: its yields matched or exceeded those at equivalent facilities in Taiwan. This was announced in late 2024 and represented a genuine inflection point in the debate about whether advanced semiconductor manufacturing could operate at world-class levels on American soil.
But a fab's yield is not the same thing as a fab's economics. The Arizona facility was more expensive to build, more expensive to staff, and more expensive to operate than its Taiwanese counterparts. The environmental commitments alone — 14.5 megawatts of solar panels, renewable energy credits covering 100% of power use, a water reclamation plant targeting "near zero liquid discharge" in the Arizona desert — reflected regulatory and social expectations that simply do not apply in Hsinchu. TSMC's Arizona operation was installing covered parking with solar panels; in Taiwan, it was building the next node.
The cultural friction was real and extensively reported. Taiwanese engineers sent to train American workers encountered expectations about work hours, hierarchy, and the pace of problem-solving that clashed with TSMC's operational DNA. The company launched apprenticeship programs, adapted its management practices, and invested heavily in cross-cultural training. Whether this adaptation succeeds at scale — whether TSMC can run an American fab with the relentless efficiency of a Taiwanese one — remains the company's most significant operational experiment.
TSMC is also building in Japan (a majority-owned subsidiary, JASM, in Kumamoto) and breaking ground in Dresden, Germany, for a specialty technology fab producing 28/22nm and 16/12nm chips. The geographic diversification is strategically necessary and economically painful. Every dollar spent building capacity outside Taiwan is a dollar that could have been spent extending the technology lead that is the source of all the company's pricing power.
The AI Supercycle
And then came the demand shock that made every prior debate about TSMC's growth trajectory look quaint.
The explosion of artificial intelligence workloads beginning in 2023 — driven by the deployment of large language models, the buildout of AI training infrastructure, and the race among hyperscalers to secure GPU capacity — created a demand environment that TSMC's leadership described in terms bordering on disbelief. High-performance computing surpassed smartphones as TSMC's largest revenue segment. Nvidia's AI accelerators, manufactured exclusively at TSMC on leading-edge nodes, became the most supply-constrained component in the technology industry. The wafer capacity TSMC had planned for 2025 and 2026 was effectively sold out before the fabs were built.
The financial results were staggering. TSMC's revenue for January through September 2024 totaled NT$2.03 trillion, up 31.9% year-over-year. September 2024 revenue alone was NT$251.9 billion, a 39.6% increase from September 2023. Full-year 2024 revenue was approximately $87 to $90 billion, with gross margins sustained above 56%. The company guided for roughly 30% revenue growth in 2025 and raised its medium-term guidance to a 25% revenue CAGR through the end of the decade — a rate of growth that, for a company of TSMC's scale, implies adding the equivalent of a major semiconductor company's entire revenue base every two years.
CEO C.C. Wei — a Yale-trained electrical engineer who had succeeded Mark Liu as chairman and CEO in June 2024, having served as CEO since 2018 — offered a characteristically measured assessment during the Q4 2025 earnings call:
You are asking if AI demand is real. I am also very nervous about it because we have to invest USD 52 billion to 56 billion. If I didn't do it carefully, that would be a disaster for TSMC. Internally, we also see productivity improvements of 1%–2% using AI. In my view, AI is real and growing into daily life. It is a megatrend.
The nervousness was itself revealing. Here was the CEO of the world's most critical semiconductor manufacturer, a man who controlled the bottleneck through which virtually all advanced AI hardware must pass, admitting publicly that betting wrong on the scale of this buildout could be catastrophic. The $52 to $56 billion in planned 2026 capex was not a comfortable number; it was a calculated wager that the AI demand curve was structural rather than cyclical, that the hundreds of billions of dollars hyperscalers were pouring into data center infrastructure reflected genuine economic value creation rather than speculative exuberance.
The evidence supported the bet, if you squinted correctly. TSMC's AI accelerator revenue was growing at a mid-to-high 30% CAGR through 2029. Advanced packaging capacity — particularly CoWoS (Chip-on-Wafer-on-Substrate), the technology required to assemble Nvidia's multi-chiplet GPU designs — had become a bottleneck commanding scarcity pricing with margins approaching, by some estimates, 80%. The customers buying this capacity — Apple, Nvidia, AMD, Qualcomm, Broadcom, Amazon, Google, Microsoft — were themselves earning extraordinary economic rents that created tolerance for TSMC price increases that would destroy a normal supplier relationship. When your customer sells a GPU for $30,000 to $70,000 and earns 75% gross margins, your leverage as the sole manufacturer of that GPU is functionally unlimited.
The Moat at the Atomic [Scale](/mental-models/scale)
TSMC's competitive position in early 2026 can be stated simply and starkly: the company controls more than 95% of the market for logic chips manufactured at 3nm and below. It is two to three years ahead of any meaningful competitor. Its nearest rivals — Samsung Foundry and Intel Foundry Services — have suffered yield collapses, management turnover, and strategic confusion that have, if anything, widened the gap during the period when they were supposed to be closing it.
