A Factory for Other People’s Ideas

In 1985, a few weeks after Morris Chang arrived in Taiwan to run the government’s Industrial Technology Research Institute, he was called in by K. T. Li, the minister who had steered much of the island’s industrial policy. Li told him that Taiwan wanted a semiconductor industry, that Chang had managed large semiconductor businesses, and that he should come back with a business plan for a new company: what it would do and how much money it needed. Chang later recalled that he was first given a week, and then told to present in four days.¹

He spent those days taking an inventory. Taiwan, he concluded, had very little strength in research and development, almost none in chip design, little in sales and marketing, and almost none in intellectual property. The only possible strength it had, and even that a potential one, was manufacturing. So he proposed a company that would do only that: make chips designed by other people, and never design or sell chips of its own. He called it, looking back, “the least evil choice.” It had one problem, which he described as possibly fatal. There was no obvious market. Almost no companies designed chips without owning a factory, and the big chipmakers would hand work to an outsider only when they ran short of capacity or wanted to stop making something unprofitable.¹

Forty years later that company, Taiwan Semiconductor Manufacturing Company, reported revenue of $46.7 billion for the three months to September 2026, up 41 percent in dollar terms from a year earlier, with a net margin of 51 percent. Chips made on its 7-nanometre and more advanced processes accounted for 79 percent of its wafer revenue.² The market research firm TrendForce estimated that TSMC took 72.5 percent of the revenue of the world’s ten largest foundries in the second quarter of 2026; the next largest, Samsung, took 5.9 percent.³ A study for the Semiconductor Industry Association estimated in 2021 that Taiwan held 92 percent of the world’s capacity to make logic chips below 10 nanometres.⁴ The processors in most smartphones, the graphics chips that train artificial intelligence and the custom chips of the large cloud companies are drawn by others and made, almost all of them, in TSMC’s fabs.

This essay follows how that happened, from the era when every chip company built its own factory, through the decisions that turned a contract manufacturer two and a half generations behind into the leader, to the present, when governments treat its factories as matters of national security. Its deeper subject is the old question of what happens when the people who design a thing stop being the people who make it: what is gained, what knowledge moves where, who must trust whom, and where the power ends up.

When every chipmaker owned its factory

For the first three decades of the industry, making chips and designing them were the same business. Texas Instruments, Fairchild, Motorola, Intel and the Japanese electronics groups each designed their chips, made them in their own fabrication plants, or fabs, tested them and sold them. The integration had good reasons. Early chips were made by processes that were closer to cooking than to engineering, and a design could only be judged by whether it could be made to work in a particular factory. Designers and process engineers sat together because they had to.

Chang’s own career shows what that meant. He joined Texas Instruments in 1958 after three years at Sylvania and was put in charge of a production line for germanium transistors whose yield, the share of devices that worked, was close to zero. He kept changing the recipe, the temperatures, pressures and diffusion times, guided by device physics, until one day the yield rose to 25 or 30 percent. “Suddenly we were in business,” he recalled. He said he could not sleep that night.¹ Yield, the fraction of what you make that works, would remain the central number of the industry, and the knowledge of how to raise it would remain the central asset.

By the late 1960s Chang ran TI’s integrated circuits business. Working with a then small consultancy, the Boston Consulting Group, he introduced what he called learning curve pricing: cutting prices every quarter in advance of falling costs, on the expectation that rising volume would teach the factory to make each chip more cheaply. Many people, he said, thought it foolish to cut prices when the market did not demand it. TI’s market share kept growing.¹ The logic, that cost falls with cumulative experience and that whoever makes the most learns fastest, would later sit at the heart of the foundry.

The cost of a factory, meanwhile, kept rising. Gordon Moore had predicted in 1965 that the number of components on a chip would keep doubling, in the paper that named the trend, and the industry organised itself around keeping the prediction true; the essay on compute follows what that did to computers. Each new generation of chips needed finer lithography, cleaner rooms and more expensive tools, and the minimum efficient size of a fab grew with them. For a designer with a good idea, that cost became a wall. Chang saw it from inside: at TI and later at General Instrument, he said, many chip designers wanted to leave and start their own companies, and what stopped them was that they could not raise the money for a fab.¹ The industry’s common sense was summed up in a remark attributed to Jerry Sanders, the founder of AMD: real men have fabs.⁵

Separating the drawing from the building

The idea that design could be pulled apart from manufacture came first from academia. In the 1970s Carver Mead at Caltech and Lynn Conway at Xerox PARC developed a method for designing very large-scale integrated circuits using simplified, scalable design rules, expressed in terms of a single unit of length, so that a designer could draw a working chip without mastering the details of a particular factory’s process. Their textbook, Introduction to VLSI Systems, appeared in 1980. Conway also organised courses whose students’ designs were gathered onto a single shared wafer, so that dozens of chips could be made for the price of one run. In 1981 the US Defense Advanced Research Projects Agency created MOSIS, a service at the University of Southern California that collected designs from universities and companies, placed them together on multi-project wafers and had them made by commercial fabs.⁶

This was more than a convenience. It defined an interface between two kinds of work. On one side of the design rules sat the designer, who needed to know only the rules; on the other sat the factory, which promised that any design obeying them would work. Chang read Mead’s writing while still at TI. Mead had made the point, Chang said, that design could be separated from the technology, but he had not advocated a company that did only manufacturing, and there did not seem to be a stable market for one.¹

The first companies to try designing chips without a factory appeared at about the same time. Xilinx, founded in San Jose in 1984 by Ross Freeman, Bernard Vonderschmitt and James Barnett, sold programmable logic chips that it designed and had made by others; it is often described as one of the first fabless semiconductor companies.⁷ But such companies depended on the spare capacity of integrated manufacturers, who were also their potential competitors, and who would push them aside when their own products needed the lines.

Taiwan’s apprenticeship

Taiwan came to chips through a deliberate apprenticeship. In 1976 ITRI, a government research institute founded three years earlier, licensed a CMOS manufacturing process from RCA, an American electronics company that was not among the industry’s leaders, and sent a group of young engineers to RCA’s facilities to learn it. In 1977 the institute opened a small demonstration factory making chips on three-inch wafers. Taiwanese accounts report that its yields soon exceeded those of RCA’s own plant.⁸ In 1980 ITRI spun off part of the team and the technology as a company, United Microelectronics Corporation, UMC, financed by the government and private investors.

Chang was blunt about how far that left Taiwan from the frontier. The RCA technology was probably a generation behind the best in the United States when Taiwan got it, he said, and without a commercial base the institute could not keep pace with companies like Intel and TI. By 1987 the technology ITRI could pass on was two and a half generations behind.¹ Taiwan had learned to run a fab. It had not learned to lead.

The least evil choice

Chang’s plan of 1985 rested on three threads of thought, as he later described them. The first was Mead’s argument that design could be separated from manufacture. The second was thirty years of watching how fierce the competition was among companies like Intel, TI and the Japanese groups, and how hard it would be for a new Taiwanese firm to carve out a niche among them. The third was his inventory of Taiwan’s strengths, which came down to a potential in manufacturing. A pure-play foundry, a company doing nothing but manufacturing for others, used the one strength and sidestepped the weaknesses: it needed no product designs, its sales were simpler, and in those years most patent disputes were about circuit designs rather than processes.¹

The market problem remained. Chang did not claim to have foreseen the rise of fabless companies. He said he only hoped for it, and that he had better reasons to hope than people inside Intel or TI, because he had seen how many designers were held back only by the cost of a factory. If a foundry removed that cost, he reasoned, those designers would start companies, and the companies would become a stable and growing market.¹ It was a bet that supply would create its own demand, made by someone who had watched the demand being suppressed.

