The Island That Runs the World

May 06, 2026

∙ Paid

You take a flight to Taipei, rent a car, and drive 80 kilometers southwest. You arrive in a city called Hsinchu. At first glance, it’s an ordinary Taiwanese city. Three hundred thousand inhabitants. Office buildings. Noodle shops. April rain.

But if you look more carefully, you see something else. The streets are unusually clean. Traffic flows smoothly. Locals speak surprisingly precise English. And to the northeast of the city, behind security gates you can’t pass, there’s something. A campus. Several enormous buildings, no windows, ventilated continuously by systems you can hear from the road.

Welcome to Hsinchu Science Park. The headquarters of Taiwan Semiconductor Manufacturing Company. TSMC.

I want you to visualize what’s happening inside. In a “fab” — a chip fabrication plant — the air is a thousand times cleaner than an operating room. Engineers wear full sterile suits. The machines that etch transistors onto silicon are EUV (Extreme Ultraviolet) systems made by a single company in the world, ASML in the Netherlands. One machine costs about $200 million. It uses a laser that hits a droplet of molten tin at temperatures comparable to the surface of the Sun, generating light at 13.5 nanometers wavelength — short enough to etch structures finer than a strand of DNA.

The most advanced transistors TSMC manufactures today — on what’s called the 2-nanometer node — measure barely a few dozen atoms across. To give you a sense of scale: if a human hair represented the distance from Paris to New York, a 2nm transistor would be roughly the length of a postage stamp.

And it’s one single company that knows how to do this at commercial scale. With yields that make it economically viable.

Here’s the information I want you to internalize before reading the rest of this edition: on May 6, 2026, the day I’m writing this, TSMC alone manufactures roughly 90% of the world’s most advanced semiconductors. It’s the company that produces 100% of Nvidia’s H100 and B100 chips. 100% of AMD’s MI300. Almost all of Apple Silicon. Every high-end chip from Qualcomm, MediaTek, Broadcom. The iPhone 18 coming soon will use A20 chips made by TSMC on the 2nm node.

Apple alone has pre-booked more than 50% of TSMC’s initial 2nm capacity. TSMC’s 2nm capacity for 2026 is fully booked. The order book is full before the machines even spin up.

And if tomorrow morning, at 4 AM Taipei time, TSMC closed its doors — for any reason — the global AI industry would stop. Not in ten years. Not in two. Within six to nine months, the time it takes for inventory at Nvidia, Apple, AMD, Microsoft to run out.

That’s what we’re going to talk about.

The numbers nobody has really internalized

OK, let’s look at the scale of this. Because when I say “TSMC is a monopoly,” I want you to understand how extreme — extreme even by modern tech standards.

On May 5, 2026, TSMC stock (TSM) closed at $401.71. Market cap: $2.06 trillion. Up +135% over 12 months. It went from the 10th largest company in the world by market value to the 7th or 8th depending on the session. And yet, I’m going to tell you something that should worry you: in my view, it’s still undervalued relative to its structural position. Not overvalued. Undervalued. We’ll come back to that.

Here are the key numbers I want you to remember:

On market share, TSMC controls roughly 70% of the global pure-play foundry market, and 90% of the “advanced nodes” segment (≤7nm) — the segment that matters for AI. No other company comes close. Samsung, its nearest competitor, has about 8% of the advanced node market. Intel Foundry is trying to come back but remains marginal in external commercial work.

On 3nm capacity, TSMC is going from 120,000-130,000 wafers per month at the end of 2025 to a target of 180,000 wafers per month by end of 2026 — a 40%+ expansion in one year. And even at that level, capacity isn’t enough. CEO C.C. Wei stated explicitly: demand for advanced-node wafers is three times available capacity. Three times.

On 2nm capacity, mass production began in Q4 2025 in the Hsinchu and Kaohsiung fabs. Current capacity is around 30,000 to 40,000 wafers/month. End-2026 target: 100,000 wafers/month. A 2.5x to 3x ramp in one year. And again: fully booked for 2026.

