Global Markets: Taiwan Semiconductors

Unit 3, Professor Carrie Elkins, Global Markets, Taiwan, Semiconductors, and a “New Cold War”?

The previous piece asked why compliance that's complete on paper can't stop window dressing. This one pulls the camera out to an island, a strait and a chip. The shield in this story is the most advanced wafer in the world. The question is whether the shield becomes a bullseye first.

Here goes — enjoy:

Case Summary

In autumn 2021, Tsai Ing-wen addressed the world in Foreign Affairs. Sorties by military aircraft into the air defence identification zone were at a record high. The defence minister judged that by 2025 the other side of the strait would be able to take military action at the lowest possible cost. She wrote: if Taiwan were to fall, the consequences for regional peace would be catastrophic. She also knew that ideals often dissolve in international relations, so she offered another reason — TSMC, and the semiconductor industry more broadly, are a “silicon shield.” Chairman Mark Liu told 60 Minutes that the whole world needs Taiwan's high technology, and so no one will let war happen in this region. Yet Beijing and Washington were both pouring in money, trying to break their dependence on Taiwan — and on each other. Some began calling this a new Cold War.

Semiconductors were never something the market grew on its own. In 1947, Bell Labs built the transistor. The Pentagon put up money, set standards and built plants. Antitrust action by the Justice Department forced AT&T to open up its patents. Venture capital grew Fairchild and Intel in the Stanford Industrial Park. Moore's Law became the metronome. In the 1970s, the US moved packaging and testing to places including Taiwan; Japan's MITI coordinated R&D and subsidised loans, and its share rose from 28% in 1978 to 46% in 1986, while the US fell from 55% to 40%. Washington did not go on “trusting the market.” It threatened anti-dumping action, slapped tariffs on US$300 million of Japanese electronics, and set up SEMATECH with fourteen companies. The US went on to win in design, not in wafers. In 2019, 38% of global semiconductor value added was still in the US, as was 74% of EDA and core IP. What it lost was the most advanced fab.

Taiwan chose a very narrow slice. Land reform, export processing zones and vocational schools got the people ready first. In 1990, 30% of university students studied engineering. ITRI was founded in 1973 to do applied research, and over twenty-seven years it fed roughly 12,000 engineers into industry. Its electronics research institute bought an already outdated 7-micron process from RCA, achieved yields higher than RCA's own US plant, and then spun off UMC. In 1985, Morris Chang put down the phone. He had spent twenty-five years at Texas Instruments, had watched the cost of building a fab roughly double every five years, and had seen a crop of fabless design companies start renting other people's capacity. The premier had asked him to lead this effort. His judgement was: don't build yet another chip company that does everything itself. Use state capital to build a pure-play foundry that only manufactures for others — no design of its own, no brand of its own. He signed with Philips, brought in technology only nine months behind the frontier, and founded TSMC. Within a year, more than forty design companies sprang up in Hsinchu. In 1992, AMD's CEO was still saying real men have fabs. In 2009, AMD sold its fabs. By 2020, TSMC held half of global foundry capacity. In 2019, 92% of logic wafer capacity below 10 nanometres was in Taiwan.

In 2014, Beijing switched to a bigger industrial policy. After the Snowden affair, the leadership concluded that core technology could not be in other people's hands. Xi Jinping's words in 2016:

Core technology is our greatest lifeline; having core technology controlled by others is our greatest hidden danger.

The National Integrated Circuit Industry Development Guidelines set an investment target of US$150 billion. Made in China 2025 wrote a ten-year self-sufficiency rate of 70%. From 2014 to 2018, 86% of state equity support for semiconductors worldwide came from China. Government ownership of SMIC rose from under 15% to over 45%. The toolkit also included forced joint ventures, US$35 billion a year in outbound M&A, and talent programmes. The scoreboard was mixed. In the first ten months of 2020, about 200 new semiconductor companies were registered every day. SMIC announced 7 nanometres. Tsinghua Unigroup walked into bankruptcy. Chinese-owned firms still produced only 5.9% of the chips China used. Huawei's Mate 60 took domestic components from 29% to 47% — real progress, but still without the extreme ultraviolet lithography machine that only the Netherlands in the whole world can make.

Washington rewrote Beijing from “responsible stakeholder” to strategic competitor. Export controls, the extraterritorial reach of the foreign direct product rule, and investment screening expanded to minority stakes, step by step removed Huawei from TSMC's customer list — it had been the second-largest customer after Apple, contributing more than 15% of revenue, and after it left, Huawei's own revenue fell by roughly a third. In September 2022, National Security Adviser Jake Sullivan said: we used to only need to stay a generation or two ahead; now, for advanced logic and memory, we must maintain as large a lead as possible. Mitt Romney, once a critic of industrial policy, turned to backing the CHIPS and Science Act: US$278 billion authorised, over US$52 billion for semiconductors. The Boston Consulting Group calculation cited in the case is that US$50 billion in subsidies would lift the US share of output only from 12% to 14%. Morris Chang called it an expensive exercise in futility. TSMC still went to Arizona, planning to invest US$40 billion, at a local construction cost roughly four times Taiwan's. He said that if a machine breaks at 1 a.m., in the US it gets fixed the next morning; in Taiwan it's fixed by 2 a.m. He also said: globalisation is almost dead, and free trade is almost dead.

