EXPLAINER

Why the world's most advanced chips are made in Taiwan

Taiwan holds 92% of the world's manufacturing capacity for the most advanced logic chips (below 10 nm, 2019 data) — most of it one company, TSMC. Here's how that happened, and what is being built to change it.

What 'most advanced' actually means

A chip is a slab of silicon covered in transistors — tiny switches that turn on and off to do arithmetic. The "node" of a chip, quoted in nanometres (3nm, 5nm, 7nm), is the generation of the manufacturing process, not a literal measurement of any single transistor. Smaller generations pack transistors closer together, so a chip can do more work in the same space while using less power per calculation.

The scale is hard to picture without an example. NVIDIA's Blackwell GPU, the chip at the centre of today's AI data centres, packs 208 billion transistors and is made on a custom TSMC process called 4NP. Two dies, each at the largest size lithography can print in one exposure, are joined into a single GPU. That chip — and the AI accelerators, phone processors and server CPUs like it — can only be built on the most advanced nodes. And those nodes are almost all in Taiwan.

What those chips are used for once they are built — training a model or running it — is a separate question; see training vs inference.

Those chips end up in racks that draw as much power as a small neighbourhood; our explainer on the 120 kW AI rack follows them there.

The figure to remember is a capacity figure, and it is worth stating precisely. A 2021 report by the Semiconductor Industry Association and Boston Consulting Group, using 2019 data, found that 100% of the world's manufacturing capacity for the most advanced logic chips — those built on processes below 10 nanometres — sat in just two places: Taiwan with 92% and South Korea with 8%. The popular version of that finding, "90% of the world's most advanced chips come from Taiwan," is a loose restatement. What was measured was capacity: the ability to make them, not the number made in any given year. But the conclusion is the same either way. If you want a leading-edge chip made, it is almost certainly made in Taiwan.

One company that designs nothing

The concentration has a name: TSMC, the Taiwan Semiconductor Manufacturing Company. It was founded in Hsinchu, Taiwan, on February 21, 1987, by Morris Chang, and it works differently from the chip companies most people know. Intel and Samsung design chips and make them. TSMC designs no chips of its own to sell. It manufactures chips designed by other companies — Apple, NVIDIA, AMD, Qualcomm — and it promises never to compete with its customers.

That promise was the invention. Before TSMC, a company that designed chips had to build its own factory, or hand its designs to a rival that also sold chips. Chang's idea, the "pure-play foundry," removed that conflict: a factory whose business is manufacturing other companies' designs. (TSMC does supply design IP — memory blocks, analog IP, standard-cell libraries — that customers build into their designs for its processes; it does not sell chips of its own.) It let hundreds of small design firms exist without building a fab, and it let TSMC focus every dollar on one thing — making chips better and cheaper than anyone else.

The result, nearly forty years later, is scale no rival can match. In 2025 TSMC held about 70% of the entire global foundry market — 69.9%, up from 64.4% the year before, according to TrendForce — with Samsung a distant second at 7.2%. TSMC's 2025 revenue was $122.54 billion. It is not just the biggest foundry. At the leading edge, it is close to the only one.

The machine only one company sells

There is a second choke point behind the first. The most advanced chips are printed with extreme ultraviolet (EUV) lithography: light with a wavelength of 13.5 nanometres, generated by firing a laser at droplets of molten tin, is bounced off mirrors to etch patterns too fine for any other method. Only one company in the world makes EUV machines — ASML, of the Netherlands.

ASML shipped 48 EUV systems in 2025 and planned to ship 60 in 2026. A standard EUV machine costs around $200 million; the newest High-NA version, for 2-nanometre processes and beyond, costs around $400 million. Each machine contains more than 100,000 parts from some 800 suppliers. There is no second vendor, and export rules keep EUV machines out of China entirely. So the chain has two single points of failure, not one: Taiwan makes the chips, and one Dutch company makes the machines that make the chips.

That is also why "just build a fab somewhere else" is harder than it sounds. A leading-edge fab costs tens of billions of dollars, takes years to build, and needs the machines, the chemicals, the masks, the engineers and the customers — all of which are already in Taiwan.

Why Taiwan, and why it is hard to copy

Taiwan's lead is not one factory. It is an ecosystem. The fabs sit in science parks in Hsinchu, Tainan and Taichung, surrounded by the suppliers that feed them: chemicals, gases, photomasks, precision parts, and the packaging firms that finish the chips. Engineers move between companies in the same city. A problem found at one fab in the morning can be fixed with a part from a supplier down the road by the afternoon.

Three things made that ecosystem, and none of them travels easily. The first is time: TSMC has been refining its processes since 1987, and each generation builds on the last — the yields, the recipes, the institutional knowledge of what goes wrong. The second is talent: Taiwan trains and retains a deep bench of process engineers, and TSMC's own workforce numbers in the tens of thousands. The third is government support: Taiwan's government backed TSMC from its founding and has supported the science parks, the power and water infrastructure, and the research institutes behind it ever since.

Other countries have the money. What they do not have is the forty years. That is the honest answer to why the concentration persists — and why the fabs now being built elsewhere are a start, not a solution. The same clustering logic applies to the buildings that use the chips — see why data centers gather in the same places.