Raising the money proved hard. The government’s Development Fund took the largest share, a little under half. Chang approached the large American and Japanese chipmakers; most declined. Philips, the Dutch electronics group, was in his words the only willing significant investor, and took a stake of about 27.5 percent in return for cash and licences; the rest came from Taiwanese private investors, many of them persuaded by the government to take part.⁹ Sources differ slightly on the exact percentages. TSMC was founded in February 1987 and began production in a six-inch fab leased from ITRI, with 2- and 3.5-micron processes inherited from the institute.¹⁰

The early years were lean. Chang said the company depended until about 1991 on what he called leftovers from large chipmakers, plus a few small customers, first in Taiwan and then in the United States. It lost money in 1987 and again in 1990. In his autobiography Chang recounts that Intel placed significant orders in TSMC’s second year and, in qualifying it as a supplier, pushed it to adopt the statistical process control and quality disciplines Intel required, which raised TSMC’s standards.¹¹ The few fabless companies that existed still went mostly to Japanese foundries. Asked whether that was for technology or capacity, Chang answered: “Technology and trust.” A foundry is the source of a customer’s products, he said, and customers could not simply trust a newcomer immediately. “We had to prove it.”¹

A promise not to compete

What TSMC offered that the integrated manufacturers could not was a promise: it would never compete with its customers. Its values statement still puts integrity first and says that customers’ success is its own.¹² The promise mattered because handing a design to a factory meant handing over the most valuable thing a chip company owned, and depending on the factory for every unit it would sell. A manufacturer with its own chips would always be tempted to learn from a customer’s design, or to favour its own products when capacity was short. A pure foundry had no such temptation, or rather had arranged its business so that the temptation could not pay.

From about 1991 the fabless industry began, in Chang’s word, to mushroom. Chip designers left large firms and started companies that needed only engineers, design software and a foundry. TSMC grew with them. It listed on the Taiwan Stock Exchange in September 1994 and in October 1997 became, by its own account, the first Taiwanese company to list on the New York Stock Exchange.¹³ Philips sold its remaining shares in 2008, as it left the semiconductor business.¹⁴

One of the companies that came looking for capacity was a small graphics start-up called Nvidia. Chang recalled that Jensen Huang, its chief executive, wrote to him around 1997 after failing to get a response from TSMC’s American office; Chang telephoned him himself. Huang predicted that his next chip would make Nvidia one of TSMC’s largest customers, which Chang thought a bold statement for a company of its size.¹⁵ Nvidia became one of TSMC’s largest customers for decades. The relationship ran both ways: Nvidia could exist because TSMC existed, and TSMC’s future growth would come from companies like Nvidia that did not yet exist when it was founded.

Learning to lead: copper, and a refusal

For its first decade TSMC followed the technology leaders. Around 2000 it began to lead, and a decision about copper marks the turn.

At the end of the 1990s the industry was replacing aluminium wiring on chips with copper, which conducts better, together with new insulating materials. IBM had pioneered copper wiring and offered to share its 0.13-micron process with partners. TSMC said in 2000 that IBM had approached it first and that it had declined, for reasons of cost and competition. Chang gave three reasons, among them that TSMC had already made great strides developing its own process and that it valued technological independence.¹⁶ IBM went on to form an alliance with Infineon and TSMC’s rival UMC to develop a common process from 0.13 down to 0.1 micron, and announced in November 2000 that the partners had begun making chips with it.¹⁷

The trade-off was speed and certainty against ownership. A licensed process would have arrived proven. A process of its own would teach TSMC’s engineers everything that went into it, and every improvement would belong to TSMC. The bet paid off. TSMC’s process went into volume production, and by early 2003 its chief financial officer was citing demand for 0.13-micron chips, from customers that reportedly included the graphics makers Nvidia and ATI, as a driver of growth. The research firm iSuppli ranked TSMC the largest foundry in 2002, with 40 percent of the market, against 16.7 percent for UMC, whose leading-edge 0.13-micron shipments were 6 percent of its revenue at the end of that year.¹⁸ The episode is told inside the company as the moment TSMC stopped being a follower. In an industry where cost falls with accumulated experience, owning the process meant owning the learning.

Water under the lens

The next turn concerned light. Chips are patterned by lithography, which projects the image of a circuit onto a wafer coated with a light-sensitive resist. The finer the features, the shorter the wavelength needed, or the larger the aperture of the lens. By 2000 the industry used deep ultraviolet light at 193 nanometres, and its roadmap called for a move to 157 nanometres. That step needed lenses made of calcium fluoride and new resists, and it was running into one difficulty after another.

Burn Lin, a lithography researcher who had spent years at IBM before joining TSMC, had long championed an alternative: put a thin layer of water between the final lens and the wafer. Because water bends light more than air does, immersion would let existing 193-nanometre tools resolve finer features, and Lin argued it could make 157-nanometre tools unnecessary.¹⁹ In October 2003 TSMC was reported to have cancelled orders for several 157-nanometre machines. That December ASML, the Dutch maker of lithography tools, announced that its first order for an immersion scanner, the Twinscan XT:1250i, had come from TSMC.²⁰ TSMC researchers reported in 2004 a roughly 50 percent improvement in depth of focus using immersion at 90 and 65 nanometres.¹⁹

The alternatives were to follow the roadmap the industry had agreed, or to wait. TSMC chose the modest extension of what existed over the new system that needed everything to go right, and it could do so because it had in-house experts able to judge the physics for themselves rather than deferring to the consensus. Immersion lithography at 193 nanometres became the industry’s standard for well over a decade. The 157-nanometre programme was abandoned.

The crash of 2009

In the global financial crisis demand for chips collapsed. TSMC’s revenue in the first quarter of 2009 was 54.8 percent lower than a year earlier, its net income 94.5 percent lower, and its gross margin, normally above 40 percent, fell to 18.9 percent.²¹ The company’s management at the time, led by Rick Tsai, whom Chang had made chief executive in 2005, cut costs, and in early 2009 several hundred employees were dismissed on the basis of performance reviews.

In June 2009 Chang, still chairman, took back the role of chief executive.²² He offered to rehire the dismissed workers, and later described the dismissals as layoffs presented as performance judgements, which he said was not how TSMC should treat people.²³ He also raised spending. Capital expenditure, $1.89 billion in 2008, reached $2.67 billion in 2009, and in October that year he told investors TSMC would be more aggressive in investing in advanced 12-inch capacity to meet rapidly growing demand for 40- and 65-nanometre chips. In 2010 it spent $5.9 billion.²⁴

The bet was that the downturn was temporary and that capacity at the new generations would be the binding constraint when demand returned. It was not obviously right at the time. Counter-cyclical spending had ruined other chipmakers: Qimonda, the German memory maker, filed for insolvency in January 2009, and Japan’s Elpida followed in 2012. TSMC could take the risk because it had a strong balance sheet and because foundry demand, spread over hundreds of customers, recovered faster than the market for any single kind of chip. When smartphones arrived in force over the next few years, TSMC had the capacity they needed.

That year the old orthodoxy also broke at its source. AMD, whose founder had said real men have fabs, spun off its manufacturing in 2009 into a new company, GlobalFoundries, owned largely by an investment company of the Abu Dhabi government, and became a fabless designer.⁵ The 2009 recovery was not smooth for TSMC either. Its 40-nanometre process had yield problems that hurt Nvidia, and Chang later settled the dispute after his return.¹⁵

The grand alliance

From 2008 TSMC organised its customers, design-software firms and suppliers of circuit building blocks into what it called the Open Innovation Platform, so that a design could be prepared for its processes long before they were ready. A fabless company choosing TSMC was choosing a whole ecosystem of tools, libraries and partners that had been validated against its factories. The more customers TSMC had, the more partners built for it, and the more partners built for it, the easier it became for the next customer to choose it.

The largest single customer came from the smartphone. Apple designed its own processors from 2010 but initially had them made by Samsung, which was also its rival in phones. It began moving production to TSMC in 2014, and in 2015 split the A9 processor between the two. For the A10 of 2016 TSMC offered something Samsung did not: a new way of packaging the chip, integrated fan-out or InFO, which made the processor package thinner and better connected. Analysts at the time reported that TSMC won most or all of the orders, and InFO was widely cited as a reason.²⁵ The episode showed the value of the promise not to compete in its sharpest form. Samsung could match TSMC in many technologies. It could not stop being Apple’s competitor.

Paying for the light

Even immersion had limits. To keep shrinking chips, the industry needed extreme ultraviolet light, with a wavelength of 13.5 nanometres, produced by firing lasers at droplets of molten tin and reflected through mirrors in a vacuum because no lens can transmit it. EUV had been in development since the 1990s and was years late. By 2012 only ASML was close to a working production tool, and its development cost was more than any one company wanted to bear.