On the island, the silicon shield theory was being used at the same time. In 2020 Taiwan was the world's twenty-second-largest economy. Exports often exceeded half of GDP, 40% of them semiconductors, and China absorbed 44% of exports. It was the economy with the second-highest exposure to global value chains. Compulsory military service had been cut to four months, which some called summer camp. Military spending fell from about 5% of GDP to about 2%. Roughly half of respondents said they were willing to defend the island, and most believed they would never really have to fight and could count on the US. In 2020, US$8.1 billion bought 66 F-16s. Analysts would rather see a porcupine: anti-ship, anti-tank and air defence, for a war of attrition the island could survive. In the winter of 2024, Lai Ching-te won the presidency. After his victory he said that Taiwan's semiconductors are a shared asset of the world, and he hoped everyone would treasure them. The silicon shield narrative lives on. The case ends with a question: is it alive and well, or dying a hard death?

The three questions in front of Carrie Elkins:

1. How have the government policies of the US, Taiwan and China each shaped the semiconductor industry?
2. How should Taiwan navigate US–China tensions? Are semiconductors a protection, or a vulnerability?
3. Does this competition mean a deglobalising world? If so, what are the consequences?

My Thinking

Three countries' industrial policies grew three different capabilities

Semiconductors are not the spontaneous order of a free market. They are capabilities selected by policy.

US policy built an open knowledge system and then outsourced the factories. The Pentagon and antitrust provided the first push; venture capital provided diffusion; the counterattack on Japan preserved design. Today the most valuable US chip companies often have no fabs of their own. Nvidia hands its products to TSMC. CHIPS's US$52 billion is an attempt to buy the factories back. A bill can buy back concrete; it can't buy back the work culture of someone coming to fix the machine at 2 a.m.

Taiwan's policy chose one segment and cut all the way down. Not design and brands — a pure-play foundry. ITRI, Hsinchu, Philips technology only nine months behind, government stakes of up to 49% — all for this one segment. Apple, Qualcomm and Nvidia look the way they do today because someone took on the manufacturing risk. A developmental state doesn't “develop electronics” in the abstract. It picks the slice it can do, and avoids the one it can't.

China's policy is bigger, broader and more urgent. A US$150 billion target, 70% self-sufficiency, 86% of the world's state equity support. Money can pile up 200 new companies a day; it can pile up the Mate 60's 47%. It hasn't yet piled up stable advanced yields. An advanced chip takes about 700 process steps. Only if every single step is right does the final product hold up. Subsidies can be written into documents. The discipline of 700 steps can't.

My own understanding is that industrial policy isn't a switch. It's a choice: at which segment of the chain are you actually prepared to become irreplaceable to the whole world?

The silicon shield is both an amulet and a bullseye

92% is a number that changes the calculus of war. In 2019, 92% of logic wafer capacity below 10 nanometres was in Taiwan, and the remaining 8% in South Korea. Only three companies in the world could make 5 nanometres. A conflict that lands on this island wouldn't be a local war. Phones, the cloud, cars and model training would all stop together. In that direction, the shield is real.

The same set of numbers is also drawing a bullseye. Both sides are spending money so that they no longer need this 92%. Beijing writes it as 70% self-sufficiency. Washington writes it as export licences and a fab in Arizona. With every piece of equipment installed, the sentence “this chip can only be made in Taiwan” becomes a little less true. Huawei once contributed more than 15% of TSMC's revenue, and it left not because of the market but because of a list. China still absorbs 44% of Taiwan's exports. Pineapples have already been used once, as a rehearsal for sanctions.

My own inclination is that being irreplaceable is not the same as being safe. A moat looks like an amulet — until the neighbour decides the moat is a threat. Then the moat becomes their reason to come over.

Tsai Ing-wen and Mark Liu cast TSMC as security policy. That made sense in 2001, when the term silicon shield was first coined — back then nobody could really afford the price of getting out. The new fact of the 2020s is that both sides have decided they can.

The sharpest layer should stay on the island

Taiwan shouldn't pick just one of “stay” or “move out.” What it should do is build a portfolio.