On wafer pricing, TSMC just made an unusual move: a 2nm wafer now costs roughly $30,000, a 50% premium over a 3nm wafer. Customers pay without flinching because there’s no alternative.

On margins, TSMC went from a gross margin of 54% in 2023 to 59% in Q3 2025. For an industrial company running $30 billion factories, that’s abnormal. That’s the margin profile of a software publisher, not a foundry.

So here’s the full picture: one company, on one island, with software-style margins on heavy industry, facing demand that’s structurally three times its capacity, in a sector that defines the next decade of the global economy.

That’s the TSMC monopoly. And it’s unprecedented.

My thesis, and it’s uncomfortable

OK, here’s what I think, and I’ll try to be fair to the complexity of it.

The dominant narrative on TSMC is: “Great company, great position, but the geopolitical risk of Taiwan makes it un-investable long-term.” You see this everywhere. And it’s partially true. The risk is real. The geographic proximity to mainland China — 180 kilometers from Xiamen, the same distance as Paris-Reims — is a fact. Xi Jinping has stated multiple times that “reunification” is “inevitable”. The Chinese military keeps multiplying encirclement exercises around Taiwan. I’m not underestimating this risk.

But here’s what I think most analysts miss: they reason as if “geopolitical risk” and “TSMC’s monopoly” were two separate things you add together. As if you could have TSMC without Taiwan, or Taiwan without TSMC.

They’re the same thing. And that’s exactly what makes the situation so stable — and so fragile at the same time.

Here’s my reasoning. No country in the world has any interest in TSMC stopping. Not the United States, which depends on TSMC for the iPhone, for Nvidia chips that run ChatGPT, for the semiconductors in F-35 fighters. Not China, which consumes massive amounts of TSMC chips for its domestic industry despite US restrictions, and whose economy would suffer enormously from a rupture. Not Taiwan, obviously. Not Japan, Europe, Korea. Not even Russia.

This concentration creates what international relations theorists call a mutual-dependency equilibrium — a bit like nuclear deterrence but without the weapons. If anyone attacks Hsinchu, everyone loses, immediately, massively, simultaneously. This is what some call the “silicon shield” — the implicit doctrine that Taiwan can’t be invaded because the global economic value destroyed would be too great.

Now, first important nuance. The silicon shield works as long as the actors are rational. That’s its strength and its fragility. Is Xi Jinping’s China in 2026 a strictly rational actor on the Taiwan question? I have my doubts. The Russian precedent in Ukraine showed that a leader can accept massive economic cost for ideological reasons of sovereignty. The silicon shield is a bet on Chinese rationality. And bets on human rationality have historically had imperfect success rates.

Second nuance, and this is the one that really interests me. The whole world is trying to reduce its dependency on TSMC. Trump is pushing to reshore manufacturing to the United States — TSMC announced in March 2025 a $100 billion investment on top of the $65 billion already committed in Arizona. The Arizona fab is producing 4nm and 3nm today. But — and this is crucial — 2nm in the US isn’t expected before late 2028, in the best case. And Arizona production suffers from labor costs 4-5x higher than Taiwan, visa issues bringing in experienced Taiwanese engineers, and a slower regulatory environment.

Meanwhile, in Taiwan, TSMC is opening five new 2nm fabs simultaneously. Five. The same year. Unprecedented in the industry’s history.

So here’s the paradox: everyone says we need to diversify away from Taiwan. Everyone is investing. And yet, TSMC Taiwan’s share of advanced production globally will probably increase, not decrease, over the next 3-4 years.Because AI demand is exploding faster than international fabs can be built. Diversification is behind on concentration.

Which brings me to the central thesis I want to give you.

The real geoeconomic question of this decade isn’t “who controls the AI models”. It’s “who controls their physical manufacturing”. And the answer today is one single actor, in one country, under permanent threat of invasion. This concentration is more extreme than OPEC’s in 1973 — when 13 countries together controlled 56% of global oil. Here, it’s one company, in one city, manufacturing 90% of a resource that’s becoming as structural as oil was in the 20th century.

And the real mystery that haunts me: why isn’t this concentration priced more aggressively in the markets?

