TSMC
| Taiwan Semiconductor Manufacturing Company Limited | |
|---|---|
| Type | Public |
| Traded as | TWSE: 2330 NYSE: TSM |
| Industry | Semiconductors |
| Founded | 21 February 1987 |
| Founder | Morris Chang |
| Headquarters | Hsinchu Science Park, Hsinchu, Taiwan |
| Key people | C. C. Wei (Chairman and chief executive officer) Morris Chang (Founder) |
| Products | Contract manufacture of integrated circuits; advanced packaging |
| Employees | Approximately 80,000 |
| Website | tsmc.com |
Taiwan Semiconductor Manufacturing Company Limited, usually TSMC, is a Taiwanese contract manufacturer of integrated circuits headquartered at the Hsinchu Science Park. It is the largest and most advanced semiconductor manufacturer in the world and produces the great majority of the most sophisticated chips made anywhere.
The company was founded on 21 February 1987 by Morris Chang, who had spent 25 years at Texas Instruments and had been recruited by the Taiwanese government to develop the island's electronics industry. Its innovation was organisational rather than technical: it manufactured chips designed by other companies and designed none itself, which had not previously been a business.
That decision created the fabless model. Before TSMC, a company wishing to sell a chip had to own a fabrication plant, which by the 1990s cost billions of dollars and which only the largest could afford. After TSMC, a company needed only designers, which is why NVIDIA, Qualcomm, Broadcom, AMD, Marvell, Apple's silicon division and several thousand smaller firms exist.
TSMC manufactures for Apple, NVIDIA, AMD, Qualcomm, Broadcom, MediaTek, Intel, which has become a customer as well as a competitor, and a very large number of others. It holds approximately 60 to 70 per cent of the global foundry market by revenue and a far higher share, above 90 per cent by most estimates, of production at the most advanced process nodes.
The Taiwanese government is the largest individual shareholder through the National Development Fund, and the majority of the company is held by foreign investors. It ranked 31st in the Forbes Global 2000 in 2026, and integrated circuits are Taiwan's largest export by a very wide margin.
Its concentration in Taiwan, approximately 130 kilometres from mainland China, makes it the single most geopolitically significant company in the world, and the phrase silicon shield describes the argument that Taiwan's centrality to semiconductor supply deters an invasion.
History
Founding
Morris Chang was born in Ningbo in 1931, moved repeatedly during the Japanese invasion and the civil war, studied at Harvard, the Massachusetts Institute of Technology and Stanford, and joined Texas Instruments in 1958, where he rose to run the worldwide semiconductor business and where he is credited with having pioneered the practice of pricing aggressively ahead of the cost curve to gain share.
Passed over for the chief executive role, he left, ran General Instrument briefly, and in 1985 accepted an invitation from the Taiwanese government to lead the Industrial Technology Research Institute and to build a semiconductor industry.
Taiwan had electronics assembly and no chip manufacturing. Chang's assessment was that Taiwan could not compete in design, where the United States and Japan led, nor in memory, where Japan and Korea were investing enormous sums, and that its advantage was disciplined process manufacturing.
His proposal was a company that would manufacture and would not design, so that it would never compete with its customers. This was the essential point. The existing manufacturers with spare capacity, principally Intel and the Japanese firms, would take outside work when convenient and would prioritise their own products, and would see the customer's designs. A pure foundry would not.
TSMC was founded on 21 February 1987 with the Taiwanese government providing approximately 48 per cent of the capital through the Development Fund, Philips providing approximately 28 per cent together with technology and patent protection, and private Taiwanese investors the remainder. Philips's participation was decisive, because it supplied process technology and because its patent portfolio protected TSMC against the infringement claims an entrant would otherwise have faced.
Growth, 1987 to 2010
The early years were difficult. TSMC's process was several generations behind the leaders, its yields were poor, and the established companies did not take it seriously.
Its customers were the companies that had no alternative: small design firms with good ideas and no capital. NVIDIA, founded in 1993, Broadcom, Xilinx, Altera and Marvell all depended on it, and TSMC's success and theirs were the same phenomenon.
