In this article
Today we get into the link that holds up the entire public conversation about artificial intelligence and that almost nobody names in Spanish. A single company manufactures more than ninety per cent of the planet's most advanced AI chips. It isn't in California, nor in Cupertino, nor in Seattle. It's on an island 130 kilometres off the Chinese coast, in a country China claims and where the Chinese military manoeuvres regularly. It's called TSMC. Why does this matter to a ChatGPT user in Madrid? Because if TSMC stops for a week, the global AI sector chokes, and the planet's digital economy goes into convulsion. My opinion is biased from having spent the last five years obsessed with this chain. Form your own, but get informed first.
Since 2020 I've been reading every TSMC quarterly report and every serious article on the semiconductor supply chain. The conclusion is disproportionate to the space the topic occupies in the Spanish press. Contemporary generative artificial intelligence rests, physically, on the factories of a single company on a single island at a single geopolitical point on the map. Let's look at it slowly.
What TSMC does and what it doesn't do
The full name is Taiwan Semiconductor Manufacturing Company. Morris Chang founded it in 1987 in Hsinchu, in the north of the island, with the support of the Taiwanese government. The business model was new at the time: a company that would only manufacture chips, without designing them. What's called a foundry. The chips are designed by the client — Nvidia, Apple, AMD, Qualcomm, MediaTek, Broadcom; TSMC manufactures them.
That division of labour, which seemed minor in 1987, has proved decisive. Designing a chip and manufacturing it are two different disciplines requiring capital, knowledge and experience that few companies can hold in parallel. When Nvidia decided in the nineties to focus on design and subcontract manufacturing, it wasn't out of weakness: it was out of specialisation. The consequence thirty years later is that Nvidia designs the H100 and the B200, but the H100 and the B200 are made by TSMC.
The same goes for Apple. The MacBook Pro's M4 chip, the iPhone's A18, the custom chip the next Vision Pro will have: all designed in Cupertino, manufactured in Taiwan. AMD, the same. Qualcomm, the same. Only Intel keeps its own manufacturing for its main processors — and it's years behind on advanced nodes, to the point that it has begun making part of its own chips at TSMC.
The exact market share varies by node. On advanced nodes — the most recent processes, called 5 nanometres, 3 nanometres and the upcoming 2 nanometres that TSMC began producing in 2025 — TSMC controls around 90% of the global market. Samsung, its only serious competitor, has spent years trying to close the gap and doesn't manage it. Each technological generation takes longer to close in, not less. The inertia is so brutal that it's hard to think of a comparable industry.
Why nobody else can
The immediate question is why this effective monopoly hasn't been challenged. The answer has three components and it's worth looking at them separately.
The first component is capital. An advanced-node semiconductor fab — a fab, in the jargon — costs today between 15,000 and 25,000 million dollars to build and equip. That's the order of magnitude of Estonia's annual GDP. The companies that can afford that investment, recover it in a decade and then repeat it for the next node can be counted on two hands. TSMC, Samsung, Intel and little else. And only TSMC has the financial and execution discipline to do it regularly.
The second component is accumulated knowledge. A fab isn't built and switched on. It's built, switched on, fine-tuned over months, adjusted process by process, learned from with each manufacturing error, with every parameter documented. That learning can't be bought or reproduced by compressing time. TSMC has been learning for thirty-eight years. Any competitor starting from scratch today would take decades to match the yield per wafer and the defective-chip ratio.
The third component is the machinery. To manufacture at 3 or 2 nanometres you need machines called EUV lithography, extreme ultraviolet. A single company in the world makes them: ASML, in Veldhoven, the Netherlands. Each machine costs more than 300 million euros and ASML produces a few dozen a year. The whole AI chain therefore rests on two bottlenecks: TSMC for manufacturing, ASML for the machinery TSMC uses. ASML doesn't sell to just anyone; its advanced machines are subject to European and US export controls, and China hasn't been able to buy them since 2023.
The combination of the three factors produces the effective monopoly. It isn't through collusion, it isn't through an explicit regulatory barrier. It's through the structural difficulty of bringing capital, knowledge and machinery together in a single operation.
Where the factories are
The vast majority of TSMC's capacity is still in Taiwan, concentrated mainly in Hsinchu, Tainan and Taichung. There are three relevant geographic-diversification initiatives and it's worth knowing them, because the general-interest press cites them as if they were already operational reality, and they aren't.
