The Round Trip

A chip that ends up in an American data centre can be manufactured in Arizona and still not be finished there.

TSMC's Fab 21, north of Phoenix, has been in volume production since the end of 2024, printing wafers on a four-nanometre process. Real output, not a showcase line. Then the wafers are loaded onto aircraft and flown to Taiwan, where the individual dies are cut apart, stacked with memory, mounted on a substrate and tested, and only after all that do the finished parts fly back out to the customer. TSMC did not have the advanced packaging capacity in Arizona that buyers like Nvidia needed, so flying them home was the only option available.

This is being fixed. In June 2026 TSMC signed a ten-year agreement with Amkor to have that work done in Arizona instead. Amkor is building a $7bn packaging and test campus in Peoria, half an hour up the road from the fab, and expects production to start in early 2028. Until then, the round trip stands.

I keep coming back to that flight, because it disposes of half the question people usually ask. The question runs roughly like this: how did the United States end up depending on one island a hundred and thirty kilometres off the coast of China for the components its economy and its military are built on, and why can't it just make them at home? The second half answers itself on the tarmac in Phoenix. America did make them at home. The factory exists, it runs, and the dependency did not go away, because the factory was never where the dependency lived.

The first half has a stranger answer. I went looking for the decision, the moment when the United States handed chip manufacturing to Taiwan, and there isn't one. No policy, no summit, no chief executive who sold the family silver. What there is instead is a long sequence of choices, nearly all of them correct at the time they were made, taken by people who mostly did not know each other and shared no plan, which compounded over thirty-five years into an arrangement nobody would have designed.

The company that promised never to compete

Morris Chang spent a quarter of a century at Texas Instruments and rose to run its semiconductor business. He did not get the top job. He left, spent a short unhappy period at another American firm, and in the mid-1980s accepted an invitation from Taiwan's government to run its Industrial Technology Research Institute with what he later described as a blank cheque to build an industry (Britannica).

Taiwan had capital, engineers and no obvious way in. Every serious chip company at the time designed and manufactured its own products. Intel, TI, Motorola, NEC, Toshiba. To compete you had to beat them at both things at once, from a standing start, on an island with no track record in either.

TSMC was incorporated in February 1987 with the Taiwanese state holding 48.3 per cent through its National Development Fund and Philips taking 27.6 per cent, along with a technology transfer and the company's first chief executive. It is worth pausing on that, since TSMC now gets described as a monument to free enterprise. It began as a government project with a European partner and an American executive at its head.

Chang's idea was not about manufacturing. It was commercial, and it was close to perverse. His company would own the factories and design absolutely nothing. Not now, not later, not as a side business. That meant giving away the half of the industry where the margins were, and everyone knew it.

What it bought was trust. Imagine you are three engineers with one good idea for a graphics chip. You cannot take that design to Intel to be manufactured, because Intel sells chips, and next year it might sell yours. A factory that competes with nobody can be trusted by everybody. TSMC's real product was never silicon. It was a credible promise not to become your rival, and that promise is the thing a company like Intel structurally cannot make, which is a large part of why Intel's own foundry business still struggles to sign the customers it needs.

In 1997 a nearly bankrupt company of about fifty people wrote to Chang asking him to manufacture its chips. It had already approached TSMC's San Jose office and been ignored, so the letter went to the head office in Taiwan and eventually landed on Chang's desk, where by his own account it made him both curious and annoyed. He phoned the number on the letterhead. The company was Nvidia, and it has been among TSMC's largest customers ever since.

The half of the business nobody wanted to keep

Once you could rent a factory, the cost of starting a chip company collapsed. You needed a design team and a bank account, not ten billion dollars and a decade. Nvidia was founded in 1993 and never owned a fab. Neither did Qualcomm, Broadcom, Marvell or, later, Apple's silicon group, which designs the most valuable processors in consumer electronics and has never built a single one.