Chang's second insight was that a foundry should improve its process as fast as the integrated manufacturers, which required spending on research and on plants at a scale nobody expected a contract manufacturer to sustain. TSMC did, and by the early 2000s its process was competitive, and by about 2015 it was ahead of everyone except Intel, and by about 2018 it was ahead of Intel as well.
The company listed in Taipei in 1994 and on the New York Stock Exchange in 1997.
Chang retired as chief executive in 2005, returned in 2009 during the financial crisis when he judged that his successor had cut capital expenditure too far, and retired finally in 2018.
The advanced node era, 2010 to 2026
The critical commercial relationship was with Apple. Apple had used Samsung to manufacture its processors and was simultaneously in litigation with it over telephones; it moved to TSMC for the A8 in 2014 and has manufactured almost exclusively there since, which gave TSMC both enormous volume and a customer prepared to fund the development of each new process by committing to buy it.
The process sequence has been 28 nanometre from 2011, which was a very large commercial success; 16 and 12 nanometre; 10 and 7 nanometre from 2016 and 2018; 5 nanometre from 2020; 3 nanometre from 2022; 2 nanometre from 2025; and the A16 and A14 angstrom-class nodes announced for 2026 and 2028.
A13 was introduced at the North America Technology Symposium in April 2026 as a direct shrink of the A14 process announced in 2025, enabling more compact designs.
Competitors fell away. Intel stumbled at 10 nanometre, losing approximately four years and its process leadership. GlobalFoundries abandoned development beyond 14 nanometre in 2018, judging that it could not fund the next step. Samsung remains at the leading edge and has struggled with yields, which has directed customers including Qualcomm and NVIDIA toward TSMC.
Financial results have been extraordinary. The company reported a 35 per cent increase in fourth-quarter profit in January 2026, a record, and in July 2026 Reuters reported a 77 per cent year-on-year increase in second-quarter profit to a record NT$706.6 billion, approximately US$22 billion, exceeding expectations, with the company raising its outlook for the year.
The cause is artificial intelligence. Every accelerator NVIDIA sells, every custom chip the hyperscalers design, and the high-bandwidth memory packaging they require are manufactured at TSMC, and the company's advanced packaging capacity, not its transistors, has been the binding constraint on the industry's supply.
In June 2025 it established the TSMC-UTokyo Lab, its first joint research laboratory with a foreign university. In August 2025 it reported that two current employees and one former employee were suspected of misappropriating 2-nanometre process information, a matter referred to prosecutors. In May 2026 it announced the sale of up to 152 million shares in Vanguard International Semiconductor through block trade, reducing its holding in that mature-node foundry.
Technology
What a foundry does
A modern logic chip is built by depositing and removing layers of material on a silicon wafer in several hundred to more than a thousand steps, each of which must be performed to tolerances measured in atoms across a wafer 300 millimetres across, and each of which must be repeated hundreds of thousands of times a month without variation.
The difficulty is not conceptual. The physics is published and the equipment is bought from the same suppliers, principally ASML for lithography, Applied Materials, Lam Research and Tokyo Electron for deposition and etching, and KLA for inspection. Every manufacturer buys the same machines.
The difficulty is yield: the proportion of finished chips that work. A process with a defect rate that produces 90 per cent working chips is profitable and one producing 50 per cent is not, and the difference lies in thousands of individually small decisions about recipes, timings, cleanliness, metrology and the handling of the very large quantity of data each wafer generates.
TSMC's advantage is that it has been doing this for more manufacturers, across more designs, at higher volume, for longer than anyone, and that a foundry's yield learning applies across every customer's product while an integrated manufacturer learns only from its own.
Processes
FinFET, from the 16 nanometre generation, builds the transistor's channel as a vertical fin surrounded by the gate on three sides, which improves control over leakage as devices shrink.
Extreme ultraviolet lithography, used from 7 nanometre in a limited form and extensively from 5, uses light at 13.5 nanometres produced by vaporising tin droplets with a laser, reflected by mirrors of extraordinary precision because no material transmits at that wavelength. The machines are made only by ASML, cost approximately US$200 million each and more for the high-numerical-aperture versions, and are among the most complex objects manufactured.
