The Taiwan Dilemma

On the evening of August 2, 2022, the cabin lights of a U.S. Air Force C-40C were dimmed for the descent into Taipei Songshan, and the seventy-fifth Speaker of the House of Representatives looked out at an island the People’s Republic of China had spent the previous week threatening to encircle if she landed on it. Nancy Pelosi had taken a circuitous route from Kuala Lumpur, looping out over the Philippines and then north between the eastern coast of Taiwan and the open Pacific to avoid Chinese airspace and the Taiwan Strait. The flight had been tracked by an estimated three quarters of a million people on Flightradar24, the largest live audience the website had ever recorded. Below the aircraft, on the densely populated west coast of the island, the lights of Hsinchu, Taichung, and Tainan blurred together in the twilight along the strip of fabs that produced, by then, almost ninety percent of the world’s most advanced logic chips. Pelosi’s traveling party had been instructed not to discuss the agenda. Almost everyone in the cabin understood that the agenda was, at bottom, those fabs.

The next day, before her formal meetings at the Presidential Office, she sat down to a working lunch at the Grand Hyatt with Mark Liu, the chairman of TSMC, and a small delegation of Taiwanese chip executives. The Washington Post would later report the conversation centered on the CHIPS and Science Act, signed a few days afterward, and on TSMC’s planned factories in Phoenix. Pelosi’s office released a sparse readout. Liu’s office released none. That evening, at the Taipei Guest House, President Tsai Ing-wen hosted a banquet in Pelosi’s honor, and Liu attended along with Morris Chang, by then ninety-one. Around the same hour, the Eastern Theater Command of the People’s Liberation Army announced live-fire exercises in six maritime zones surrounding Taiwan beginning the day after Pelosi’s departure, the largest in the strait since the 1995–96 crisis. By August 4, eleven Dongfeng ballistic missiles had been fired toward the waters off northern, southern, and eastern Taiwan; four of them, per Japan’s Ministry of Defense, took trajectories directly over the island, the first time the PLA had ever overflown Taiwan with ballistic missiles. Taiwan’s military tracked the launches and chose not to intercept. Nearly four hundred and fifty Chinese aircraft entered the island’s ADIZ over the course of August, the highest monthly total ever recorded. The live-fire ban was lifted on August 11. The condition the exercises had created did not lift with it.

What had happened, the analysts at the German Marshall Fund and at the Council on Foreign Relations and at the Chinese-language papers in Taipei agreed in the weeks afterward, was a normalization. The median line of the Taiwan Strait, the unofficial center-of-the-channel boundary PLA aircraft had crossed only sporadically before 2020, was now crossed routinely. The exercise pattern improvised for Pelosi became a template. The Joint Sword exercise of April 2023, called in response to a Tsai transit through Los Angeles, used many of the same maritime zones. Joint Sword 2024A in May, called the day after William Lai’s inauguration as Tsai’s successor, deployed a hundred and eleven aircraft and forty-six naval vessels and put PLA Coast Guard cutters into Taiwan’s offshore waters. Joint Sword 2024B in October, called after Lai’s National Day speech, set a single-day record for ADIZ incursions. Each iteration was rehearsed faster and mounted with less notice than the one before. The PLA was, U.S. Indo-Pacific Command’s then commander Admiral John Aquilino testified, treating the strait as if conflict were a question of when rather than whether. Phil Davidson, his predecessor, had told the Senate Armed Services Committee in March 2021 that he believed China could be in position to attempt unification by force as soon as 2027, the centenary of the People’s Liberation Army. That date, however contested by analysts who thought it more notional than operational, had stuck.

The ground Davidson’s date crossed was an industrial park, not a military one. Hsinchu Science Park, on the northwest plain of Taiwan an hour’s high-speed-rail south of Taipei, was a four-square-mile rectangle of low office buildings, parking structures, cafeteria towers, and the windowless gray fabs whose air-handling stacks vented into the constant onshore breeze off the strait. By the early 2020s, more than five hundred companies operated inside the park, directly employing roughly a hundred and sixty thousand people. TSMC had its corporate headquarters there, in a glass building on Li-Hsin Road named for Chang’s mentor K.T. Li, the minister who had made the company possible in 1985. It also had Fab 12, the original 300-millimeter facility where every leading-node TSMC process had first been proven. A hundred kilometers down the expressway in central Taiwan stood Fab 15 outside Taichung, a 28-nanometer giga-fab the size of a small airport. Two hundred kilometers further south, on the alluvial flats outside Tainan that had once been rice paddies, stood Fab 18, an eight-phase complex whose floor area exceeded nine hundred and fifty thousand square meters and whose phases by 2024 were running at 5-nanometer and 3-nanometer nodes. Fab 18 had begun 3-nanometer volume production at the end of December 2022, two days before the calendar year closed; the ceremony had been small, the consequence enormous. Every advanced TSMC node under volume manufacture in 2024, from N5 to N3, ran on Taiwanese soil, in three locations on a single island, within easy range of any cruise missile launched from the Chinese mainland eighty miles to the west.