That July ASML offered its largest customers a minority stake in return for funding its research. Intel joined first. In August 2012 TSMC agreed to pay €838 million for 5 percent of ASML and to commit €276 million over five years to its research programmes on EUV and larger wafers; a TSMC executive described the purpose as controlling the escalating cost of making wafers.²⁶ The decision spread the cost of a technology no single firm could carry and kept TSMC close to the first tools. In October 2019 TSMC announced that its N7+ process, in volume production since the second quarter, was the first EUV process delivering customer products to market in high volume.²⁷

The leader stumbles

For most of the industry’s history the most advanced manufacturing in the world belonged to Intel. In the late 2010s that changed, and the way it changed explains why TSMC came to stand alone.

Intel defined its 10-nanometre process to increase the density of transistors 2.7 times over its 14-nanometre process, a larger step than the 2.4 times of the previous generation and well above the roughly 1.5 to 2 times typical for the industry. Because EUV was not ready when the process was defined, Intel had to reach that density with deep ultraviolet tools by patterning critical layers several times, which multiplied the opportunities for defects. Mass production, first planned for the second half of 2016, slipped repeatedly. Intel’s chief executive later acknowledged that the target had been too aggressive, and another senior executive linked the problems to the choice of aggressive scaling without EUV.²⁸ In July 2020 Intel announced that its 7-nanometre process was running about twelve months behind its internal yield target. Its chief executive, Bob Swan, said Intel would be “pragmatic and objective” about using its own process, an external foundry or a combination.²⁹ The company that had built the world’s most advanced fabs was considering renting someone else’s.

It is easy to judge this in hindsight. When the 10-nanometre target was set, Intel had a long record of delivering ambitious nodes on schedule, and nothing in that record suggested its engineers could not do it again. But a company that makes chips only for itself has a different relationship with risk from one that makes chips for hundreds of customers. TSMC’s schedules were disciplined by customers with their own product launches, who needed a predictable process more than a heroic one. TSMC moved in smaller, regular steps and, after N7+, adopted EUV as soon as it was usable.

The other contenders dropped away. In August 2018 GlobalFoundries put its 7-nanometre programme on hold indefinitely, saying customers wanted more from existing processes. AMD, already moving its leading products to TSMC, said it would focus its 7-nanometre portfolio there.³⁰ That left three companies at the leading edge: Intel, Samsung and TSMC. Samsung was first to production with a new kind of transistor, the gate-all-around design in which the gate wraps the channel on every side, announcing 3-nanometre mass production in June 2022.³¹ TSMC kept its proven FinFET transistors for its 3-nanometre generation and moved to nanosheets, its version of gate-all-around, with its 2-nanometre process, which it said entered volume production on schedule in the fourth quarter of 2025.³² By then most of the world’s demand for the most advanced chips was flowing to one company.

The chokepoint

Concentration had been a commercial fact. It became a political one in 2020.

On 15 May 2020 the US Commerce Department extended its controls on Huawei, the Chinese telecommunications company, to foreign-made chips produced with American equipment or software, which covered virtually every advanced fab in the world. Within hours of each other TSMC announced plans for a $12 billion fab in Arizona to make 5-nanometre chips from 2024.³³ TSMC stopped shipping to Huawei on 15 September 2020, when the rule took effect.³⁴ Huawei had been one of its largest customers; one estimate put its share of TSMC’s 2019 revenue at 14.5 percent. In September 2023 TechInsights found that a new Huawei phone used a 7-nanometre processor made by China’s SMIC without EUV tools, showing that the controls slowed Chinese progress without stopping it.³⁵ In October 2024 TSMC halted shipments to a customer after one of its chips turned up inside a Huawei product; Reuters later reported the customer was Sophgo.³⁴

The idea that Taiwan’s chip industry protected it had a name long before. In 2001 the journalist Craig Addison argued in Silicon Shield that the world’s dependence on Taiwan’s chips made an attack by China self-defeating, because it would destroy what China and the United States both needed.³⁶ By the 2020s the argument was being turned around: if the shield made Taiwan valuable, it also made the world vulnerable to anything that happened there, from war to earthquakes. When the strongest earthquake in twenty-five years struck near Hualien in April 2024, TSMC evacuated its fabs; it said more than 70 percent of its tools had recovered within ten hours, and that none of its EUV tools was damaged.³⁷

Building abroad

Arizona became the test of whether the foundry’s knowledge could be moved. At a ceremony in December 2022 marking the arrival of tools, the company raised its plans to $40 billion and two fabs. Chang, by then retired, told the audience that globalisation was almost dead and free trade almost dead, and that he did not think they would return for a while.³⁸ The United States had passed the CHIPS and Science Act that August; in November 2024 the Commerce Department finalised a grant of up to $6.6 billion for TSMC Arizona, plus up to $5 billion in loans, after TSMC had raised its plans to $65 billion and three fabs.³⁹

The work was harder than in Taiwan. In 2023 TSMC delayed the first fab’s production to 2025, blaming a shortage of workers skilled in installing equipment in a semiconductor-grade facility, and sought visas for several hundred Taiwanese technicians. Arizona construction unions disputed the explanation, saying local workers had built fabs for Intel and that the problems lay in management and safety practices.⁴⁰ In 2024 current and former employees filed a proposed class action alleging discrimination against workers who were not of East Asian or Taiwanese origin; it was refiled in 2025 with more plaintiffs. TSMC has said it hires and promotes without regard to race or nationality. The allegations had not been tried at the time of writing.⁴¹ The first fab nevertheless began volume production on TSMC’s N4 process in the fourth quarter of 2024, with yields the company described as comparable to its Taiwan fabs.

The commitments kept growing, under steady political pressure. In March 2025 TSMC added $100 billion, for a total of $165 billion; in July 2026, announcing second-quarter results, it added another $100 billion for further advanced fabs, bringing the total to $265 billion.⁴² Taiwan and the United States had agreed in January 2026 on a trade deal under which Taiwanese companies pledged at least $250 billion of investment in American chipmaking, and those building fabs in the US could import chips free of the new semiconductor tariffs up to a multiple of their American capacity.⁴³ TSMC also opened a fab in Kumamoto, Japan, in 2024 and began building one in Dresden. Intel, meanwhile, had become a ward of industrial policy: in August 2025 the US government converted grants into a 9.9 percent stake in the company.⁴⁴

The AI factory

The newest demand came from artificial intelligence. The graphics processors that train and run large AI models are among the largest chips made, and the most advanced of them combine a processor with stacks of high-bandwidth memory on a silicon base, using a packaging technique TSMC calls CoWoS. In 2023 packaging, more than wafer capacity, became the bottleneck. TSMC’s chairman said that September that it could meet only about 80 percent of customers’ demand and expected the shortage to last about a year and a half.⁴⁵ TSMC said it would double CoWoS capacity in 2024 and again in 2025, and still described demand as well beyond what it could supply. Packaging, once a low-margin step handed to subcontractors, had become part of the leading edge, as it had in the Apple deal a decade earlier.

The money followed. TSMC’s revenue in 2025 was about NT$3.81 trillion, roughly $122 billion, and it spent $40.9 billion on capital equipment; in January 2026 it guided spending for 2026 of $52 to 56 billion, then the largest in its history, and raised the range again in July.⁴⁶ In its third quarter of 2026, 2-nanometre chips already made up 8 percent of wafer revenue.² The company also had to guard what it knew. In April 2026 a Taiwanese court sentenced a former employee of TSMC and of the equipment maker Tokyo Electron to ten years in prison, and three former TSMC engineers to between two and six years, for leaking trade secrets about its 2-nanometre process; Tokyo Electron’s Taiwanese unit was fined. It was the first case brought under Taiwan’s law protecting core national technologies.⁴⁷

What the foundry is today

Today a chip is usually the work of at least two companies that never share a building. A designer, which might be Apple, Nvidia, Qualcomm, AMD, Broadcom or one of the cloud companies, defines the chip in software, using tools from a few design-software firms and building blocks licensed from others, against design rules supplied by the foundry. The foundry turns the design into masks and processes wafers through months of steps on tools from a handful of suppliers, above all ASML. The finished dies may then be stacked and packaged with memory, often by the foundry itself. TSMC is not the only foundry. Samsung, Intel, GlobalFoundries, UMC and China’s SMIC all make chips for others, and most of the world’s chips by number are made on older processes in many places. But at the leading edge, where the most valuable chips are made, TSMC has no close rival, and most of its leading-edge capacity is in Taiwan.