Gate-all-around nanosheet transistors, introduced at 2 nanometre, surround the channel on all four sides, further improving control.
Advanced packaging has become as important as the transistors. CoWoS, chip-on-wafer-on-substrate, places a logic die and several stacks of high-bandwidth memory on a silicon interposer, which is how every artificial intelligence accelerator is built. TSMC's CoWoS capacity has been the constraint on the supply of those accelerators since 2023, and the company has expanded it repeatedly. SoIC bonds dies directly without solder, and InFO is used in Apple's processors.
Node names no longer describe any physical dimension. The 3 nanometre process contains no feature 3 nanometres across; the numbers are marketing designations indicating a generation, and comparisons between manufacturers' nodes require examining transistor density rather than the name.
Customers and the fabless ecosystem
TSMC manufactures for approximately 500 customers producing several thousand distinct products a year, and the relationships determine much of the semiconductor industry's structure.
Apple is the largest customer by revenue, accounting for approximately a quarter of TSMC's sales, and has first access to each new process node, which it obtains by committing volume and capital in advance. Its A-series and M-series processors, among the most sophisticated chips made, have been manufactured exclusively at TSMC since 2016.
NVIDIA is now the second largest and has been a customer since its founding in 1993, when it had no alternative. Its accelerators are manufactured on TSMC's advanced nodes and packaged using CoWoS, and its growth since 2023 has been the principal driver of TSMC's results.
AMD moved its manufacturing to TSMC after spinning off its own plants as GlobalFoundries in 2009, and its subsequent recovery against Intel is substantially attributable to having access to a better process than its competitor's.
Qualcomm, Broadcom, MediaTek, Marvell, Analog Devices and a very large number of others are substantial customers.
Intel has become a customer, manufacturing some products at TSMC while competing with it in foundry services, an arrangement that would have been unthinkable a decade ago and that is the clearest measure of what happened to Intel's process leadership.
The hyperscalers design their own accelerators, Google's Tensor Processing Units, Amazon's Trainium and Graviton, Microsoft's Maia and Cobalt, Meta's MTIA, and each is manufactured at TSMC, generally through Broadcom or Marvell as design partners.
The concentration works in both directions. TSMC depends on a small number of very large customers; those customers depend entirely on TSMC, since no alternative exists at the leading edge. The relationship is therefore cooperative rather than adversarial in a way that is unusual between a supplier and a buyer, with customers funding capacity, committing to volumes years in advance and in some cases prepaying.
The economics of a fabrication plant
A leading-edge fabrication plant costs approximately US$20 billion to US$30 billion and takes three to four years to build and qualify.
Its cost structure is almost entirely fixed. Once built, the marginal cost of a wafer is a modest fraction of the price, which means the plant is enormously profitable at high utilisation and loses money at low utilisation, and which is why the industry's cycles are severe.
The capital must be committed before the demand is known, for a process that has not been completed, for customers whose products do not yet exist. A manufacturer that builds too little loses customers permanently, because a customer who has redesigned for a competitor's process does not return; one that builds too much destroys its returns.
TSMC's ability to make those commitments correctly for thirty years, funded from its own cash flow rather than from state subsidy for most of that period, is the underappreciated part of its achievement.
Wafer pricing at the leading edge has risen substantially, with a 3 nanometre wafer reported at approximately US$20,000 against roughly US$5,000 at 28 nanometre, which reflects the increasing cost of equipment and of process complexity and which has changed the economics of chip design: a chip that costs several hundred million dollars to design and tape out can only be justified by very large volumes, which concentrates the industry further.
Mature nodes, from 28 nanometre upward, remain a very large business serving automotive, industrial, power management and display applications, where the advanced nodes are unnecessary and where the plants are fully depreciated. TSMC's sale of its Vanguard International Semiconductor holding in May 2026 reduces its exposure to that segment, in which Chinese capacity is expanding rapidly.
The equipment and materials supply chain
TSMC is the largest customer of an industry whose concentration exceeds its own.
ASML of the Netherlands is the only supplier of extreme ultraviolet lithography machines and holds an effective monopoly on the most advanced patterning. Its machines contain optics from Zeiss produced to tolerances that have no equivalent in any other manufactured object.