The fabs being built outside Taiwan

Governments are spending heavily to spread chipmaking out. Here is what is actually being built, and what stage each project is at. Announced is not the same as producing.

Arizona, United States — TSMC Fab 21, Phoenix. The most advanced fab project outside Taiwan. The first fab is in volume production on 4-nanometre chips, starting in late 2024 — the first leading-edge chips ever made in volume in the United States. A second fab, for 3-nanometre chips, has finished construction and is targeted for N3 volume production in the second half of 2027. A third fab, for 2-nanometre and more advanced processes, is under construction for the end of the decade. TSMC's total planned investment in Arizona has grown to $265 billion. The U.S. Commerce Department finalized up to $6.6 billion in direct CHIPS Act funding for the project in November 2024, plus up to $5 billion in proposed loans. One limit is worth stating plainly: the first NVIDIA Blackwell wafer made in Arizona, announced in October 2025, still had to travel back to Taiwan for advanced packaging — the step that joins the chip to its memory — because the United States does not yet have that packaging capacity. TSMC's first advanced-packaging fab in Arizona is in early construction — so for now, packaging still happens in Taiwan. Packaging itself is becoming a bottleneck — see our brief on silicon interposers, the layers that connect chips to memory.

Kumamoto, Japan — JASM. TSMC's joint venture with Sony, Denso and Toyota. The first fab began volume production in December 2024, making chips on 12- to 28-nanometre processes for cars and industrial uses — mature nodes, not leading-edge ones. A second fab began construction in October 2025 and was upgraded to a 3-nanometre process, with volume production scheduled for 2028. A magnitude-7.1 earthquake struck Kumamoto in July 2026; the first fab was undamaged, operations returned to normal after inspections, and the second fab remains scheduled for 2028.

Dresden, Germany — ESMC. A joint venture led by TSMC (70%) with Bosch, Infineon and NXP (10% each). The investment is over €10 billion, including €5 billion in German state aid approved by the European Commission in August 2024. The building topped out in September 2026. It will make chips on 28/22-nanometre and 16/12-nanometre processes — automotive and industrial chips, not leading-edge logic — at about 40,000 wafers a month, with production targeted for the end of 2027.

The pattern is the same in all three places: the fabs being built outside Taiwan are either a generation or two behind the leading edge, or years away from producing it. Arizona is the exception, and even there TSMC's most advanced processes stay in Taiwan for now. (Intel's competing 18A process ships from Oregon — see the FAQ below — but at a much smaller scale.)

What happens if the supply is disrupted

The risk is not theoretical, and it is not only about war. An earthquake, a drought (fabs use enormous amounts of water), a power outage, or an export restriction could interrupt supply from an island that makes nearly all of the world's most advanced chips. The SIA/BCG report put it bluntly: an extreme scenario of a complete disruption of Taiwanese foundries for one year could bring the global electronics supply chain to a halt.

Governments are responding with money. The U.S. CHIPS and Science Act, signed in August 2022, appropriated $52.7 billion for semiconductors, including $39 billion in manufacturing incentives run by the Commerce Department. The European Union has its own Chips Act, and Japan has committed trillions of yen to its fabs. But money does not shorten the timeline. A fab takes years to build, years more to reach the yields that make it economic, and the ecosystem around it longer still. The honest summary is that diversification has started, it is real, and it will not remove the concentration this decade.

Questions readers ask

Does TSMC design any chips?

Not chips of its own to sell. TSMC is a pure-play foundry: it manufactures chips designed by other companies and sells none under its own name. It does supply design IP — memory blocks, analog IP, standard-cell libraries — that customers build into their designs for its processes. That is the point of the model — its customers trust it with their designs because it does not compete with them.

What does '3nm' actually measure?

Not the size of a transistor. Node names are generation labels: each new generation packs transistors more densely than the last. A "3nm" chip and a "5nm" chip differ in density, power and cost — the number itself is a marketing label for the generation, and different manufacturers' numbers are not directly comparable.

Why can't the United States just build the same fabs?

It is building them — in Arizona — but a fab is not a building. It is the machines, the chemicals, the engineers, the suppliers and the customers, and the yields that only come with years of running the process. Arizona's first fab is producing 4-nanometre chips. The most advanced processes, and the advanced packaging that finishes AI chips, are still in Taiwan.

Is Samsung a real alternative to TSMC?

Samsung is the only other company making leading-edge chips at scale, with about 7% of the foundry market against TSMC's roughly 70%. It is a genuine second source, but a much smaller one, and its most advanced processes have lagged TSMC's in volume. Intel is ramping its own leading-edge process, 18A — in July 2026 ASML confirmed Intel had become the first company to ship high-volume logic patterned with High-NA EUV, with selected 18A layers (Panther Lake laptop chips) dual-qualified in Oregon — but it is not yet at comparable scale.

What does this have to do with data centres?

Everything in a data centre runs on chips, and the most demanding chips — the GPUs that train AI models — are made on TSMC's most advanced nodes. A data centre is, among other things, a building full of Taiwanese silicon. Our explainers on what a GPU does and why more AI chips aren't always faster cover what those chips do once they arrive.

Sources and further reading

Glossary — plain definitions for words used here.