How the foundry changed the way chips are made and decided

From owning the factory to renting the future

The foundry turned the most expensive part of a chip company into a service. Designers could start companies with engineers and software, and the industry’s capital spending was concentrated in a few firms that served everyone. Decisions about which technologies to build moved from hundreds of companies to a few.

From the company to the ecosystem

Choosing a foundry came to mean choosing an entire system of design tools, libraries, packaging and partners. TSMC’s advantage lay as much in that system, and in the trust of the firms within it, as in any single process.

From commodity to chokepoint

Contract manufacturing once sounded like the least strategic part of an industry. At the leading edge it became the narrowest point, because the learning needed to make the most advanced chips accumulated in very few places, and the cost of each new generation drove out all but a few.

From shrinking to stacking

For decades progress meant making transistors smaller. As shrinking slowed and grew dearer, more of the gains came from new transistor shapes and from packaging several chips together. The boundary between making a chip and assembling one blurred, and the foundry absorbed both.

From markets to ministries

Where chips are made, and for whom, became a matter for governments. Export controls, subsidies, tariffs and investment pledges now shape where fabs are built as much as costs do.

What the history suggests

The foundry repeats several patterns seen across this site. A small working thing beat a grand design: TSMC began far behind, on leftovers, and grew by serving customers no one else would serve well. Abstraction won, as design rules let designers ignore the factory. A defining strength became a trap for Intel, whose excellence as a self-contained manufacturer left it exposed when its own process faltered. Who pays shaped what got built, as customers funded ASML’s research and governments funded fabs abroad. And making something cheap moved the cost elsewhere: the designers’ freedom from capital was paid for by concentrating the capital, and the risk, in one place.

The foundry also shows something more particular. Its product is a promise, that any design obeying its rules will be made faithfully, in volume, without the maker taking advantage of what it sees. Behind that promise lies knowledge that is learned through repetition, hard to write down and hard to move, and the decisions that made TSMC were, again and again, decisions to own that learning. Separating design from manufacture freed the designers. It did not make the making less important. It made it less visible, and more concentrated. That is where the philosophy begins.

Understanding the philosophy

The rest of this essay examines eight ideas beneath the foundry, one at a time. Each is explained in plain terms, placed in the tradition of thought that has examined it, traced in the history above, tested against its strongest objection and related to how people and organisations decide today. Two of them come from psychology. The connections to economists, philosophers, sociologists and psychologists are interpretations offered to make the ideas clearer. The engineers and executives in this history were solving practical problems, and did not, as a rule, set out to answer these thinkers.

1. Dividing the work: designing apart from making

The idea in plain terms

A chip used to be designed and made by the same company, often by people in the same building. The foundry split the work in two. One firm decides what the chip should do and draws it; another firm makes it. Splitting work in this way lets each side specialise, and specialists usually get better faster than generalists. The question is what the split costs: what each side stops knowing about the other, and whether the side that draws can keep drawing well once it no longer makes.

The tradition

Adam Smith opened The Wealth of Nations (1776) with a pin factory. One worker alone, he wrote, could make perhaps a single pin a day, and certainly not twenty; ten workers dividing the job into about eighteen distinct operations could make upwards of forty-eight thousand. Smith gave three reasons: each worker became more dexterous at one task, no time was lost moving between tasks, and people focused on a single operation were more likely to invent machines to do it. He added that the division of labour is limited by the extent of the market. A specialised pin-head maker needs enough customers for pin heads.

In 1832 Charles Babbage, in On the Economy of Machinery and Manufactures, added a point Smith had missed. Dividing a process lets a firm pay for each skill only in the quantity it needs, instead of paying a highly skilled worker to do the unskilled parts too. What is now called the Babbage principle explains why the division of labour reorganises costs as well as raising speed.

The modern version of this tradition concerns how to divide design itself. In Design Rules (2000) the economists Carliss Baldwin and Kim Clark argued that the computer industry grew by modularity: breaking a system into parts that interact only through visible, agreed interfaces, while everything else stays hidden inside each part. Once the interface is fixed, firms can work on the modules independently, and the value of the whole system can grow through many experiments at once. Mead and Conway’s design rules were such an interface. They defined, in a few pages, everything a designer needed to know about the factory, and hid the rest.

Where it shows up

Chang’s plan was a bet on Smith’s condition about the extent of the market. A firm that only manufactured needed many customers, and there were almost none in 1985. He bet that the division would create the market by removing the cost that kept designers from starting companies. It did. From 1991 the fabless industry multiplied and TSMC’s specialisation deepened: yields, process generations, packaging, each improved by people who did nothing else. The Open Innovation Platform extended the modular logic to design tools and building blocks. Intel, which kept design and manufacture together, kept a tight link between its designers and its process engineers, which helped it for decades and then bound its products to a process that was late.

The strongest objection

In 2009 the Harvard Business School professors Gary Pisano and Willy Shih argued in the Harvard Business Review that separating design from manufacture can destroy what they called the industrial commons: the shared know-how, suppliers and skilled workers that grow around making things, and on which future innovation depends. When an industry’s manufacturing leaves a region, they argued, its ability to invent the next generation of products often follows, because many innovations come from problems discovered on the factory floor. The division may be efficient for each firm and damaging for the economy that loses the making. On this view, the American companies that went fabless gained for themselves and lost, collectively, the capacity that the United States has since spent hundreds of billions of dollars trying to rebuild. The objection is strong, and the subsidies of the 2020s are, in a sense, an admission of it.

What it means today

Before an organisation gives a part of its work to a specialist, it should ask what knowledge lives in that part and whether it will need that knowledge to design the next thing. Interfaces like design rules are powerful because they let people ignore what is on the other side. That is also their danger. A leader deciding what to outsource is deciding what the organisation will no longer be able to learn.

2. Learning by doing, and the curve that rewards whoever makes the most

The idea in plain terms

The more of something a factory makes, the better it gets at making it. Each doubling of cumulative output tends to cut the cost of each unit by a roughly constant fraction, because workers and engineers discover thousands of small improvements that no one could have planned. If that is true, the firm that makes the most learns the fastest, and its lead grows. Being big is a cause of getting better, and getting better is a cause of being big.

The tradition

In 1936 Theodore Wright, an aeronautical engineer, published a paper in the Journal of the Aeronautical Sciences showing that the labour needed to build an aircraft fell by a predictable percentage each time cumulative production doubled. In 1962 the economist Kenneth Arrow gave the observation a theory in “The Economic Implications of Learning by Doing,” arguing that knowledge is a by-product of production itself, accumulated through experience, and that this helps explain why economies grow. Arrow noted a consequence: because the learning spills over and cannot be fully captured, markets may produce too little of it.

In the 1960s Bruce Henderson’s Boston Consulting Group turned the curve into strategy. If costs fall with cumulative volume, a firm should price low early, win share, accumulate experience faster than rivals and enjoy lower costs ever after. Chang’s learning curve pricing at Texas Instruments, developed with BCG in the late 1960s, was one of the first systematic uses of the idea in semiconductors.

Where it shows up

The foundry is a learning curve that serves everyone. By pooling demand from hundreds of customers, TSMC accumulated volume on each process faster than any single chipmaker could, and with it the yield improvements that lowered costs. The refusal of IBM’s process in 2000 was a decision to own its learning rather than rent someone else’s. The capital spending of 2009 was a decision to keep the curve moving when the market said stop. Even Arizona fits: TSMC’s argument that its American fabs cost more rests partly on their having less accumulated experience, and their yields depend on how fast learning can be transferred from Taiwan.

The strongest objection

In 1974 William Abernathy and Kenneth Wayne published “Limits of the Learning Curve” in the Harvard Business Review. Their example was Ford’s Model T, whose cost fell dramatically as Ford pursued the curve with ever more specialised plants, until the product could not change when customers wanted closed bodies and variety, and General Motors overtook Ford. The pursuit of learning on one product, they argued, builds rigidity, and the curve can lead a firm efficiently into a dead end. A second objection is that learning curves describe the past and do not guarantee the future: physics, not experience, may set the limit, as the end of easy shrinking suggests.