Applied Materials, Lam Research and Tokyo Electron supply deposition, etching and cleaning equipment; KLA supplies inspection and metrology; Nikon and Canon supply older lithography.
Materials include photoresists, principally from Japanese suppliers including JSR, Tokyo Ohka and Shin-Etsu; specialty gases; silicon wafers, from Shin-Etsu and SUMCO; and chemicals of extraordinary purity.
Electronic design automation software, from Synopsys, Cadence and Siemens, is required to design any chip and to verify it against TSMC's process rules, and the process design kits TSMC supplies to its customers are among its most closely held assets.
The dependencies run in every direction and are the mechanism through which export controls operate: American controls apply to ASML's machines because they contain American components and software, which is why a Dutch company's exports to China are determined in Washington.
Taiwan and the company
TSMC's relationship with Taiwan is unlike that of any comparable company with any state.
It accounts for a very large proportion of the Taiwan Stock Exchange's capitalisation, of the island's exports and of its corporate tax revenue. Integrated circuits are Taiwan's largest export category by a very wide margin, and semiconductors and their supply chain account for a substantial share of gross domestic product.
The Hsinchu Science Park, established in 1980 on the model of Stanford's industrial park and with returning Taiwanese-American engineers as its founding population, is the origin of the cluster, and the concentration of suppliers, equipment engineers, materials companies and trained staff within a small area is the practical reason TSMC's yields are what they are.
The company's demands on the island are substantial. It consumes a very large proportion of Taiwan's industrial electricity, which is a political issue given the phase-out of nuclear generation, and very large quantities of ultrapure water, which became critical during the drought of 2021 when the government diverted agricultural irrigation to the science parks.
Taiwanese graduates in engineering are recruited to TSMC in preference to almost any alternative, and the resulting concentration of talent has been described domestically as both the island's greatest asset and a distortion that has starved other industries.
Competition
Samsung Foundry is the only other manufacturer at the leading edge. It has the advantage of an integrated business supplying its own products and the disadvantage that customers compete with it, which is the problem TSMC was founded to avoid. Its yields at 3 nanometre have been reported as substantially worse than TSMC's, and customers including Qualcomm have moved work away.
Intel Foundry is the strategic attempt to make Intel a contract manufacturer, announced in 2021. It has secured some customers, has very large government support, and faces the same structural difficulty as Samsung's, compounded by Intel's financial position and by the years lost at 10 nanometre.
GlobalFoundries abandoned the leading edge in 2018 and competes profitably at mature nodes.
SMIC of China is the most-watched competitor for political rather than commercial reasons, and has produced 7 nanometre chips using older lithography with multiple patterning, at yields and costs that are not disclosed and that are probably poor. Export controls prevent it acquiring extreme ultraviolet equipment, which constrains it to a limited number of further generations at increasing cost.
UMC, Tower, PSMC and Vanguard compete at mature nodes.
The structural position is that TSMC has no competitor at the leading edge and that every attempt to create one requires a state to fund what a market will not.
Geography and expansion
TSMC's manufacturing is concentrated in Taiwan, principally at Hsinchu, Taichung and Tainan, with its most advanced plants in the Southern Taiwan Science Park.
International expansion has been driven by customer requests and by government subsidy rather than by economics, since TSMC's own position is that concentration in Taiwan produces the clustering, the supplier network and the labour availability on which its yields depend.
Arizona hosts plants at Phoenix supported by the American CHIPS and Science Act, with the first producing 4 nanometre chips from 2024 and further plants planned at 3 and 2 nanometre. The company has committed a very large sum, raised several times and reported above US$165 billion, and has described difficulties with construction cost, with labour practices and with regulatory process that have delayed it.
Japan hosts a plant at Kumamoto, opened in February 2024, built with Sony, Denso and Toyota as partners and with substantial Japanese government subsidy, producing mature nodes for automotive and image sensor customers, with a second plant following.
Germany hosts a plant at Dresden, announced in 2023 with Bosch, Infineon and NXP as partners and with European Union subsidy, for automotive and industrial chips.
China hosts plants at Nanjing and Shanghai producing mature nodes, constrained by American export controls which prevent the installation of advanced equipment.