Samsung's foundry business has struggled with yields on its 3nm Gate-All-Around (GAA) transistor architecture, losing key customers — most notably Qualcomm, which shifted its flagship Snapdragon processors to TSMC. Intel, under Pat Gelsinger's ambitious IDM 2.0 strategy, invested heavily in foundry capabilities but could not overcome the fundamental disadvantage of running a foundry that also designs competing products, alongside persistent execution challenges at advanced nodes. Gelsinger departed Intel in late 2024, leaving the foundry strategy in limbo.
The moat is not merely technological. It is the compounding result of three decades of capital investment, process learning, ecosystem lock-in, and customer trust. It is the $200 million EUV machines from ASML — of which TSMC buys a disproportionate share — combined with the process recipes that turn those machines into working chips at commercial yields. It is the design kits and IP libraries that thousands of engineers worldwide have spent careers optimizing for TSMC processes. It is the institutional knowledge embedded in a workforce of over 70,000, many of whom have spent their entire careers in TSMC's fabs.
And it is, perhaps most importantly, the simple fact that the cost of entry is now prohibitive. Building a single leading-edge fab costs $20 billion or more. Developing a new process node costs $10 billion-plus. The entire program — from node development through high-volume manufacturing — requires $50 to $100 billion in cumulative investment and five to seven years of execution before a single chip is sold to a customer. The number of organizations on earth with both the capital and the institutional capability to attempt this can be counted on one hand, and most of them are already TSMC's customers.
Chris Miller, the author of
Chip War: The Fight for the World's Most Critical Technology, put it precisely when describing Hsinchu Science Park: "This park has been at the center of Taiwan's efforts to first build the technological capacity to have a chief industry, to second train the workforce that the industry requires, and third incubate most of the key firms that have emerged in Taiwan. It's a really interesting and successful mix of education and training programs where it's easy for firms to get established, and find partners and skilled workers." The ecosystem is the moat. You cannot replicate a single company; you certainly cannot replicate an entire industrial ecosystem developed over four decades.
The Man Behind the Machine Behind the Machine
Morris Chang retired from TSMC's board in June 2018 at the age of eighty-six, after serving as chairman for the second time (he had briefly retired in 2005, returned in 2009 during the financial crisis, and remained through the Apple transformation). His successor as CEO, C.C. Wei, was in many ways his operational mirror — a deeply technical leader, Yale-educated like Chang, who had spent time at Texas Instruments, ST Microelectronics, and Chartered Semiconductor before joining TSMC in 1998. Where Chang was the strategist and institution-builder, Wei was the process engineer who understood, at the molecular level, how to push transistor geometries to their physical limits.
The transition was seamless in a way that belied its difficulty. Chang had built TSMC so thoroughly around his vision that the vision could survive his departure — the culture of customer trust, the relentless capital reinvestment, the willingness to cannibalize existing nodes to pursue newer ones, the deep institutional relationships with ASML and the EDA ecosystem. Wei did not need to reinvent TSMC. He needed to execute the roadmap at a pace that was accelerating beyond anything Chang had planned for, as AI demand transformed the company from a cyclical manufacturer into something closer to a perpetual growth engine.
Chang himself, in his eighties and nineties, continued to shape public discourse about the semiconductor industry with characteristic directness. His memoir — published in Taiwan, not yet fully available in English — became an instant bestseller. His public comments about the impossibility of replicating Taiwan's chip ecosystem in the United States were simultaneously inconvenient and irrefutable. He had spent fifty years in the American semiconductor industry before building TSMC in Taiwan. He knew both systems intimately. When he said the cost differential was 50%, he was not guessing.
I had the dream of a great semiconductor company in Taiwan. Taiwan had certain advantages — hardworking, talented people, and the government was willing to help. But Taiwan needed a new kind of company. Not a design company that also manufactured. Not a manufacturer that also designed. A company that only manufactured, for everyone.
The Physics of Money
The financial architecture of TSMC in the AI era reveals a company operating at a level of economic performance that most business models cannot sustain. Return on invested capital estimated near 40% against a weighted average cost of capital around 12 to 13% — a spread of roughly 27 percentage points that, sustained over time, represents one of the most powerful compounding engines in global capitalism. Return on equity of 39%.
Free cash flow of approximately $30 billion in FY2024, suppressed by the massive capex program and poised to expand as depreciation cycles from earlier buildouts complete.
The margin structure tells its own story. Gross margins above 56% for a capital-intensive manufacturer are extraordinary — comparable to software companies, not hardware companies. The explanation lies in TSMC's pricing power at the leading edge, where it is effectively a monopoly supplier, combined with the depreciation dynamics of semiconductor equipment. A fab's tools depreciate over roughly five years, but production from those tools can continue for a decade or more at escalating prices as the process matures and yields improve. The company is simultaneously collecting monopoly rents on its newest nodes and earning margin windfalls on fully depreciated older nodes.