The foundry partly answers Abernathy and Wayne, because TSMC rides many curves at once, one per process, and abandons each as customers move on. But the objection describes Intel’s trouble well. Its excellence on one integrated path made the path hard to leave.

What it means today

Experience compounds, and so does rigidity. Leaders should ask where their organisation’s learning is accumulating, whether they own it, and whether the curve they are riding is still pointing at what customers will want. Volume is a way of learning, which is why giving up volume, by outsourcing it or cutting it in a downturn, often costs more than the spreadsheet shows.

3. Knowledge that will not travel

The idea in plain terms

Some knowledge can be written in a manual. Much cannot. How to bring a new process to high yield, how to read a drift in a tool before it ruins a batch, how a team coordinates through the night when something fails: this knowledge lives in people and in their habits of working together, and it moves only when people move, slowly. That is why a fab cannot simply be copied, and why the most advanced chipmaking has stayed in so few places.

The tradition

The chemist and philosopher Michael Polanyi argued in Personal Knowledge (1958) and The Tacit Dimension (1966) that “we can know more than we can tell.” A skilled person relies on a vast body of awareness that they attend from rather than to, as a cyclist keeps balance without being able to state the rules. Such knowledge is passed on by apprenticeship, by working alongside someone who has it, more than by instruction. The site’s idea We Know More Than We Can Tell traces this further.

The sociologist Harry Collins tested the idea in laboratories. In the 1970s he studied scientists trying to build a new kind of laser and found that no group succeeded from published papers alone; those who succeeded had spent time with a group that had already built one. In a 2001 study of physicists measuring the quality of sapphire crystals, he found the same: a technique that worked in a Moscow laboratory could not be reproduced in Glasgow until people had visited, watched and built trust in each other’s results. Collins concluded that the transfer of skill is a social process, dependent on contact and on trust.

Where it shows up

Taiwan’s industry began with exactly this kind of transfer. ITRI did not simply buy RCA’s documents; it sent engineers to work in RCA’s plant, then built a demonstration factory and learned by running it. TSMC began with ITRI’s people as well as its processes. When Intel qualified TSMC as a supplier, it taught disciplines that Chang described as raising the company’s standards. Arizona’s difficulties show the transfer in reverse: TSMC sought visas for Taiwanese technicians because it believed that installing and ramping a fab needed people who had done it before. The 2026 trade-secrets case shows the other side. What was allegedly leaked was information that could help another firm learn faster, which suggests that some of what TSMC knows can be written down after all.

The strongest objection

The objection is that tacit knowledge is always on the way to becoming explicit, and that the mystique around it can protect incumbents and excuse failures. Modern fabs are among the most instrumented places on earth: every tool logs its data, and yield is managed with statistics and machine learning. SMIC produced 7-nanometre chips without EUV and under sanctions, showing that determined learners can close gaps faster than their rivals predict. And the claim that only Taiwanese workers can run a fab is precisely what the Arizona unions disputed. The objection is serious. The best reply is that codification and tacit skill grow together: each new layer of data creates new judgements about what the data mean, which are learned in practice.

What it means today

What an organisation knows is held as much by its people working together as in its documents. Leaders moving work to a new site, a new team or a new system should budget for apprenticeship, not only for documentation, and should be wary of anyone who says the knowledge is all in the manual, and equally of anyone who says none of it can ever be written down.

4. Trust, and the promise not to compete

The idea in plain terms

Handing your design to another company makes you vulnerable. The company could copy it, favour its own products, or raise its prices once you depend on it. TSMC’s founding promise, that it would never sell chips of its own, removed some of those fears by removing the motive. Trust in business is often treated as a feeling. Here it was built into the structure of the company.

The tradition

The economist Oliver Williamson, building on Ronald Coase’s question of why firms exist at all (the subject of the site’s idea Why Firms Exist), argued in The Economic Institutions of Capitalism (1985) that the main reason to bring an activity inside a firm is asset specificity. When one party must invest in something valuable only within a particular relationship, it becomes exposed to the other party’s opportunism, “self-interest seeking with guile,” in Williamson’s phrase. In 1978 Benjamin Klein, Robert Crawford and Armen Alchian described the danger as hold-up: once one side has sunk its investment, the other can demand a larger share. Their famous example was General Motors buying Fisher Body, after disputes over a supplier that had built plants for GM alone. The usual remedy for hold-up is integration, the very structure the foundry took apart.

Thomas Schelling, in The Strategy of Conflict (1960), showed another remedy. A party can make itself trustworthy by binding itself: by giving up options in a way others can see, so that it cannot later act against them even if it wanted to. A commitment is credible when breaking it would be costly or impossible for the one who made it.

Where it shows up

The fabless model looked, by Williamson’s logic, dangerous. A chip designer invests years in a design tuned to one foundry’s process and cannot easily move it, a textbook case of asset specificity. The industry’s answer was not integration but commitment. By staying out of chip products altogether, TSMC gave up the profits a Samsung could earn and, with them, the main motive to exploit what it saw. Chang’s two words, “technology and trust,” describe what customers were buying. The Apple case shows the value of the commitment in a single decision: Samsung could compete on technology, but it could not stop being Apple’s rival in phones.

The strongest objection

The objection is that a promise not to compete does nothing about the other kind of hold-up: price and allocation. A supplier with more than 70 percent of a market, and no close rival at the leading edge, can raise prices and decide which customer gets scarce capacity first, without designing a single chip. In 2023, when packaging capacity met only about 80 percent of demand, someone was choosing whose 20 percent went unmet. On this view trust in TSMC is partly a polite name for the absence of alternatives. And the commitment is enforced by reputation, not law: it holds only as long as TSMC judges that keeping it pays. The reply is that TSMC’s long record of restraint on pricing and allocation is itself a form of commitment, repeated every year. But the dependence is real, which is why customers and governments keep trying to create alternatives.

What it means today

The most valuable assurance a supplier can give is often what it will not do. For buyers, the question is not only whether a partner is trustworthy now but what keeps it trustworthy once you depend on it: its structure, its reputation, your alternatives. For firms that sell to others, giving up a business that would compete with your customers can make you the partner everyone can choose.

5. Fear in the downturn

The idea in plain terms

When revenue falls by half, every instinct says cut. Spending looks like loss, and losses feel much worse than equal gains feel good. Yet in a cyclical industry the decisions that determine who leads the next upturn are often made at the bottom of the downturn, when the future looks worst. Acting against the instinct is partly a matter of judgement about the cycle and partly a matter of psychology: of whether leaders can see a temporary loss as temporary.

The tradition

In 1979 Daniel Kahneman and Amos Tversky published prospect theory in Econometrica. People judge outcomes as gains or losses relative to a reference point, not as final states of wealth, and losses loom larger than gains, by a factor often estimated at around two. People are risk-averse when choosing among gains and risk-seeking when trying to avoid a sure loss. Later research has qualified the size and generality of loss aversion, and some studies find it weaker than first reported, but the asymmetry is among the most robust findings in behavioural economics. The site’s idea Loss Aversion and the Pull of the Status Quo covers it in detail.

In 1993 Kahneman and Dan Lovallo applied these findings to managers in “Timid Choices and Bold Forecasts,” in Management Science. Managers, they argued, are at once too optimistic in their forecasts and too timid in their choices: they evaluate each risky option alone, as if it were the only one, and so refuse bets that, taken as a portfolio over time, would pay. Barry Staw’s research from 1976 on escalation of commitment shows the opposite failure: people who feel responsible for a losing course of action tend to invest more in it, to justify what they have already spent, a danger described in the site’s idea Sunk Costs and the Escalation of Commitment.

Where it shows up

In 2009 TSMC’s revenue fell by 54.8 percent in a quarter. The first response, cutting staff, followed the frame of loss. Chang’s return reframed the situation: the reference point became the capacity the company would need when demand returned, against which not spending was the loss. He took back dismissed employees and raised capital spending to $2.67 billion and then $5.9 billion. In Kahneman and Lovallo’s terms, he treated the downturn as one draw in a long series of cycles rather than as a single, frightening bet.