The number that traveled was the ninety percent. By the early 2020s every American think-tank report, every Brookings memo, every Senate Armed Services hearing, every Reuters dispatch from Taipei had converged on a version of the same line: TSMC produced approximately ninety percent of the world’s leading-edge logic, defined for these purposes as 7-nanometer and below. The figure was useful shorthand for an underlying reality that did not always survive precise auditing. Samsung’s foundry division still ran a small share of leading-edge wafers, mostly captive to its parent. Intel was at varying stages of bringing 18A online in Arizona and Ireland. SMIC in Shanghai had pushed to a 7-nanometer-class node by 2023, although yields and EUV access remained constrained. But the practical truth the shorthand pointed at was real. If a chip’s design was in the most demanding tier of the industry, the chip sold in volume was, with overwhelming probability, made by TSMC in Hsinchu or Tainan. The integrated circuits at the core of the iPhone, the Nvidia GPUs that trained the largest language models, the AMD CPUs that powered the cloud, the Qualcomm modems that linked five billion cellular devices, and the embedded ASICs at the heart of every American precision-guided weapon system fielded in the post-Iraq era passed, at some point in their fabrication, through a TSMC clean room. The clean room was on an island one of the world’s two largest militaries had identified, in the words of its current head of state, as the unfinished business of Chinese national rejuvenation.

That phrase, national rejuvenation, had a date attached. Xi Jinping had since 2012 framed his presidency around the project he called the great rejuvenation of the Chinese nation, pegged to 2049 and the hundredth anniversary of the People’s Republic. He had repeatedly described unification with Taiwan as the essence of that rejuvenation, and in the 2022 work report to the Twentieth Party Congress he had pointedly declined to renounce the use of force. Chinese scholars debated whether 2049 should be read as a deadline within a non-deadline, a pliable target meant to motivate without committing. Xi himself, the consensus of the better Sinologists ran, had constructed the timeline so it could be slipped if circumstances required. He had also made the project personal in a way none of his predecessors quite had. Hu Jintao had spoken of unification as a long-term goal. Xi had made it a test of his own historical seriousness. The window in which Xi would still be in power, even on the most generous Politburo arithmetic, did not extend much past the 2030s.

Hal Brands and Michael Beckley, scholars at Johns Hopkins and Tufts, had captured the strategic implications of that arithmetic in a 2022 book called Danger Zone. China, they argued, was a peaking power: its working-age population had crested, its productivity gains had compressed, its property bubble had begun unwinding. A peaking power, on their reading of the historical literature, was more dangerous than a still-rising one because the leadership knew it could only get weaker. The window in which Beijing could plausibly attempt to take Taiwan by force was, on the Brands-Beckley account, opening now and closing in the 2030s. Taiwanese scholars at Academia Sinica observed that nationalism would drive a decision well before any technocratic optimization of relative military power did. Bonnie Glaser at the German Marshall Fund, who had been writing about cross-strait relations since the 1980s, cautioned that the timeline depended on Beijing’s reading of American resolve as much as on any internal Chinese economic curve. What none of them disputed was that the question of whether China would move on Taiwan, and when, had passed in the previous decade from a problem analysts published papers about to a problem the Pentagon now wargamed continuously.

In a secured room at the Center for Strategic and International Studies on K Street in Washington, over the course of 2022, a team led by Mark Cancian, a retired Marine colonel, ran the same wargame twenty-four times. The scenario was a Chinese amphibious invasion of Taiwan in 2026. The map was a hex-and-counter battlespace covering the western Pacific from the Bashi Channel to Hokkaido. The forces were modeled on declared and projected orders of battle for the PLA, Japan’s Self-Defense Forces, U.S. Pacific commands, and Taiwan’s military. The team published its findings in The First Battle of the Next War in January 2023. In most of the runs, Cancian wrote, the United States, Taiwan, and Japan defeated the invasion and Taiwan retained its autonomy. The cost of victory was sobering. The U.S. Navy lost between ten and twenty surface combatants, two carriers among them. The U.S. Air Force lost between two hundred and four hundred and fifty aircraft. American casualties ran into the thousands within three weeks. Japanese bases were struck. Taiwan’s economy, the report noted in the paragraph that drew most of the financial press’s attention, was a smoking ruin. The PLA Navy lost ninety percent of its amphibious lift and fifty-two major surface combatants. The fabs of Hsinchu and Tainan, in every run that involved meaningful exchange of fires, ceased to exist as functioning industrial facilities by the second week of fighting.