The first is Arizona. TSMC announced in May 2020 a fab in Phoenix with an initial investment of 12,000 million dollars, later expanded to more than 65,000 million across three phases. The first phase began producing commercial wafers in late 2024, on the N4 node. The second phase is under construction and should produce N3 from 2027 or 2028. The third phase was announced in 2024 with the N2 node. The Arizona fab is real, but as of today it manufactures non-leading-edge nodes and will produce the advanced nodes several years behind Taiwan.
The second is Dresden, Germany. TSMC signed in 2023 a joint-venture agreement with Bosch, Infineon and NXP to build a factory called ESMC. Construction began in 2024. The planned nodes are mature — 28 to 12 nanometres — not advanced. The fab is meant to serve the European automotive and microcontroller industry, not generative AI.
The third is Kumamoto, in Japan, in collaboration with Sony and Denso under the name JASM. The first phase came into operation in 2024, with mature nodes too. The second phase is under construction.
The operational conclusion: the manufacturing of the most advanced AI chips — the ones in Nvidia's H100, H200 and B200, in the Tesla Dojo chips, in Google's TPUs — still depends almost exclusively on Taiwan's factories. The fabs outside Taiwan exist, are real, but don't produce the chips that matter for the AI sector, nor will they over the next five years. That's today's reality. The press that tells it as if it were something else lies out of ignorance, not bad faith.
The geopolitical risk, in its real size
Here comes the part hardest to narrate without alarmism or minimisation, so I'll tread carefully.
Taiwan is an autonomous island with its own democratic government. The People's Republic of China formally claims Taiwan as Chinese territory under the "one China" principle. That claim isn't rhetorical: the Chinese military conducts naval and air manoeuvres around the island with growing frequency since 2022, including simulated-blockade exercises in August 2022, May 2024 and October 2024. The United States, without officially recognising Taiwan as a state, maintains a formal defence commitment through the Taiwan Relations Act of 1979, though the exact nature of that commitment has always been deliberately ambiguous.
The situation is stable because the three parties — China, Taiwan, the United States — prefer it to be stable. China doesn't want to break TSMC because it depends on the global tech sector, where the chips made in Taiwan are also used. The United States doesn't want direct conflict with China because the economic consequences would be devastating. Taiwan obviously prefers not to be invaded. The balance holds as long as the costs of breaking it exceed the benefits of maintaining it.
There's a phrase that has circulated for years in foreign-policy circles, attributed with variants to Morris Chang and other Taiwanese executives: "Taiwan's best protection is being indispensable." The idea is that as long as TSMC manufactures chips critical to the global economy, no power will have an incentive to destroy that capacity. The doctrine is informally known as the silicon shield. It's a deliberate doctrine, not accidental.
But the shield cuts both ways. If the global sector managed to depend less on Taiwan — through the fabs in Arizona, Dresden, Kumamoto and others — indispensability would drop and the balance would change. That explains why successive Taiwanese governments have supported TSMC's geographic diversification, but only for the non-most-advanced nodes. The leading-edge nodes stay on the island.
The immediate consequence for the European user
At first glance it seems this matters to the average European user about as much as the geology of the Gulf's oil wells. In the short term, that's true: it affects nothing about the daily use of ChatGPT, Claude or Gemini. The subscriptions keep working, the models keep improving, the price stays relatively stable.
In the medium term, it matters more. The political decisions being made now — the European AI Act, the US export restrictions on China, the massive subsidies of the CHIPS and Science Act of 2022 (52,000 million dollars in the US), the European Chips Act of 2023 (43,000 million euros) — are a direct consequence of the concentration of manufacturing in Taiwan and the associated risk. The European user pays, via taxes, part of those subsidies, without the public debate having explained why.
In the long term, it matters even more. If a serious disruption ever occurs in the supply of Taiwanese chips — for military reasons, an industrial accident, an earthquake in a seismically active zone — the global economic consequences would be on the order of magnitude of the 2020 pandemic, possibly greater. Whole sectors would grind to a halt in a matter of weeks, not months. The price of electronics would rise dramatically. Investment in AI, already high, would be forced to reorient toward critical infrastructure with five-to-ten-year timelines before being able to absorb the shock.
This isn't alarmism. It's a baseline scenario every serious economic-intelligence agency — from the US Department of Commerce to the German Bundesnachrichtendienst to the Spanish CCN-CERT — has factored into its analyses.