The AI demand wave was amplifying this dynamic. Advanced packaging revenues — CoWoS in particular — were growing faster than wafer revenues as multi-chiplet architectures required increasingly sophisticated assembly techniques. The packaging business, once a commodity afterthought, was transforming into a scarcity-priced bottleneck with margins that would have seemed hallucinatory five years earlier. TSMC's Q4 2025 results delivered gross margins of approximately 59 to 61%, against Wall Street consensus expectations of 56 to 57%. The company guided Q1 2026 gross margins to roughly 64%.
This was not a blip. It was a structural shift — a function of TSMC's monopoly position at the leading edge, the extraordinary willingness of AI-rich customers to pay scarcity premiums, and a depreciation cliff that was releasing billions in annual expense as equipment purchased for the 5nm buildout completed its accounting life while production continued at escalating prices. The market, trained by decades of semiconductor cyclicality to expect margin reversion, was systematically underpricing the durability of TSMC's new economics.
The Stagehand Takes the Stage
Owen Lin, the Taiwanese journalist who has become TSMC's unofficial historian, described the transformation with a metaphor that captures the company's peculiar position: "Taiwan used to be like the black-clad stagehands who move props unseen, working behind the scenes while companies like Apple and Nvidia took center stage. Now, these stagehands have become crucial players."
The metaphor is precise. For most of its history, TSMC was invisible to anyone outside the semiconductor industry. Its customers got the glory — the product launches, the stock market valuations, the magazine covers. TSMC got the wafer starts. This invisibility was, in a sense, by design. The pure-play foundry model works best when the foundry is infrastructure — as essential and unremarked-upon as electricity or water. You don't think about your utility company until the power goes out.
The pandemic changed that. The global chip shortage of 2020-2021 — which shut down automobile production lines, delayed consumer electronics shipments, and forced world leaders to suddenly care about semiconductor supply chains — made TSMC visible in a way that Morris Chang had specifically avoided. Suddenly, this company in Hsinchu that most people had never heard of was the most discussed strategic asset on earth. The U.S. Congress was debating subsidies to bring its fabs to American soil. The Chinese government was spending hundreds of billions trying to replicate its capabilities. European leaders were calling for "semiconductor sovereignty." And through it all, TSMC continued doing what it had always done: manufacturing chips, investing in the next node, and serving 534 customers across 12,682 products.
In 2025, TSMC served customers across high-performance computing, smartphones, IoT, automotive, and digital consumer electronics. Its annual capacity exceeded 17 million 12-inch equivalent wafers across six 12-inch GIGAFAB® fabs and four 8-inch fabs in Taiwan, plus facilities in Nanjing, Arizona, and Kumamoto. It operated with the quiet efficiency of a machine that has been optimized for decades by people who believe, with genuine conviction, that manufacturing excellence is a form of moral commitment.
The production cost analogy that Owen Lin cited from his reporting illuminates TSMC's value proposition with startling clarity: "The production cost of Viagra accounts for only 0.7% of its price. If you reduce manufacturing costs by half, it only decreases the overall cost by 0.35%, which has a minimal impact. In contrast, semiconductor costs account for a much larger portion of the final product price. When TSMC uses the best technology to produce Nvidia's GPUs, the cost reduction is substantial." This is the core of TSMC's power — it is not making a commodity input. It is making the thing that determines the performance, power efficiency, and cost of every electronic device on earth. The value it creates for its customers is enormous, and it captures only a fraction of that value in its own pricing. This is what makes its moat sustainable: its customers get rich from TSMC's manufacturing excellence, which makes them willing to pay more, which funds the next round of investment, which makes the manufacturing better.
In the basement of a fabrication facility in Hsinchu, machines worth more than $200 million each fire extreme ultraviolet light through molten tin droplets at 50,000 pulses per second, etching circuit patterns smaller than the wavelength of visible light onto silicon wafers. Each pulse must land within 0.05 nanometers of its target — roughly the width of a single atom. The tolerances required exceed those of the Hubble Space Telescope. The process cannot be executed at commercially viable yields anywhere on Earth outside of TSMC. A company founded in 1987 with government money and Philips technology, in a science park that its own workers called a graveyard, now manufactures the computational substrate of artificial intelligence, the silicon nervous system of the global economy, and the most strategically consequential industrial product since petroleum. Its founder, a refugee from a country that no longer exists in the form he left it, solved a problem that no one else had identified — that the semiconductor industry needed a manufacturer with no products — and built the answer in a place that the rest of the world had barely noticed. The fabs run twenty-four hours a day, seven days a week. The Nightingales are still working.