That last finding was almost incidental in the CSIS report. It was the central finding of every honest reading of the report. The wargame had not been designed to model the chip economy; it had been designed to model military outcomes. The military outcomes implied what happened to the chip economy as a side effect. A war in the strait, fought with anything resembling the order of battle the PLA had assembled by the mid-2020s, would produce, within the time it took for the U.S. Seventh Fleet to surge from Yokosuka, the destruction of TSMC. It did not matter who won the war. The factories sat in the path of the artillery and the missiles. The high-purity gases and resists came in tankers that would not sail. The ASML EUV machines, which by the 2020s required real-time telemetry from the Netherlands to operate and which both ASML and TSMC had quietly engineered to be remotely disabled, would lose their telemetry within hours of any kinetic exchange and could not be restarted thereafter without engineers physically present. A clean room with a shattered HEPA filter or a contaminated water loop or a missing supply of fluorinated etch chemistry was, in industrial terms, scrap. Five hundred and forty-four engineers walking off Fab 18 on a Tuesday morning would, by the following Tuesday, have an unrecoverable line. Within the imaginative reach of a single Eastern Theater Command exercise, the chip industry the world had spent forty years building could be wiped out.

This was the conclusion Mark Liu had been saying out loud, in his quiet way, since 2021. In a CNN interview on the eve of the Pelosi visit, he told Fareed Zakaria that nobody could control TSMC by force. The fab was not a building. It was a real-time entanglement of European lithography vendors, Japanese chemicals, American software, and Taiwanese process engineering, all of which depended on uninterrupted contact with the rest of the world to keep operating. Sever the contact and the fab stopped. Bomb the fab and the fab stopped harder. There were, he was implying without quite saying, no winners in any scenario. Anything that ended with the PLA at the Hsinchu gate ended with no chips. Anything that ended with the U.S. Seventh Fleet in the strait ended, by the same logic, with no chips. The only outcomes Liu’s framing did not eliminate were the ones in which no one did anything at all.

Morris Chang gave the most pointed version of the same argument on April 14, 2022, in a conversation hosted jointly by Brookings and CSIS. The interviewer was the historian Robert Kagan. Chang had agreed to the interview, his first major American appearance in years, partly to push back against what he saw as the misreading of TSMC’s Phoenix project as a real solution to the concentration risk. Kagan asked him about the Pat Gelsinger line, then circulating in Washington, that Taiwan was not safe and the United States therefore needed its own leading-edge fabs. Chang gave an answer the Taipei Times paraphrased as “exercise in futility.” Fabbing in Arizona cost fifty percent more than in Taiwan and would remain that way because the supporting ecosystem the U.S. had let migrate over forty years would take a generation to rebuild. The talent did not exist on the necessary scale. The cost premium was something he believed the U.S. would eventually be unwilling to pay. And, as a kind of grace note his American audience took longer to absorb than his Taiwanese one did, if there was no war, the entire effort to onshore advanced semiconductor manufacturing in the United States was a wasteful and expensive exercise in futility, and if there was a war, everyone was going to have a great deal more to worry about than chips. The thing the silicon shield was supposed to deter was also the thing American policy was now hedging against, and the hedge was incoherent on its own terms because the moment the hedge mattered, nothing the hedge protected would be left.

The hedge, by the middle of the 2020s, was nevertheless real, expensive, and globally distributed. TSMC’s Arizona Fab 21, announced in May 2020 with an initial commitment of twelve billion dollars, had broken ground in north Phoenix in 2021, weathered cost overruns and labor disputes, taken delivery of EUV machines in 2022 and 2023, and entered high-volume N4 production at yields the company described as comparable with Taiwan in early 2025. The total investment had grown to $65 billion across three planned fabs by November 2024, when the Department of Commerce finalized $6.6 billion in CHIPS Act direct funding and up to $5 billion in additional loans. It was the largest greenfield foreign direct investment in U.S. history, and it bought Phoenix, in steady state, around three percent of TSMC’s leading-edge capacity. In Kumamoto, on Kyushu, TSMC had partnered with Sony, Denso, and Toyota in a joint venture called Japan Advanced Semiconductor Manufacturing, opening its first fab on February 24, 2024, with a second committed for 2027. Kumamoto would run 28-, 22-, 16-, and 12-nanometer nodes for the automotive and image-sensor markets Sony’s CMOS lines and Denso’s automotive controllers required. In Dresden, on August 20, 2024, TSMC, Bosch, Infineon, and NXP had broken ground on the European Semiconductor Manufacturing Company, with Ursula von der Leyen and Olaf Scholz on the dais and a five-billion-euro state-aid package signed off by Brussels. ESMC would target 2027 for first production at 22- to 28-nanometer nodes serving European automakers. Each project was strategically necessary. None was leading-edge. Every wafer of N3 and below the world bought between 2022 and 2030 would still be made on Taiwan.