The most advanced process nodes, and the yield know-how that only Hsinchu can really explain, stay on the island. That's the genuinely scarce leverage. CHIPS money can buy a fab; it can't buy the feel that keeps all 700 steps stable. Arizona can exist, and should — the fourfold construction cost is an insurance premium paid to the US Congress and US customers, its purpose political entanglement, not cost. Morris Chang has already said the unpleasant part out loud: globalisation is almost dead, so you still have to go and build. Just don't move the frontier along with it. The day it moves is the day the shield expires.

Nor can the shield go on serving as the navy. Four months of compulsory service, units only about 80% staffed, military spending down to about 2% of GDP, and then US$8.1 billion spent on fighter jets — this is a rich island using a foundry in place of defence. The porcupine the analysts describe isn't pretty, and it won't lift market cap. But it is the second line that has to be added once the shield starts to depreciate.

Politically, the polls in the case are clearer than any slogan. What most people want is the status quo, not an immediate move to either end. After Hong Kong, the island's elections have already given one answer. Ambiguity here isn't weakness. It's the product the middle has already bought.

My own choice: keep the sharpest layer, pay an expensive American premium, reshape the military into something that can survive, and gradually bring down the 44% of exports going to a single counterparty. Every step is expensive. The other path is more expensive — a shield that both sides are making obsolete, sitting in the same typhoon track.

This isn't deglobalisation; it's permission being weaponised

Deglobalisation would grow three complete stacks: the US designing, manufacturing and supplying its own equipment; China able to survive without Dutch lithography machines and American software; and a third stack in between that sells to no one. That world isn't in the case.

What's in the case is: wafers are still crossing the sea, software is still being licensed, talent is still flying. In 2023, China's imports of chips and equipment were still near record highs. Nvidia didn't leave; it modified its cards to sit just below the threshold. The Netherlands and Japan followed with controls, not by bringing factories home. The US still makes more than 70% of design software. The only company that can make EUV lithography machines is still in the Netherlands.

What has changed is permission. The foreign direct product rule claims jurisdiction over chips that contain “no American atoms.” Investment screening can review a Korean company with almost no US footprint. A single entity list can make your second-largest customer disappear within a quarter.

My own understanding is that the globalisation of atoms is still running. The globalisation of permission is being rewritten. Those are two sentences. The first explains why a conflict on this island would still bring the world to a halt. The second explains why both great powers feel that, given time, they can care a little less.

Morris Chang says globalisation is almost dead. What he's describing is the gates. Not the road.

The bill will land on mature nodes first

Even without full deglobalisation, the price is already being paid.

Arizona's fourfold construction cost will sit inside every phone and every server, or inside a fiscal subsidy that can't be cleanly carved out. When every capital pours money into the easiest segment of the chain, mature nodes end up in oversupply — TSMC's CEO already said on an earnings call in 2024 that too much mature-node capacity may have been built. The truly scarce things — EUV, yields, the discipline of 5 nanometres — will stay scarce for longer. The old model put manufacturing risk on one specialist island so a hundred design companies could iterate. Three half-finished islands will waste capital on concrete instead of on the next nanometre.

Families, carmakers and cloud providers will start maintaining two bills of materials. One that can get into this side, one that can get into the other. This isn't two planets. It's two sets of passes for the same road.

My own understanding is that the word deglobalisation is too big, and too convenient. It makes people think the world is splitting into three blocs that don't deal with each other. The more accurate picture: the road is still there, the checkpoints have multiplied, the premiums have gone up, and the checkpoints will move once more just after you think you've already diversified.

Takeaways for Arcus42

Reading about this island that wrote its chips into a shield, what came to mind wasn't whether we should talk about geopolitics, but whether we've been treating concentration as an amulet that expires.

First, being irreplaceable is not the same as being safe. One fab holding 92% looks like a moat. Moats can also become bullseyes. A family that puts its operating company, its listed entity and its living expenses all in one place is a scaled-down version of the same story.

Second, what makes you rich doesn't necessarily keep you safe. Cash flow isn't a navy. A listed company isn't defence policy. When your core business becomes someone else's reason to make policy, you need a second line that doesn't make money — structure, liquidity, jurisdictions you can switch between.

Third, permission is now a due diligence item. Clients, licences and listing venues can vanish within a quarter because of a list. Before asking what an asset is worth, ask: if permissions change, can it still be serviced, sold, and handed to the next generation?

Fourth, diversification depends on whether the underlying risk is the same. Ten products with six identical exposures isn't a portfolio. Two bills of materials, two custody locations, two settlement currencies — that's when it starts to look like a portfolio. Things that look globalised are often just different brand names on the same chain.

Fifth, don't bet on which capital wins the tech war. A government can spend US$52 billion to move output share by two points. A family can't spend money like that. Our job is to make sure the family can still function under any outcome. The premiums will be expensive. Expensive, because the shield is depreciating.

Final Thoughts

Before class, what I agree with most is this: “the silicon shield is turning into a bullseye faster than it can go on being a shield.”