Intel’s 10-nanometre programme may show the other side, though it calls for caution. A company with a long record of success may find it hard to scale back a goal it has announced, and each delay may seem the last. Whether that happened inside Intel is not documented in the sources used here.

The strongest objection

The objection is outcome bias. We praise bold counter-cyclical bets that worked and forget the ones that failed. Qimonda and Elpida also invested in capacity through downturns and went bankrupt; their leaders might have described their choices just as Chang did. The difference may have been luck, the structure of the memory market, or TSMC’s balance sheet, not superior psychology. The site’s idea Moral Luck and Outcome Bias warns against judging a decision by how it turned out. This is right, and it is why the decision is graded here on process: TSMC had the cash, the diversified customer base and the evidence of past cycles that made the bet reasonable at the time, not merely successful afterwards.

What it means today

In a downturn, leaders should ask what reference point they are using. Measured against last year’s revenue, every expense is a loss; measured against the position the organisation needs when conditions recover, some cuts are the real losses. Judging a decision as one of many similar decisions over time, rather than as a single gamble, is a practical defence against timidity. It is no defence against a weak balance sheet.

6. Attention at the scale of a nanometre

The idea in plain terms

A modern chip passes through more than a thousand process steps over several months, and a single particle or a drift of a few atoms in one step can ruin it. Making chips at high yield is therefore a discipline of attention: noticing small deviations before they become large failures, and keeping that vigilance up, shift after shift, without becoming numb to warnings. That attention is a psychological achievement as much as a technical one, and it has human costs.

The tradition

During the Second World War the British psychologist Norman Mackworth studied radar operators, who were missing signals late in their watches. In experiments published in 1948 he showed that people watching for rare signals detect fewer of them after only half an hour: the vigilance decrement. Attention to rare events decays, and the decay grows with monotony and fatigue.

Organisations can compensate for what individuals cannot sustain. The organisational theorists Karl Weick and Kathleen Sutcliffe, studying aircraft carriers, nuclear plants and air traffic control, described in Managing the Unexpected (2001) what they called high-reliability organisations. Such organisations are preoccupied with failure, treat small anomalies as possible signs of larger problems, resist simplifying explanations, stay close to operations, and defer to whoever has the most expertise on a problem, whatever their rank. Weick and Karlene Roberts had earlier described this as a “collective mind,” in which people act heedfully, aware of how their actions connect to others.

The sociologist Diane Vaughan described the opposite process in The Challenger Launch Decision (1996): the normalisation of deviance, in which small departures from expected performance, repeatedly survived, come to be accepted as normal until they are not. The site’s idea When Small Risks Become Normal follows it further.

Where it shows up

Chang’s first success at TI, raising a yield from almost nothing to 25 or 30 percent by changing a recipe step by step, was an act of sustained attention guided by physics. Intel’s insistence that TSMC adopt statistical process control in its early years was a lesson in turning individual vigilance into a system. The recovery after the 2024 earthquake, with more than 70 percent of tools back within ten hours, depended on practised routines. The demands of such work are also visible in reports of long hours and night calls at TSMC, and in the friction in Arizona, where expectations about hours and hierarchy formed in Taiwan met a different workforce. High reliability depends on deference to expertise; a workplace where people feel unable to speak, which some Arizona employees alleged, works against it.

The strongest objection

The objection is that this romanticises human attention in what is now an automated process. Tools monitor themselves; statistical process control, invented by Walter Shewhart at Bell Labs in the 1920s, flags drifts without anyone watching; machine-learning systems predict defects. Vigilance has been delegated to machines, which do not suffer Mackworth’s decrement. That is largely true and is part of why yields are so high. But automation moves attention rather than removing it. People must still decide which alarms matter, and automation can breed its own complacency, as the research described in Trusting the Machine Too Much shows. And there is a second objection: that the culture of vigilance can become one of overwork, with costs that do not appear in yield data.

What it means today

Reliability is built from small signals taken seriously, and from systems that keep attention alive when individuals tire. Leaders should ask which deviations their organisation has learned to live with, and whether the people closest to the work can halt it. They should also ask what their reliability costs the people who sustain it, because exhausted attention is the kind that fails.

7. Interdependence as a weapon, and as a shield

The idea in plain terms

When countries trade, they come to depend on each other. That dependence can keep the peace, because war would destroy what both need. It can also be used as a lever: whoever controls a point through which everyone must pass can threaten to close it. Advanced chips pass through one company on one island, using machines from one company in the Netherlands and software from a few American firms. Those chokepoints have made the foundry an instrument of power.

The tradition

The idea that commerce softens conflict goes back at least to Montesquieu, who wrote in The Spirit of the Laws (1748) that the natural effect of commerce is to lead to peace, since two nations that trade with each other become mutually dependent. Immanuel Kant argued in Perpetual Peace (1795) that the spirit of commerce, which cannot coexist with war, would sooner or later take hold of every nation.

In 1945 the economist Albert Hirschman, a refugee from Nazi Germany, showed the darker side in National Power and the Structure of Foreign Trade. Germany in the 1930s had deliberately built trade with small southeastern European countries so that they depended on the German market far more than Germany depended on them, which gave Germany political influence over them. Dependence, Hirschman argued, is power when it is asymmetric. Robert Keohane and Joseph Nye developed this in Power and Interdependence (1977), distinguishing a country’s sensitivity to another’s actions from its vulnerability, the cost of finding alternatives. In 2019 Henry Farrell and Abraham Newman argued in International Security that global networks, from finance to the internet, concentrate around hubs, and that states controlling the hubs can use them to watch and to choke others. They called it weaponised interdependence.

Where it shows up

The rule of May 2020 was a textbook case of weaponised interdependence. The United States did not make Huawei’s chips; it controlled the equipment and software that every advanced foundry needed, and so could order TSMC, a Taiwanese company, to stop serving a Chinese one. SMIC’s 7-nanometre chip of 2023 shows Keohane and Nye’s distinction: China was sensitive to the controls, but not wholly vulnerable, because it could find costly alternatives. Addison’s silicon shield is Montesquieu’s argument applied to Taiwan: the world needs the island’s fabs too much to let them be destroyed. The American push to build fabs in Arizona is a response to Hirschman: it aims to reduce the United States’ own vulnerability, which, if it succeeds, may also weaken the shield.

The strongest objection

The objection to the shield is that interdependence has never reliably prevented war. Europe’s great powers traded heavily with each other in 1914. The economist Norman Angell argued in The Great Illusion (1909) that war between industrial nations had become economically irrational; five years later they fought one anyway. States go to war for reasons of security, identity and pride that trade does not cancel. The objection to weaponised interdependence runs the other way: using chokepoints teaches others to build around them, so each use reduces the weapon’s future power. China’s investment in its own chip industry after 2019 suggests that both objections hold. The shield may be weaker than its admirers hope, and the weapon may wear out with use.

What it means today

Any organisation that sits at a chokepoint holds power it may not want, and will be pressed by governments to use it. Any organisation that depends on one should know which of its dependencies are merely sensitive and which are truly vulnerable. Concentration that is efficient in peacetime is, for governments, a strategic exposure, and they will pay heavily to reduce it.

8. The hidden maker

The idea in plain terms

The chip in a phone is advertised under the name of the company that designed it. The company that made it rarely appears on the box. The foundry model makes the maker invisible to the people who use what it makes, and this shapes how we value work: we tend to credit the idea and overlook the making, as if designing were thought and manufacturing mere execution. The foundry’s history suggests that the making is where much of the hardest thinking happens.

The tradition

In the first volume of Capital (1867), Karl Marx described what he called the fetishism of commodities. When goods are bought and sold, the social relations between the people who make them appear as relations between things, as prices. The labour and the human arrangements inside a product disappear from view, and the product seems to have its value by nature. One does not have to accept Marx’s economics to see the observation: a market tends to hide the making behind the thing.

Hannah Arendt, in The Human Condition (1958), distinguished labour, the endless, repetitive activity of sustaining life, from work, the making of durable things, and both from action. She worried that modern societies esteem the visible product and undervalue the processes that make and sustain it. Matthew Crawford, a philosopher who left a think-tank job to repair motorcycles, argued in Shop Class as Soulcraft (2009) that the separation of thinking from doing, which management theory since Frederick Taylor had pursued, rested on a false idea that manual and practical work contains no thought. In his experience, diagnosing a fault was as demanding an act of reasoning as most office work.