Then the direction of travel reversed for the companies that already had factories. AMD spun its manufacturing out into a separate business, GlobalFoundries, in 2009 and became a design house. It was not a defeat. Fabs eat capital, run at brutal utilisation risk and depress return on assets. Shedding them made the American design companies dramatically more profitable, and that profitability is not an accident of the last few years. It is the direct consequence of the trade.

So the framing of America "losing" its chip industry is wrong in a way that matters. American companies still account for roughly half of global semiconductor revenue and spend around seventy billion dollars a year on research, a share of turnover second only to pharmaceuticals (SIA). The United States did not lose the industry. It sold the capital-intensive, low-margin, physically demanding half of it, kept the half that prints money, and got richer doing so for three decades.

Then the ante went up

While that was happening, the cost of staying at the front of manufacturing was climbing in a way that eventually settled the question by arithmetic.

A leading-edge fab now runs to something like twenty billion dollars. A single low-numerical-aperture EUV lithography machine costs somewhere between $180m and $220m, and the next-generation High-NA version runs about $380m each. Developing a new process node costs billions before a single saleable wafer comes off the line, and the only way to recover that is volume that almost nobody has.

You can watch the field thin out. In August 2018 GlobalFoundries abandoned its 7nm programme, having worked out that it would need to spend two to four billion dollars just to build enough capacity to have a shot at a return on the node. UMC made the same call. IBM, Motorola's spin-off, Texas Instruments and most of the Japanese giants had already stepped off at earlier nodes. What was once a crowded field is now three companies: TSMC, Samsung and Intel, and the gap between the first and the other two is not small.

The result is that TSMC makes about 70 per cent of all chips manufactured under contract, and at the leading edge, the processors inside phones and AI accelerators, its share is somewhere north of 90 per cent.

Two things get lost when those numbers are quoted, and both matter. The first is that Taiwan's share of total global wafer capacity is nothing like 90 per cent. It is around 18 per cent, roughly level with South Korea. The largest single holder of raw capacity today is China, at about 24 per cent, almost all of it at older, cheaper nodes. Concentration in Taiwan is a leading-edge phenomenon, not a general one.

The second is that the shortage that frightened everybody was not a leading-edge problem at all. When carmakers could not build cars in 2021, at an estimated cost of $210bn in lost revenue, they were short of unglamorous parts made on twenty-year-old processes, the kind that cost fifty cents. The industry got its scare from one end of the market and drew its strategic conclusions about the other.

Intel had the one job

None of this had to end with a single foreign supplier at the frontier, because for most of that period there was an American company that was better at manufacturing than TSMC was. Intel's 14nm process, which arrived in 2014, was excellent. It was supposed to be followed by 10nm around 2016.

10nm did not reach real volume until 2019.

The technical account is that Intel bet against EUV lithography and tried to reach its density targets with conventional light and multiple patterning instead. In the worst cases it was exposing a wafer six times to draw a single feature, which lengthens the process, raises defect density and destroys the economics. It also switched to cobalt in the interconnect layers. TSMC, meanwhile, committed to EUV earlier and moved in smaller steps. Intel spent four or five years stuck, and it has not led since.

Underneath the technical story is a commercial one that I think matters more. In 2013, on his way out, Paul Otellini told The Atlantic why Intel had not made the processor for the original iPhone. Apple wanted a chip at a fixed price and not a cent more, and that price sat below Intel's own forecast of what the chip would cost to build. He could not see it working. "The forecasted cost was wrong and the volume was 100x what anyone thought," he said, and the reporter noted that this was the only moment in the interview when regret came into his voice.

Missing the iPhone was not just a missed contract. Mobile was the volume that paid for the next node, and it went to TSMC via Apple and Qualcomm. This is where the integrated model quietly turns into a liability: Intel's factories could only ever be filled by Intel's products, so when Intel's product strategy misfired, its fabs starved at exactly the moment they needed to be fed. TSMC's fabs are filled by the whole industry at once, which means TSMC does not have to be right about any particular market. It only has to be there when somebody else is.

The chain is longer than the island

Here is where the popular framing needs its second correction. Suppose Taiwan vanished from the map tomorrow. Would America then be free to make its own chips? No, and not because of anything Taiwanese.