The international plants produce a small proportion of output and are not expected to change the concentration substantially within a decade. The most advanced production remains in Taiwan, and the company's research and development is entirely there.
Geopolitics
TSMC is at the centre of the most significant technological conflict of the period.
The silicon shield is the argument that Taiwan's indispensability to semiconductor supply deters an invasion, because an attack would remove the chips on which the Chinese economy, the American economy and the militaries of both depend, and because the facilities could not be captured intact given the dependence on foreign equipment, software and service engineers.
The argument is contested. Its critics observe that it assumes rational calculation, that the plants are within missile range regardless, and that the dependence creates an incentive to seize rather than to deter. Morris Chang has described the concept as overstated.
American export controls, from October 2022 and tightened since, prohibit the supply of advanced chips and equipment to China, and TSMC must comply because it uses American equipment and software. It ceased manufacturing for Huawei in 2020 and has restricted advanced work for Chinese customers since.
Subsidy competition has produced the CHIPS Act in the United States, the European Chips Act, and Japanese, Korean, Indian and Chinese programmes, all attempting to relocate manufacturing that has concentrated for economic reasons. TSMC is the principal beneficiary and the principal reluctant participant.
Tariffs. The second Trump administration's threats of tariffs on imported semiconductors have been directed at accelerating American manufacturing, and TSMC's expanded Arizona commitment followed.
China's domestic effort through SMIC and others has advanced faster than American controls intended, reaching 7 nanometre production without extreme ultraviolet equipment, and remains several generations behind.
Corporate affairs
Morris Chang founded the company, led it from 1987 to 2005 and from 2009 to 2018, and is regarded in Taiwan as a national figure. His decision to make the company a pure foundry, and his willingness to spend on process development at a scale competitors regarded as reckless, are the company's two defining choices.
C. C. Wei has been chief executive since 2018 and chairman since 2024, having joined in 1998 and risen through operations. Mark Liu was co-chief executive and then chairman until 2024.
TSMC employs approximately 80,000 people, the great majority in Taiwan, and its working culture is demanding by international standards, with engineers on call and shift patterns that its American plants found difficult to replicate, which was among the sources of the Arizona delays.
The National Development Fund of Taiwan holds approximately 6 per cent and is the largest single shareholder; the majority is held by foreign institutions, principally through the American depositary receipts.
Capital expenditure runs at approximately US$40 billion a year or more, among the largest of any company in the world, and research and development at approximately US$6 billion.
Morris Chang
Morris Chang is among the small number of people who can be said to have created an industry, and his career is worth describing because the company's decisions follow from it.
He was born in Ningbo in 1931 and moved six times before he was eighteen, fleeing the Japanese invasion and then the civil war, through Nanjing, Guangzhou, Chongqing, Shanghai and Hong Kong. He went to the United States in 1949, studied mechanical engineering at Harvard and then at the Massachusetts Institute of Technology, failed the doctoral qualifying examination twice, and took a job in industry instead, which he has described as the best thing that happened to him.
He joined Texas Instruments in 1958, completed a doctorate at Stanford while working, and rose to run the worldwide semiconductor business. His principal contribution there was a pricing strategy: cutting prices ahead of the cost reductions that volume would bring, on the reasoning that share gained early compounds through the learning curve. It was aggressive, it was correct, and it is the standard method in the industry now.
He was passed over for the chief executive role, left in 1983, ran General Instrument for a year, and accepted the Taiwanese government's invitation in 1985.
His decision to build a pure-play foundry was taken because of what he could not do rather than what he could. Taiwan had no design talent, no memory expertise and no chance of competing with Japan on capital. What it could do was manufacture reliably, and manufacturing for others avoided competing with the customers whose designs would fill the plant.
He led the company from 1987 to 2005, returned in 2009 when he judged that his successor had cut capital expenditure too deeply during the financial crisis, and retired in 2018. He has since spoken frequently and bluntly on the industry, including scepticism about American reshoring, which he has described as expensive and futile, and about the silicon shield.
Sustainability
Semiconductor manufacturing is energy and water intensive, and TSMC's scale makes it a significant environmental actor in Taiwan.
