The argument of this essay is narrow and specific. It is that the Taiwan Semiconductor Manufacturing Company, TSMC, is now the single point in the global semiconductor supply chain whose disruption would do the most damage to the world’s most economically and strategically important computing products. The same disruption would hit American technology giants harder than those giants themselves can hit back. It is not an argument that TSMC is the most innovative chip company, the most profitable, or the best known. By market capitalisation NVIDIA passed US$4 trillion in 2025 and US$5 trillion later that year, and Apple, Microsoft, Alphabet, Amazon, Meta and Broadcom are individually worth more than TSMC. The argument is about a particular kind of centrality, the kind that sits between a design and a physical object.
By “essential” I mean that, without TSMC, the leading-edge logic chips at the heart of mainstream American technology products could not, today, be made at scale by anyone else. By “American technology giants” I mean Apple, NVIDIA, Microsoft, Alphabet, Amazon, Meta, Intel, Qualcomm, AMD and Broadcom. The argument is not that TSMC has no competitors. Samsung Foundry is a serious second, GlobalFoundries and United Microelectronics Corporation cover older nodes, Intel is rebuilding a foundry business, and SMIC in mainland China is climbing up the curve. The argument is that none of them can replicate, in 2026, what TSMC does at the most advanced nodes that Apple, NVIDIA and the rest depend on.
How TSMC began inside a Taiwanese industrial policy
TSMC did not begin as a private startup in a garage. It began as a piece of industrial policy. The Industrial Technology Research Institute, ITRI, was set up by the Taiwanese government in 1973 with the explicit purpose of building technological capability that the local economy did not yet have. In 1980 ITRI spun out the United Microelectronics Corporation, UMC, the island’s first semiconductor company. The Hsinchu Science Park, the high-tech enclave that straddles Hsinchu City and Hsinchu County in northern Taiwan, was formally opened by the government on 15 December 1980 as a home for the new industry.
TSMC itself was founded in 1987. Morris Chang, who had spent nearly three decades at Texas Instruments in the United States and then run ITRI, was the founder. The new company was structured as a joint venture between the Taiwanese government, ITRI and a group of private investors, with the Dutch electronics group Philips taking a 27.5 percent stake and providing the production technology, intellectual property and patents that made the first line possible. Philips also provided TSMC’s first chief executive, James E. Dykes, who had worked at Philips North America. Chang took over as chief executive after Dykes left and ran the company for the next eighteen years.
The choice of model mattered more than the choice of location. TSMC was set up from the start as a pure-play foundry, a contract manufacturer that would make chips designed by other companies and would never sell chips under its own brand. At the time this was unusual. Most chip companies of that era, including Intel, Texas Instruments, NEC, Hitachi and Toshiba in Japan, and the early Samsung semiconductor operation, were integrated device manufacturers, IDMs, that both designed and made their own products. The pure-play bet was not obvious, and for several years the new Taiwanese firm lost money. What changed was that the economics of a new fab, a single fabrication plant, made it prohibitively expensive for any one chip design company to keep building the latest generation of factory every two or three years. The foundry offered a way to spread that cost across many customers.
The pure-play foundry model and why it won
The foundry model rests on a simple division of labour. A fabless designer such as AMD, NVIDIA, Apple, Qualcomm, Broadcom or MediaTek handles the chip architecture and the layout of the transistors. A foundry handles the manufacturing process, the photolithography, the etch, the deposition, the metrology, the packaging. The two halves depend on each other, but they do not own each other.
This separation had three consequences that compounded over time. First, the foundry could accumulate process know-how across many customers and many product types, which is what gives a leading-edge fab its yield advantage. Second, the fabless designer could focus capital on design talent and on EDA licences rather than on billion-dollar fabs. Third, the foundry could charge for capacity and for process recipes, turning what had been a cost centre inside an IDM into a revenue line on its own.
TSMC was the first company to commit fully to that model, and for two decades it had little serious competition at the leading edge. Samsung Foundry followed with a similar offering. GlobalFoundries, created in 2009 when AMD spun out its manufacturing arm to the Abu Dhabi fund Mubadala, decided in 2018 to stop developing nodes beyond 14 nanometres and to focus on mature and specialty processes, a recognition that the leading-edge race had become uneconomic for anyone outside the top two or three players. UMC made the same kind of choice. Intel, the only American IDM with a serious process roadmap, kept its fabs in house for its own CPUs and only dipped in and out of contract work.