The political question
This is personal opinion, but I hold it with the available literature. The current structure of AI chip manufacturing is geopolitically fragile and economically efficient. That's a bad deal in the long run. A society that delegates its technological backbone to a single company on a single island at a single point of conflict is accepting a structural vulnerability in exchange for low prices.
What Europe can do, within its means, is demand that the advanced fabs built on the continent with public money actually come to produce leading-edge nodes, not just mature ones. The Dresden fab as conceived doesn't solve the problem; it reinforces the dependence. The European fab that would solve it would need a direct partnership with TSMC for N3 or N2 nodes, and that requires a level of negotiation Europe's governments haven't shown they have.
What the United States is doing, with the Arizona fabs, is the right thing from a national-security standpoint, though with notable delays and overruns. American industrial policy has recovered in this sector an aggressiveness that seemed forgotten since the seventies.
What the individual reader can do is little in the short term, beyond understanding the system. But understanding the system changes the vote, changes expectations about regulation and changes the public conversation. Today that conversation centres on whether ChatGPT writes sonnets. It should centre on whether the supply chains producing the chips that produce ChatGPT are resilient to a conflict scenario in the Taiwan Strait.
A concrete figure closes it. According to the Boston Consulting Group and Semiconductor Industry Association report Strengthening the Global Semiconductor Supply Chain (May 2021, updated upward in 2023 and 2024), a one-year interruption in Taiwan's semiconductor manufacturing capacity would produce an estimated global GDP loss of roughly one trillion dollars and a technical setback of several years in the development of generative AI. That figure is the real dimension of the dependence the Spanish press doesn't name and the entire sector assumes every morning.
Definitions
Foundry: a company that manufactures chips designed by third parties, without designing them itself. TSMC is the world's largest foundry. The model is contrasted with vertically integrated companies like Intel, which designs and manufactures.
Node: in semiconductor jargon, the commercial designation of the manufacturing process that roughly measures the minimum size of the structures etched on the chip. The current advanced nodes — 3 nm, 2 nm — no longer correspond literally to length measurements, but to process families.
EUV lithography: an extreme-ultraviolet lithography technique used to etch circuits on silicon wafers at the most advanced nodes. The machines are made only by ASML in the Netherlands.
Silicon shield: a Taiwanese foreign-policy doctrine holding that the global dependence on Taiwanese chip manufacturing is the best protection against a military conflict with China. Operational, not formally codified.
References
Chris Miller, Chip War: The Fight for the World's Most Critical Technology (Scribner, 2022). The central reference on the geopolitics of the semiconductor chain; winner of the Financial Times Business Book of the Year 2022.
Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (Houghton Mifflin Harcourt, 2018). Background on the technological competition between the United States and China.
Graham Allison, Destined for War: Can America and China Escape Thucydides's Trap? (Houghton Mifflin Harcourt, 2017). A geopolitical frame for understanding the strategic conflict between Washington and Beijing.
Boston Consulting Group & Semiconductor Industry Association, Strengthening the Global Semiconductor Supply Chain in an Uncertain Era (report, April 2021, 2023–2024 updates). Quantitative estimates of the global impact of a Taiwanese disruption.
TSMC, Annual Report 2024 (Taiwan Semiconductor Manufacturing Company, March 2025). Official figures on share, capacity and investment.
Financial Times, sector coverage of Taiwan Semiconductor Manufacturing Company (2022–2025). Recurring analysis of the chain and the geopolitical debate.
Office of the U.S. Secretary of Commerce, CHIPS and Science Act Implementation Reports (2022–2025). Documentation of the US semiconductor-manufacturing subsidy programme.
Going deeper
Vaclav Smil, How the World Really Works (Viking, 2022). A material frame on why the physical infrastructure of the digital economy is still more important than the public conversation admits.
Yasheng Huang, The Rise and Fall of the EAST (Yale University Press, 2023). A comparative institutional history that helps in understanding the Taiwanese model.
Dan Wang, Breakneck: China's Quest to Engineer the Future (Norton, 2025). A recent analysis of China's industrial bet and its impact on semiconductor geopolitics.
Centre for Strategic and International Studies, Mapping the Semiconductor Supply Chain (csis.org, periodically updated). An interactive visual map of the global supply chain.
You might also like
- ASML, the Dutch company ChatGPT, Claude and the rest depend on
- Why the United States banned chip sales to China
- Nvidia is worth more than Spain's economy
- DeepSeek and the end of the American monopoly
Comments0
No comments yet.
Leave a comment