In 1992 Stan Shih, the founder of the Taiwanese computer maker Acer, drew what he called the smile curve: in electronics, he argued, the value lies at the two ends, in research and design at one end and brands and marketing at the other, while manufacturing in the middle earns the least. The curve became a common explanation of why designing pays and making does not.

Where it shows up

The foundry history overturns the smile curve at its own centre. TSMC, a manufacturer in Shih’s own country, earned a net margin of 51 percent in the third quarter of 2026, comparable to the most profitable of the designers it serves and far above most of them. Its value lay in the hardest part of making, where the learning was deepest and the competitors fewest. Yet the invisibility persists. Users know the processor by its designer’s brand. Mead and Conway’s design rules were built precisely so that designers would not need to think about the factory, and the abstraction worked so well that the factory dropped out of public view until a pandemic shortage and a war threat brought it back.

The strongest objection

The objection is that invisibility is the abstraction succeeding, not a moral failing. Users do not need to know who made their chip any more than they need to know who generated their electricity. TSMC is well paid, its workers among the best paid in Taiwan, and its shareholders well rewarded; the market has not undervalued it. Most manufacturing does still earn thin margins, so the smile curve describes the general case and TSMC is the exception that comes from being the only one at the frontier. These points are fair. The deeper claim is not that the maker is underpaid but that a culture that credits ideas over making will misjudge where its capabilities lie, as many countries did when they let their factories go and later found that their designers depended on someone else’s floor.

What it means today

Organisations should know who actually makes what they sell and what those makers know. Valuing the people who make things, inside and outside the organisation, is partly a matter of respect and partly one of strategy, because the knowledge that is hardest to replace often sits in the work that is least visible. Abstractions that let us forget the making are useful. It is dangerous to forget that we have forgotten it.

Bringing the threads together

The foundry began as a division of labour that took design out of the factory, and it succeeded because it created the market Smith said such a division needs. It grew by learning, owning its processes and its curves of experience, and the deepest of that learning was tacit, carried by people and hard to move. It won customers with a commitment, giving up a category of profit so that designers could trust it with their secrets, and that trust shaded, as it grew, into a dependence that no promise could fully address. Its leaders resisted the fear that a downturn invites, and sustained a vigilance that yields demand, at a cost to the people who keep it. And the concentration that efficiency produced became, in the hands of governments, a lever and a shield at once, while the people who made the world’s most advanced chips remained, to most of the people who used them, invisible.

A factory for other people’s ideas

When Chang wrote his plan in four days in 1985, he described it as the least bad option for a place with almost nothing to build on. Its premise was humble: designers had ideas and lacked factories, and a company that made their chips faithfully, and never competed with them, might find enough of them to survive. The premise turned out to be one of the most consequential business ideas of the late twentieth century. Nvidia, Qualcomm, Apple’s chip division and many of the companies building artificial intelligence exist in their present form because someone else would make what they drew.

The history suggests some working principles. Decide what to outsource by asking what you will stop learning, not only what you will stop paying for. Own the learning that compounds. Build trust into the structure of a relationship, by what you give up, rather than relying on goodwill. In a downturn, measure losses against the position you will need when it ends. Take small anomalies seriously, and protect the attention of the people who notice them. And remember that every efficient concentration is also a single point of failure, which someone, eventually, will try to use. The foundry made it possible to design without making. It did not make making matter less. It gathered the making into fewer hands, and the world now depends on how carefully those hands work.

Selected bibliography

• Smith, Adam. An Inquiry into the Nature and Causes of the Wealth of Nations. 1776.
• Babbage, Charles. On the Economy of Machinery and Manufactures. Charles Knight, 1832.
• Marx, Karl. Capital: A Critique of Political Economy, vol. 1. 1867.
• Angell, Norman. The Great Illusion. 1909 (first published as Europe’s Optical Illusion).
• Wright, T. P. “Factors Affecting the Cost of Airplanes.” Journal of the Aeronautical Sciences 3, no. 4 (1936): 122–128.
• Hirschman, Albert O. National Power and the Structure of Foreign Trade. University of California Press, 1945.
• Mackworth, N. H. “The Breakdown of Vigilance during Prolonged Visual Search.” Quarterly Journal of Experimental Psychology 1, no. 1 (1948): 6–21.
• Polanyi, Michael. Personal Knowledge. University of Chicago Press, 1958.
• Arendt, Hannah. The Human Condition. University of Chicago Press, 1958.
• Schelling, Thomas C. The Strategy of Conflict. Harvard University Press, 1960.
• Arrow, Kenneth J. “The Economic Implications of Learning by Doing.” Review of Economic Studies 29, no. 3 (1962): 155–173.
• Polanyi, Michael. The Tacit Dimension. Doubleday, 1966.
• Abernathy, William J., and Kenneth Wayne. “Limits of the Learning Curve.” Harvard Business Review 52, no. 5 (1974): 109–119.
• Staw, Barry M. “Knee-Deep in the Big Muddy: A Study of Escalating Commitment to a Chosen Course of Action.” Organizational Behavior and Human Performance 16, no. 1 (1976): 27–44.
• Keohane, Robert O., and Joseph S. Nye. Power and Interdependence. Little, Brown, 1977.
• Klein, Benjamin, Robert G. Crawford and Armen A. Alchian. “Vertical Integration, Appropriable Rents, and the Competitive Contracting Process.” Journal of Law and Economics 21, no. 2 (1978): 297–326.
• Kahneman, Daniel, and Amos Tversky. “Prospect Theory: An Analysis of Decision under Risk.” Econometrica 47, no. 2 (1979): 263–291.
• Mead, Carver, and Lynn Conway. Introduction to VLSI Systems. Addison-Wesley, 1980.
• Williamson, Oliver E. The Economic Institutions of Capitalism. Free Press, 1985.
• Kahneman, Daniel, and Dan Lovallo. “Timid Choices and Bold Forecasts: A Cognitive Perspective on Risk Taking.” Management Science 39, no. 1 (1993): 17–31.
• Weick, Karl E., and Karlene H. Roberts. “Collective Mind in Organizations: Heedful Interrelating on Flight Decks.” Administrative Science Quarterly 38, no. 3 (1993): 357–381.
• Vaughan, Diane. The Challenger Launch Decision. University of Chicago Press, 1996.
• Baldwin, Carliss Y., and Kim B. Clark. Design Rules, vol. 1, The Power of Modularity. MIT Press, 2000.
• Collins, H. M. “Tacit Knowledge, Trust and the Q of Sapphire.” Social Studies of Science 31, no. 1 (2001): 71–85.
• Addison, Craig. Silicon Shield: Taiwan’s Protection against Chinese Attack. Fusion Press, 2001.
• Weick, Karl E., and Kathleen M. Sutcliffe. Managing the Unexpected. Jossey-Bass, 2001.
• Chang, Morris. Oral history, interviewed by Alan Patterson, 24 August 2007. Computer History Museum, reference 102658129.
• Pisano, Gary P., and Willy C. Shih. “Restoring American Competitiveness.” Harvard Business Review 87, nos. 7–8 (2009): 114–125.
• Crawford, Matthew B. Shop Class as Soulcraft. Penguin Press, 2009.
• Farrell, Henry, and Abraham L. Newman. “Weaponized Interdependence: How Global Economic Networks Shape State Coercion.” International Security 44, no. 1 (2019): 42–79.
• Miller, Chris. Chip War: The Fight for the World’s Most Critical Technology. Scribner, 2022.
• TSMC. Quarterly results and management reports, 2009–2026.