The pure-play bet made TSMC into a piece of shared infrastructure. By 2026 the company was reporting that approximately 70 percent of the world’s merchant foundry revenue ran through its fabs, and that it served as the main supplier for NVIDIA, Apple, Broadcom and Qualcomm. That share is not a marketing claim. It reflects a decade-by-decade accumulation of customer relationships, design rules, IP libraries, electronic design automation tools, photoresist chemistries, gas and slurry suppliers, and process engineers, none of which can be moved in a year.
The customer concentration on the leading edge
The leading edge is where the dependency is sharpest. Apple’s A-series processors for the iPhone, the M-series processors for the Mac and the iPad, and the S-series processors for the Apple Watch are all manufactured by TSMC. The Apple A14 Bionic, which appeared in the iPhone 12 in 2020, was the first mass-market 5 nanometre chip in the world, and the Apple M1, the first of the Apple silicon Mac processors, was manufactured at the same node. The Apple M2 followed on TSMC’s N5P, an enhanced 5 nanometre recipe. Every generation of Apple’s custom silicon has been a TSMC part.
NVIDIA’s data-centre GPUs are also TSMC parts. The A100 accelerator of 2020, the H100 of 2022, the Blackwell generation of 2024 and 2025, all depend on the same Taiwanese fabs. AMD’s Ryzen and EPYC CPUs and its Radeon GPUs are largely manufactured by TSMC; AMD’s most successful product generations of the last decade, including the Ryzen 7000 series and the EPYC Rome server line, were all built on TSMC’s 7 nanometre and 5 nanometre nodes. Qualcomm’s Snapdragon system-on-chips for smartphones are TSMC products, as are the high-end mobile processors from MediaTek and most of Broadcom’s networking and custom silicon.
The dependency reaches into the American cloud, which is the other half of the story. The hyperscalers, Amazon Web Services, Microsoft Azure, and Alphabet’s Google Cloud, do not run on chips they designed themselves. They run large numbers of NVIDIA GPUs that were manufactured by TSMC. Apple’s Mac, iPhone and Watch business, the bulk of high-end Android smartphones, the bulk of the global AI compute footprint, and a large share of the world’s networking and data-centre silicon all converge on a small number of factories in Hsinchu, Tainan and, more recently, Central Taiwan Science Park near Taichung.
Why the American fabless giants cannot easily move production off Taiwan
The first reason is capital intensity. A modern fab costs on the order of twenty billion US dollars before it produces a working wafer. TSMC’s capital expenditure guidance for 2026 was raised in mid-2026 from a range of US$52 billion to US$56 billion to a range of US$60 billion to US$64 billion, and the company announced in July 2026 an additional US$100 billion of investment in its United States operations on top of that. No fabless American company, and arguably no single American company other than Intel, can match that scale of capital commitment on its own.
The second reason is extreme ultraviolet lithography. EUV, the 13.5 nanometre wavelength light source that lets foundries print the smallest features at 5 nanometre, 3 nanometre and below, is produced by a single company. That company is the Dutch firm ASML, originally spun out of Philips and ASM International in 1984. ASML is, in the words of its own industry, the only firm in the world that produces and sells EUV systems for chip production, and Intel, Samsung and TSMC all bought equity stakes in it during a 2012 co-investment programme to fund its development. Without EUV, foundries cannot economically reach the most advanced nodes.
The third reason is the learning curve. A leading-edge fab does not produce good chips on day one. It takes years of incremental tuning to bring yields up to commercially viable levels, and the tuning is specific to a particular set of tools, materials and process recipes. The yield learning that TSMC has accumulated across N7, N5, N3 and N2 is the product of a decade and a half of continuous effort, not a thing that can be bought or replicated quickly.
The fourth reason is the surrounding ecosystem. A modern fab needs a dense network of suppliers, photoresist manufacturers, gas and chemical suppliers, tool service engineers, water treatment, power, and process engineers who can be recruited and retained in one place. The Taiwanese cluster around Hsinchu, Tainan and Taichung is dense in a way that is not yet matched in Arizona, Kumamoto or Dresden.