Electricity consumption is very large and growing, and TSMC alone accounts for a substantial proportion of Taiwan's industrial demand. The island's phase-out of nuclear generation, completed in 2025, and its limited renewable resources make supply a political question, and TSMC has signed very large renewable power purchase agreements, including for offshore wind, and has committed to using entirely renewable electricity by 2040 and to net zero by 2050.
Water is required in enormous quantities and to extraordinary purity. The 2021 drought, the worst in Taiwan in half a century, required the government to divert agricultural irrigation to the science parks and TSMC to truck water in, an episode that made explicit a trade-off between agriculture and semiconductors that the island's water policy had not previously had to confront. The company recycles a very high proportion of its process water and has built reclamation plants.
Chemicals used in manufacturing include perfluorinated compounds with extremely high global warming potential, and the industry has committed to abatement.
The company reports comprehensively and is generally rated among the better performers in its sector, and the underlying fact remains that producing the most advanced chips in the world requires a great deal of energy and water in a place that has limited quantities of both.
Risks
TSMC's position carries risks of several distinct kinds.
Geopolitical. An invasion or blockade of Taiwan would remove the world's supply of advanced chips, and a great many analyses of the scenario conclude that the economic consequences would exceed those of any event since the Second World War. The company cannot mitigate this and its international expansion addresses a fraction of its capacity.
Seismic. Taiwan is on an active plate boundary. The earthquake of April 2024, the strongest in twenty-five years, damaged wafers in process and required the evacuation of plants; TSMC's facilities are engineered to very high standards and recovered within days, and a larger event would not necessarily be survivable.
Concentration of customers. A small number of customers account for a very large proportion of revenue, and a change in Apple's or NVIDIA's position would be material.
Cyclicality. The industry's demand is cyclical and the company's costs are fixed, which produces severe swings. The artificial intelligence demand of 2023 to 2026 has been unusually strong and is a function of capital expenditure decisions by a small number of buyers.
Technical. Each process generation is harder than the last, and the physical limits of transistor scaling are approaching. The industry's response, advanced packaging and three-dimensional integration, plays to TSMC's strengths, and the underlying difficulty is real.
Espionage. The August 2025 case involving two current and one former employee suspected of misappropriating 2 nanometre process information is one of several, and Taiwanese prosecutors have brought a number of cases involving Chinese recruitment of semiconductor engineers.
In summary
TSMC manufactures approximately 90 per cent of the world's most advanced semiconductors, employs 80,000 people, spends more on capital expenditure than almost any company in the world, and is almost entirely unknown to the people who use its products daily.
It exists because one man concluded in 1985 that Taiwan could not compete in chip design and could compete in chip manufacturing, and because he identified that a manufacturer which never competed with its customers would be trusted with designs that an integrated manufacturer would not.
That decision created the fabless industry, which created NVIDIA, Qualcomm, Broadcom, AMD's recovery and Apple's silicon, and which is therefore responsible for a substantial proportion of the value created in technology over thirty years.
The consequence is a dependency that every government now regards as intolerable and that none has a plan to resolve within a decade.
Assessment
TSMC's importance is difficult to overstate and is also difficult to see, because it makes no product anyone buys.
Its founding decision, to manufacture and not to design, created an industry. Every chip company founded since 1990 that does not own a plant exists because TSMC does, which includes the companies now most valuable in the world.
Its execution has been the most sustained manufacturing achievement of the modern era. Improving a process by a generation every two years for three decades, while raising yields on each, while serving hundreds of customers with different designs, is a discipline that its competitors, including Intel, which had every advantage and a decade's head start, have not matched.
Its position is now uncomfortable in a way that has nothing to do with its performance. A single company, on a single island, in a contested strait, manufactures essentially all of the most advanced chips in the world, and every government that has examined the dependency has concluded that it is unacceptable and that reducing it will cost hundreds of billions of dollars and take a decade.
TSMC's own view, expressed consistently, is that the concentration is the source of its capability rather than an accident of it, and that distributing manufacturing across subsidised plants in higher-cost countries will produce less capable manufacturing at greater expense. It is probably correct, and it is not an argument that governments weighing strategic risk are likely to accept.