Research limitations

Much of the account of TSMC’s founding rests on Morris Chang’s own recollections, chiefly his 2007 oral history for the Computer History Museum, which is a participant’s account given twenty years after the events; his autobiography is cited here through a published summary, not the original Chinese text. Chang dated the RCA licence to 1975, while ITRI and most histories give 1976. The founding shareholdings of the government’s Development Fund and Philips vary slightly between sources, and one later account says Philips once held a majority, which this essay does not adopt. The report that ITRI’s demonstration factory exceeded RCA’s yields comes from Taiwanese accounts. The reasons TSMC gave for declining IBM’s process in 2000 are the company’s own. The date Burn Lin first proposed immersion to the industry could not be confirmed; the essay relies on contemporary trade press for TSMC’s 2003 orders. Accounts of the 2009 dismissals differ in the number of employees involved. Apple’s split of chip orders between TSMC and Samsung in 2015 and 2016 was never disclosed and is known only from analysts’ reports. Intel’s explanations of its 10-nanometre delays come from executives’ public remarks and press accounts, and the internal decision process is not documented. Estimates of Taiwan’s share of advanced chip capacity differ by year and method: the 92 percent figure is for 2019, and a later study put it at 69 percent for 2022. Market shares are TrendForce estimates. Arizona investment totals are company announcements of plans, not money spent. The labour disputes in Arizona involve contested claims by the company and by unions, and the discrimination lawsuit had not been decided at the time of writing. Capital spending guidance for 2026 was revised during the year and is reported here from secondary sources. The philosophical connections to Smith, Babbage, Baldwin and Clark, Pisano and Shih, Wright, Arrow, Henderson, Abernathy and Wayne, Polanyi, Collins, Williamson, Klein, Crawford and Alchian, Schelling, Kahneman, Tversky, Lovallo, Staw, Mackworth, Weick, Sutcliffe, Roberts, Vaughan, Shewhart, Montesquieu, Kant, Hirschman, Keohane and Nye, Farrell and Newman, Angell, Marx, Arendt, Matthew Crawford and Shih are the author’s interpretation; their works are paraphrased apart from short, attributed phrases. The decisions in the decisions box are judged by the author, separating what could be known at the time (process) from what happened (call).

Footnotes

1. Morris Chang, oral history interviewed by Alan Patterson, Computer History Museum, recorded 24 August 2007. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

2. TSMC, “TSMC Reports Third Quarter EPS of NT$29.43,” 15 October 2026. ↩ ↩²

3. TrendForce, “Global Foundry Revenue Approaches US$53.49 Billion in 2Q26,” 9 September 2026. ↩

4. Semiconductor Industry Association and Boston Consulting Group, Strengthening the Global Semiconductor Supply Chain in an Uncertain Era, April 2021; for the 2022 figure, Communications Today, summarising the 2024 SIA and BCG report. ↩

5. Wikipedia, “Jerry Sanders (businessman)”; TechPowerUp, June 2017, on the GlobalFoundries spin-off and its Abu Dhabi investors. ↩ ↩²

6. “MOSIS: the 1980s DARPA silicon broker,” The Good Science Project; Wikipedia, “MOSIS”. ↩

7. IEEE, “Milestones: The Field Programmable Gate Array (FPGA), 1984,” Engineering and Technology History Wiki; Wikipedia, “Xilinx”. ↩

8. “A Short History of Semiconductor Technology in Taiwan during the 1970s and the 1980s,” Taiwan Insight, May 2024; “The institute behind Taiwan’s chip dominance,” Asterisk. ↩

9. Karina Bao, “How TSMC was funded, VC-free,”; Wikipedia, “TSMC”. ↩

10. TechInsights, Logic blog on TSMC’s process history; TSMC, “TSMC to Decommission Fab 1,” 2001. ↩

11. “Summary of Dr. Morris Chang’s Autobiography (Volume 2),” 2024. ↩

12. TSMC, “Values”. ↩

13. TSMC, “TSMC First Taiwan Company to List on NYSE,” October 1997; TSMC, 2013 annual report, notes to financial statements. ↩

14. TSMC, “Philips Completes Exit of TSMC Shareholding,” August 2008. ↩

15. “TSMC founder Morris Chang,” Acquired podcast interview. ↩ ↩²

16. “TSMC weighs pain versus gain of 130-nm technology,” Design & Reuse, 2000. ↩

17. Infineon, “IBM, Infineon, UMC Roll Out World’s Most Advanced Chip Foundry Manufacturing Process,” 2000; UMC, press release, 28 November 2000. ↩

18. “Asia foundries envision cloudy ’03,” Design & Reuse; EE Times, 2003. ↩

19. “TSMC preps 193nm immersion ‘risk’ production,” Solid State Technology, 2004; EE Times, on Burn Lin and 157-nanometre lithography. ↩ ↩²

20. “ASML sells first immersion scanner,” Silicon Semiconductor, December 2003. ↩

21. TSMC, Form 6-K, first-quarter 2009 results, April 2009. ↩

22. TSMC, Form 6-K, “Dr. Chang Returns as Chief Executive Officer”, June 2009. ↩

23. Karina Bao, “The memory of layoffs reverberates for decades at TSMC”; Chang’s account in the Acquired interview. ↩

24. TSMC, fourth-quarter 2009 conference call transcript and third-quarter 2009 transcript; third-quarter 2010 call, transcript. ↩

25. “TSMC reportedly sole supplier of next-gen iPhone A-series chip,” AppleInsider, February 2016; “Will packaging make the difference for TSMC?” Solid State Technology, April 2016. ↩

26. ASML, “TSMC joins ASML’s Customer Co-Investment Program for Innovation,” 5 August 2012. ↩

27. TSMC, “TSMC’s N7+ Technology is First EUV Process Delivering Customer Products to Market in High Volume,” 7 October 2019. ↩

28. “Intel’s 10nm Is Broken, Delayed Until 2019,” Tom’s Hardware, April 2018; “Intel: EUV-Enabled 7nm Process Tech is on Track,” AnandTech, December 2018. ↩

29. “Intel’s 7nm chips are delayed,” The Register, 24 July 2020. ↩

30. “GlobalFoundries Halts 7-Nanometer Chip Development,” IEEE Spectrum, August 2018. ↩

31. “Samsung kicks off mass production of 3 nm chips,” Korea Economic Daily, 30 June 2022. ↩

32. “TSMC says started mass production of ‘most advanced’ 2nm chips,” eNCA, from AFP. ↩

33. “TSMC to put USD 12 bln into building 5 nm chip plant in Arizona,” Telecompaper, May 2020; “Trump’s Whiplash on China Leaves Chip Industry Spinning,” Bloomberg, 16 May 2020. ↩

34. “TSMC won’t sell chips to Huawei in fourth quarter,” Caixin, October 2020; “TSMC cuts off client after discovering chips diverted to Huawei,” BNN Bloomberg, October 2024. ↩ ↩²

35. TechInsights, “TechInsights finds SMIC 7nm (N+2) in Huawei Mate 60 Pro,” September 2023. ↩

36. Review of Craig Addison’s Silicon Shield, Taipei Times, May 2009; US Institute of Peace. ↩

37. “TSMC evacuates fabs and suspends construction in earthquake aftermath,” DatacenterDynamics, April 2024; TechRadar. ↩

38. Nathan Gardels, “The Cost Of Deglobalization,” Noema, quoting Chang’s remarks of December 2022. ↩

39. “US finalizes $6.6 billion chips award for TSMC,” Korea Times, 15 November 2024. ↩

40. “TSMC Arizona chip plant delays,” Quartz; “‘They would not listen to us’: inside Arizona’s troubled chip plant,” AOL, from the Financial Times. ↩

41. “US staff sue TSMC,” The Register, November 2024; “Discrimination lawsuit refiled against TSMC Arizona,” Arizona’s Family, July 2025. ↩

42. City of Phoenix, “TSMC Announces Additional $100 Billion Investment in Arizona,”; “TSMC adds $100 billion to Arizona investment,” Quartz, July 2026; CNBC, 20 July 2026. ↩

43. “Taiwan will invest $250 billion in U.S. chipmaking under new trade deal,” CNBC, 15 January 2026. ↩

44. “Intel announces $8.9 billion investment from US government, which will own 9.9% of chipmaker,” Yahoo Finance, August 2025. ↩

45. “TSMC: Short Supply of HPC GPUs to Persist for 1.5 Years,” AnandTech, September 2023; TrendForce. ↩

46. “TSMC announces 2026 capex spend of $56bn,” DatacenterDynamics, January 2026. ↩

47. “Taiwan court hands out jail terms of up to 10 years in TSMC trade secrets case,” MarketScreener, from Reuters, April 2026; Focus Taiwan, July 2026. ↩