The machines that print the smallest features are made by one company, ASML, in the Netherlands. There is no second supplier and no near-substitute. The optics inside them come from Carl Zeiss SMT in Germany, which has no other customer for that work, and the drive lasers come from TRUMPF, also German. Japanese firms make more than 90 per cent of the photoresist those machines expose. The blank masks come from two Japanese companies. A single Japanese food company, Ajinomoto, makes the insulating film used in almost every high-performance chip package, a by-product of its work on amino acids.

And the packaging itself, the step that sends Arizona's wafers back across the Pacific, is concentrated in Taiwan at around 45 per cent of global capacity, then South Korea and China. Nvidia has reportedly booked more than half of TSMC's advanced packaging output for 2026 and 2027, which is why the constraint on AI hardware right now is not the printing of the chip but the assembling of it.

Now turn it around. The tools that deposit, etch and inspect the layers come substantially from Applied Materials, Lam Research and KLA, all American. The software used to design essentially every advanced chip on earth comes from two companies headquartered in California, Synopsys and Cadence, which between them hold about 61 per cent of that market. In May 2025 the United States required licences for the sale of that software to China and the affected companies immediately suspended their financial guidance. The restriction was lifted weeks later as part of a trade truce. That is what a chokepoint looks like when it is yours.

And those links do get pulled. Since October 2022 Washington has been progressively barring China from buying advanced chips, the equipment to make them and, at intervals, the software to design them. The results are real and partial at the same time. SMIC did reach a 7nm process without an EUV machine, by exposing wafers repeatedly with older tools, and Huawei shipped phones built on it, which was widely read as proof the controls had failed. Then it stopped there, because the next step down is where doing it the hard way stops paying for itself. The same object is a policy instrument when you are holding it and a vulnerability when somebody else is.

So "America depends on Taiwan" is true and incomplete. The accurate version is duller and more interesting: no country can make an advanced chip alone, everybody holds one irreplaceable link, and the reason this feels like a Taiwan problem rather than a Dutch problem or a Japanese problem is that Taiwan is the only link sitting next to a state that claims it.

What the money actually bought

The American response has been the largest industrial policy programme in two generations, and it is far enough along to score honestly.

The CHIPS and Science Act of August 2022 provided $52.7bn, of which the Commerce Department had committed up to $33.7bn in direct awards plus $5.5bn in loans across nineteen companies by early 2025. The following administration renegotiated. In August 2025 Intel's unpaid grants and its Secure Enclave money were converted into equity: $8.9bn for 9.9 per cent of the company, making the federal government Intel's largest shareholder.

On the evidence so far, the pessimists were half right.

They were wrong about the hard part. In April 2022 Morris Chang told a Brookings audience that trying to manufacture chips in America would be "a very expensive exercise in futility", citing a 50 per cent cost premium at TSMC's older Oregon fab and a shortage of American manufacturing talent he traced back to the 1970s. Arizona has not borne that out. Fab 21 reached volume production on schedule, TSMC's Arizona subsidiary swung to a $150.1m profit in the first half of 2025 against a $143.4m loss the year before, and the company's own US president has claimed the Phoenix site runs 4 per cent better yields than comparable fabs in Taiwan. That last figure is TSMC talking about TSMC, so treat it accordingly, but the direction is not in dispute. Advanced chipmaking outside Taiwan is demonstrably possible.

They were right about the money and the clock. Arizona's first fab reportedly came in 30 to 50 per cent over the cost of an equivalent Taiwanese building. TSMC has raised its Arizona commitment to $165bn across five fabs, packaging plants and a research centre, and says that when it is all finished roughly 30 per cent of its most advanced capacity will sit in the United States. The second Arizona fab starts equipment installation in the third quarter of 2026 for production in 2027; the third, for the 2nm generation, broke ground in April 2025. Intel's Ohio project, announced in 2022 for a 2025 start, has slipped to 2030 or 2031.