Intel’s foundry bid and its limits
Intel is the only American company that combines leading-edge process design with chip design of its own, and it has been trying to break into contract manufacturing for more than a decade. In 2013 the company signed a foundry deal with Altera to make 14 nanometre parts. It then built a small custom foundry business around FPGA makers Achronix, Tabula, Microsemi and Panasonic and the network processor firm Netronome. The custom foundry effort was closed in 2018 as Intel struggled with its own process delays.
In March 2021 Pat Gelsinger returned to Intel as chief executive and announced the IDM 2.0 strategy. The plan included a new Intel Foundry Services business, a renewed commitment to internal process leadership, an Ohio fab megaproject that was originally announced at US$20 billion, and a separate German megaproject at Magdeburg that has since been deferred and partly cancelled. The Intel 18A process, the company’s 1.8 nanometre-class node, entered high-volume manufacturing in late 2025. In October 2023 Intel confirmed it would be the first commercial user of ASML’s high-NA EUV tools.
The execution has been uneven. In August 2024 Intel reported a US$1.6 billion loss for the second quarter, announced 15,000 job cuts and warned of a long road back to process leadership. In July 2025 the company announced a further 24,000 job cuts, about 15 percent of the workforce, and said it would scrap tens of billions of dollars of planned European facility investments. In September 2025 NVIDIA invested US$5 billion in Intel to jointly develop data centre and personal computing CPUs, a deal that speaks more to NVIDIA’s need for an American partner for some products than to Intel’s success in winning external foundry customers. In October 2025 Intel confirmed it was in early talks to add AMD as a foundry customer.
The lesson is the gap between foundry announcements and foundry production. Building a foundry customer base is a multi-year effort. Even if every reported Intel 18A customer commitment materialises, Intel will still be a generation behind TSMC at the most advanced nodes, and its process track record has been poor enough that customers discount its roadmap heavily.
The CHIPS Act, export controls, and the new geography of fabs
The US policy response to the concentration of leading-edge manufacturing in Taiwan has two halves. The first is the CHIPS and Science Act, signed by President Biden on 9 August 2022. The act authorises roughly US$280 billion in new funding and appropriates US$52.7 billion, of which US$39 billion is for chip manufacturing subsidies and a 25 percent investment tax credit for manufacturing equipment, and US$13 billion is for research and workforce training. The second is a rolling programme of export controls administered by the Bureau of Industry and Security at the US Department of Commerce, beginning with the October 2022 rule that restricted the export of certain advanced computing chips and semiconductor manufacturing items to mainland China, with further updates in October 2023 and December 2024 that tightened the thresholds and aligned allied controls.
For TSMC, the consequence has been a wave of offshore investment. The Phoenix, Arizona site was first announced in May 2020 at a US$12 billion scale, then tripled to about US$40 billion in December 2022. In April 2024 the Biden administration announced a preliminary award of up to US$6.6 billion in direct funding and up to US$5 billion in loans under the CHIPS Act to support more than US$65 billion of investment across three Phoenix fabs. In November 2024 the Commerce Department finalised the US$6.6 billion direct funding award and TSMC agreed to deploy its most advanced A16 manufacturing technology in Arizona and to forgo stock buybacks for five years under an upside sharing arrangement. In July 2026 the company announced an additional US$100 billion of US investment, taking the planned total to about US$265 billion across twelve fabrication and advanced-packaging facilities in Arizona.
In Japan, TSMC established Japan Advanced Semiconductor Manufacturing, JASM, a subsidiary with Sony, Denso and Toyota as minor shareholders. Fab 23 in Kikuyo, Kumamoto, began commercial operations in December 2024, produces 12, 22 and 28 nanometre parts, and was built at a cost of US$8.6 billion with a 476 billion yen subsidy from the Ministry of Economy, Trade and Industry. A second Japanese fab is being built alongside it for 6 and 12 nanometre production at a cost of about US$13.9 billion, with 732 billion yen of Japanese government funding.