This was the fact that turned the silicon shield from a metaphor into an unresolved question. The shield thesis, in its most articulate version, ran as follows: Taiwan’s chip dominance made any war the United States could not afford to lose, because losing meant the loss of a strategic resource without which contemporary military and economic life was impossible; therefore, the United States would defend Taiwan; therefore, China would not attack. The vulnerability thesis, articulated by writers like Chien-Huei Wu at Academia Sinica and developed in policy form by Lev Nachman at Brookings, ran the opposite direction: Taiwan’s chip dominance made the island a higher-value target, narrowed the strategic options of all sides, and crowded out the diplomatic flexibility cross-strait management had historically required. The two theses agreed on the facts. They differed on whether concentration produced deterrence or temptation. There was no empirical experiment that could resolve the difference short of the war neither thesis wanted.

The American policy response, after 2022, hedged both theses simultaneously and somewhat awkwardly. Phoenix and Kumamoto and Dresden were the diversification leg. The CHIPS Act subsidies to Intel, Micron, Samsung, and Texas Instruments were the domestic-capacity leg. The export controls on advanced semiconductor equipment to China, expanded in October 2022 and again in October 2023, were the deny-the-adversary leg. The arms sales to Taipei, the Taiwan Enhanced Resilience Act of December 2022, the rotation of U.S. Special Forces trainers to Taiwanese ground units, the Strait transits by U.S. and allied warships, and the Pentagon contingency reviews were the deterrence leg. The persistent ambiguity over whether the United States would actually intervene in a Taiwan contingency, an ambiguity successive presidents both relied on and undermined, was the hedge against the hedge. Inside this lattice, the Phoenix fab was small in ordinary industrial terms and large in symbolic ones. It was insurance against an event the policy was simultaneously trying to prevent. It would not, by any honest metric, replace TSMC’s Taiwanese capacity. It would mean that, in the days after a war, the world would still possess the institutional knowledge of leading-edge logic manufacturing inside a U.S. jurisdiction, and that knowledge could in principle be scaled. The hedge was not a substitute. It was a residue.

In Hsinchu, on a weekday morning in the middle of the decade, the engineers walked from the parking garages on Li-Hsin Road into the lobbies of TSMC’s headquarters and Fab 12 in unbroken streams. The cleanroom-suit lockers filled and emptied in twelve-hour shift cycles. The fabs ran twenty-four hours a day, seven days a week, every week of the year except for short scheduled maintenance windows on the Lunar New Year. The R&D engineers known internally as the Night Hawks, the after-midnight teams Chang had built in the 1990s and Liu had institutionalized, kept their pagers next to their beds. The wafers came off the line at the rates the customers required: an N3 wafer for an Apple A18, an N4 for an Nvidia Hopper successor, an N5 for an AMD MI300 or a Qualcomm 8-series, an N7 for the embedded controllers of the next generation of automotive radar. Each step in the production line was performed by a tool built somewhere else, calibrated against samples shipped overnight from Tokyo or Eindhoven, supplied by chemicals from Osaka or Antwerp, instrumented by software written in California. The fab was the visible portion of an iceberg whose submerged mass was the entire integrated technology base of the developed world. Taiwan, in this picture, was less a country than a node. Or, from the other side, less a node than a country.

The question Morris Chang had unintentionally posed in 1987, when he carved a 2-micron foundry out of an ITRI building under the protection of K.T. Li, had become the strategic question of the 2020s and beyond. The world had bet that economic interdependence would produce political stability. It had received, as the bet’s first dividend, four decades of unprecedented growth. It had received, as the bet’s second dividend, a single point of failure of a kind no industrial system had ever before constructed. The single point sat eighty miles off the coast of a country that claimed it. Whether the dependency deterred Beijing or attracted it, whether the diversification effort could be scaled fast enough to relieve the pressure, whether Xi or his successor would treat 2049 as a target or a metaphor, whether the next Joint Sword exercise would be an exercise or something else, were questions to which no one had a confident answer. The CSIS wargame had been run twenty-four times. The runs had not been run for real. The lit fabs of Hsinchu and Tainan kept running, and the ASML telemetry kept flowing, and the wafers kept coming off the line, on a frequency that had not yet been interrupted. In Taipei, the politicians worked the diplomatic angles their predecessors had worked since 1979. In Phoenix, the EUV machines were craned into Fab 21 and the engineers, many on rotation from Hsinchu, set about the long work of reproducing in the Sonoran Desert the ecosystem the original engineers had built in Hsinchu, knowing the work would take a generation, and not knowing whether they had a generation. The story that began in the postwar laboratories of New Jersey, that ran through Tokyo and Seoul and Hsinchu and Beijing, that turned a hand-soldered point-contact transistor into the substrate on which the next century would be built, had arrived at a single island and a single open question. No one alive yet knew the answer.