Why TSMC is probably still undervalued

I know that’s going to sound counterintuitive. A company up 135% in a year is rarely described as undervalued. But look at the math.

TSMC trades at roughly 32-33x trailing earnings. Expected earnings growth is 33% per year compounded through 2028 per consensus estimates. The PEG ratio (P/E divided by growth) lands around 1, which is low for a company of this quality.

Compare to Nvidia: 30x forward earnings, more volatile expected growth. AMD: 38x forward. Apple: 30x but with growth barely at 8%.

TSMC trades like a cyclical industrial company even though it has the characteristics of a software monopoly. I read this discount as a permanent geopolitical risk premium implicitly priced in. The market applies a 30-40% discount versus what an equivalent US or European company would pay. And this discount has remained roughly stable for 5 years — it has not compressed despite the AI boom.

Here’s what that means concretely. If geopolitical risk resolves peacefully (lasting status quo, no invasion, accelerating diversification), the discount compresses and the stock can gain another 50-100% over 3 years. If the risk materializes (invasion, blockade, serious disruption), you lose 60-80% in days. It’s a fundamentally asymmetric trade — but not in the usual sense. It’s positive-asymmetric, not negative-asymmetric.

The implicit bet you take when buying TSMC in 2026 is: will the rationality of mutual dependency hold for another 5 to 10 years? If yes, it’s one of the best long-term positions in the market. If no, you’ll lose a lot. But — and this is the point that fascinates me — the entire world is taking that bet too. Apple is. Nvidia is. The American administration, despite its rhetoric, is. You’re not more exposed than the global tech civilization.

Except you can be paid for that exposure.

What you should be asking yourself this week

If you read Public Markets, you’re probably thinking about how to build your long-term portfolio. So here are the real questions, not the headline ones:

Why does the market apply a geopolitical risk discount to TSMC that hasn’t compressed despite increasing diversification efforts? Is that discount rational, or a bias?
If you don’t have TSMC in your portfolio in 2026, is that an active choice or an omission by default? Most European retail investors don’t own it. Why?
What percentage of your tech portfolio is actually exposed to TSMC, without you knowing it, through your positions in Nvidia, Apple, AMD, or an S&P 500 / MSCI World ETF?
How do you evaluate the “Taiwan” risk rationally? Which indicators would tell you before the market that it’s materializing — and which are noise?
What’s the scenario where TSMC loses its monopoly position? Is Intel 18A a real threat? What happens if Samsung succeeds at 2nm?
And the genuinely uncomfortable question: if you refuse TSMC on geopolitical principle, do you also have to refuse the entire ecosystem (Nvidia, Apple, etc.) that depends entirely on it? If not, why not?

🔒 Inside the premium edition

Here’s what I’ve put together for Public Markets subscribers. Concrete, line by line:

→ The full economics of a TSMC fab — how much a 2nm fab costs, how much it returns, how the ROI is calculated, and why no competitor can enter economically at these investment levels.

→ The competitor map — Intel 18A vs Samsung Foundry vs Rapidus (Japan) vs SMIC (China). For each: where they actually are (not their PR), their technology gap, their probability of catching up to TSMC at different horizons.

→ The geopolitical indicator dashboard — the 8 concrete signals I track to evaluate Taiwan risk in real time (ADIZ incursions, PLA exercises, Taipei arms purchases, undersea cable incidents, etc.) with my alert thresholds.

→ The 5 underpriced derivative plays — beyond TSMC itself, the less obvious companies that capture rent on its position: ASML (lithography), Tokyo Electron, Lam Research, Applied Materials, and one Japanese name nobody is watching.

→ The 3-scenario analysis — status quo, limited escalation (blockade), black scenario (invasion). For each: price impact on TSMC, on Nvidia, on Apple, on indices, on the dollar, on gold. And the buy/sell levels I’m setting.

→ The tipping-point calendar — the 7 dates between now and end of 2027 where Taiwan-China-US can shift concretely (Taiwanese elections, Chinese plenums, Trump tariff deadlines, US arms deliveries to Taipei).