In Europe, TSMC committed in August 2023 to a joint venture with Bosch, Infineon and NXP for a fab in Dresden, Germany, under the name European Semiconductor Manufacturing Company, ESMC. The total project value is more than €10 billion, with TSMC contributing €3.5 billion in equity, €5 billion of German government subsidy, and a 10 percent stake from each of the three European partners. The fab was planned to come fully online in 2029 with a monthly capacity of 40,000 twelve-inch wafers, producing 28 nanometre and 12 and 16 nanometre FinFET parts. The building cannot host ASML’s EUV scanners because they are too tall, and ground was not broken until August 2024.
Taiwan concentration, power and water, and what diversification cannot do
None of this changes the underlying geography. TSMC’s most advanced production, the 3 nanometre and 2 nanometre nodes that Apple, NVIDIA and the hyperscalers depend on, still runs overwhelmingly from Hsinchu, Tainan and Central Taiwan Science Park. The Arizona fabs, even at full ramp, will produce a fraction of TSMC’s total output and will run a node or two behind the leading edge for some years. The Japanese fabs run mature and mid-range processes. The Dresden fab runs older processes still, and without EUV.
The concentration has prompted the argument, sometimes called the silicon shield, that Taiwan’s role in chip manufacturing is so important to the global economy that it deters a Chinese attempt to take the island by force, or at least guarantees that the United States and its allies would intervene. Critics point out that the same concentration creates a single point of failure that a blockade, a typhoon, an earthquake, or a prolonged drought could exploit. Taiwan’s power grid and water supply are already under strain, and TSMC uses very large volumes of both. Geographic diversification reduces this risk but does not eliminate it, and a fab in Phoenix does not move a leading-edge process across the Pacific.
The mainland Chinese alternative and why SMIC cannot replace TSMC
The mainland Chinese answer to TSMC is the Semiconductor Manufacturing International Corporation, SMIC, which was founded in 2000 and is partially owned by the Chinese state. As of 2024 SMIC was the world’s third-largest contract chip maker, behind TSMC and Samsung Foundry, and the largest in mainland China. It offers processes from 350 nanometre up to 7 nanometre.
SMIC’s most prominent recent product is the Kirin 9000s system-on-chip designed by HiSilicon, the chip design arm of Huawei, which appeared in Huawei’s Mate 60 Pro smartphone in August 2023. The launch was a political event as much as a commercial one, because it followed several rounds of US export controls aimed at preventing Chinese firms from accessing the most advanced chip technology, including a specific cutoff that in 2020 had stopped HiSilicon from sourcing chips from TSMC.
The Kirin 9000s is widely reported by analysts, including TechInsights, as a 7 nanometre-class part manufactured by SMIC using deep ultraviolet multi-patterning rather than EUV lithography. The absence of EUV access is the central constraint on SMIC’s leading-edge output. EUV systems are sold only by ASML, and the Dutch and US governments have agreed to restrict sales of the most advanced lithography tools to mainland China. Without EUV, a foundry can in principle reach 7 nanometre by other means, but only at lower yield and at higher cost, and the next nodes are not realistically accessible. SMIC is reported by the Financial Times to have been aiming for a 5 nanometre offering, but the gap between SMIC’s leading edge and TSMC’s is measured in process generations and in unit economics, not in marketing numbers.
There is no public evidence that any mainland Chinese foundry can today produce a 3 nanometre chip at commercial scale, and there is no public evidence that any such foundry will be able to do so in the next several years.
Closing
Strip the argument back to its mechanism. The world’s most advanced logic chips are extraordinarily expensive to design and extraordinarily expensive to manufacture. The expense of manufacturing is concentrated in a small number of factories, in a small number of countries, using a small number of lithography systems from a single supplier. The leading-edge logic process requires extreme ultraviolet light, a per-tool yield learning curve measured in years, and a surrounding ecosystem of chemicals, gases, IP libraries, EDA tools and process engineers that has accumulated in one place over four decades. The American technology giants are fabless or near-fabless on the leading edge, and they depend on TSMC for the part of their products that is hardest to substitute. None of the alternative suppliers, not Intel, not Samsung, not SMIC, can replicate that combination in a year, and none of the alternative geographies, not Arizona, not Kumamoto, not Dresden, can replicate the surrounding ecosystem of a Taiwanese science park. Until one of those alternatives is built, the most strategically important chips in American and global computing will continue to be made in a small number of fabs on an island off the coast of mainland China. That is what the centrality of TSMC means in plain language.