The scoreboard, then. America held 37 per cent of global manufacturing capacity in 1990. It holds about 11 per cent now, and the forecast for 2032 is 14 per cent. Hundreds of billions of dollars have arrested a thirty-year slide and bought back three percentage points over a decade. Whether that is a triumph or a rounding error depends entirely on what you thought the money was for.

Then came the blunter instruments. On 15 January 2026 the United States imposed a 25 per cent tariff under Section 232 on a narrow class of advanced computing chips, with a long list of exemptions for anything destined for American data centres, research or domestic manufacturing. The same day it announced a deal with Taipei: Taiwanese companies to invest at least $250bn in American production, the Taiwanese government to guarantee another $250bn in credit, and Taiwan's tariff rate down from 20 to 15 per cent. Whatever else that is, it is not the market allocating capital.

The part Taiwan is not sending

There is one more reason Arizona will not close this, and it is written down.

In December 2025 a deputy minister at Taiwan's National Science and Technology Council told legislators that the island would apply what he called an N-2 rule: semiconductor process technology may only be deployed overseas if it is at least two generations behind the most advanced process running in Taiwan. Earlier reporting had suggested one generation, so the exact setting has moved, but the principle is now policy rather than preference. Whatever TSMC is doing in Hsinchu, Arizona will be doing something older, permanently, by law.

Taiwan calls this the silicon shield, and the logic is that a country whose factories the world cannot do without is a country the world will protect. It is a real argument and I think it is half true. The other half is the part Taipei argues about domestically: indispensability also makes you worth seizing, and it gives your protector a standing incentive to relocate the very capability that makes you indispensable. Taiwanese debate about the industry being hollowed out is not paranoia. Every fab TSMC builds abroad, at American insistence, moves a little of the shield off the island.

What that means in practice is that Washington and Taipei have overlapping but genuinely different objectives here, and both of them know it. America wants the capability duplicated. Taiwan wants the capability duplicated slowly and never completely. The N-2 rule is what that disagreement looks like when it has been written into law and handed to a regulator.

What it was, all along

The thing I find hardest to shake, after reading all this, is how little blame there is to hand out.

Morris Chang was right that a factory which competes with no one has an asset no integrated manufacturer can copy. American design companies were right that shedding fabs would make them richer, and they were right for thirty years running. GlobalFoundries was right that it could not earn back the cost of 7nm. Otellini's arithmetic on the iPhone was defensible on the numbers he had; he was simply wrong about a market that did not exist yet. Even Intel's 10nm bet was a reasonable engineering judgement about a technology, EUV, that really was late and really was unproven.

Concentration was not the failure mode here. It was the success mode. The best manufacturer attracts the volume, the volume funds the next node, the next node widens the lead, and the loop tightens until one company on one island is making almost everything that matters. That is what a well-functioning market looks like when the product has enormous fixed costs and a steep learning curve. The system optimised beautifully for cost and speed across thirty-five years and never had to price the possibility of the island becoming unavailable, because that possibility does not appear in anyone's quarterly numbers until the quarter in which it does.

Which makes the current moment genuinely unusual, and worth stating plainly rather than dressing up as strategy. Several governments and some of the wealthiest companies in history are now spending enormous sums to make chips more expensively than they need to. That is the actual content of the CHIPS Act, of Arizona, of Kumamoto in Japan, of the European fabs. It is a bet that some duplication is worth more than the savings it destroys, made by people who spent their whole careers being rewarded for believing the opposite.

In early 2028, if Amkor holds its schedule, the flight from Phoenix stops. The dies will be cut and stacked and tested in Peoria, and the chip will leave Arizona finished. It will still have been printed by a machine from Veldhoven, using light from a German laser focused through German mirrors, coated in Japanese resist, on a wafer that was very likely Japanese as well, laid out in software from two companies in California, at a process node that Taiwanese law requires to be two generations behind whatever Hsinchu is running that year.

You can shorten a supply chain. Nobody has yet ended one. What all that money is buying is something more modest than independence, which is a second place the world can go if the first one stops, and after the last thirty-five years that turns out to be worth quite a lot.