The Moat and the Strait

The widest moat in technology just posted a record quarter — and the stock fell. It is the closest to a fair price it has been in a year. It is still not there, and the reason has a hundred miles of water in it.

Hold a modern AI chip in your mind — a Blackwell from Nvidia, a Trainium from Amazon, an M-series from Apple. They are designed by rival companies in rival countries and they compete ferociously. But they share a secret: every one of them is physically etched into silicon by a single company, on a single island, about a hundred miles off the coast of mainland China.

That company is Taiwan Semiconductor Manufacturing Company, and this week it did something that ought to have sent the stock up. Second-quarter revenue hit a record $40.2 billion, up 34% on a year earlier; net income rose 77%; the gross margin reached 67.7%; and management raised its full-year growth outlook above 40%. The shares fell anyway — down to $398, a two-month low, caught in the worst week for chip stocks in over a year. A record quarter, and the market sold it. That gap between the business and the price is the whole of what follows.

The company beneath the companies

Start with what TSMC is, because the word for it is unglamorous and the fact is staggering. It is a foundry — a contract manufacturer that builds chips other firms design and sells fabrication to everyone while competing with no one. Nvidia is fabless: it owns almost no factories and lets someone else bear the cost of building. TSMC is that someone, and its moat is the mirror image of Nvidia’s — it is made of capital intensity. A single leading-edge fab now costs north of $20 billion; TSMC spent $40.9 billion on capital in 2025 and guided this week to $60–64 billion in 2026, a sum no rival has come close to matching.

Here is the inversion that reorganizes the whole AI boom. TSMC also makes Google’s TPUs, Amazon’s Trainium, AMD’s accelerators, Broadcom’s custom silicon — the very chips trying to take share from Nvidia. The custom-silicon threat that hangs over Nvidia is, one layer down, simply another TSMC customer. TSMC does not have to win the AI war, because it is paid by everyone fighting it. It is the arms dealer to every side, which is why its position is more durable than any single combatant it serves.

Horizontal bar chart of foundry revenue share in early 2026: TSMC about 70%, Samsung about 7%, everyone else about 23%; leading-edge share exceeds 90%.

And capital is only the visible half of the moat. The deeper half is the learning curve — thirty years of tacit knowledge of how actually to make these chips, held in the heads of tens of thousands of veteran engineers. A competitor can buy the same lithography machines from ASML; it cannot buy the thirty years of learning that tell TSMC how to run them. That is why the gap has widened even as rivals spent billions to close it: TSMC put its 2-nanometer node into mass production at the end of 2025, while Samsung and Intel remain a node behind, with worse yields.

Capital that compounds, not capital that imprisons

Graham and Buffett were rightly wary of capital-hungry businesses: an airline pours cash into plant year after year just to stand still, and the owner sees little of it. TSMC will spend more than $60 billion this year and looks, on the surface, like exactly that trap — yet it still throws off tens of billions in free cash flow and pays a rising dividend, because 60%-plus gross margins let it earn far more on each fab than the fab costs to run. The distinction is the whole point: capital intensity destroys value when it buys no advantage and creates value when it buys a lead rivals have spent tens of billions chasing and not reached. The test is not “how much must it spend?” but “what does the spending buy?” In plain terms, this is the rare business where the bill for staying ahead is smaller than the reward for being ahead.

It can still betray you. Most of that spending is not optional growth but the maintenance of leadership, and the demand beneath it is cyclical: as recently as 2023, in the post-pandemic glut, TSMC’s revenue fell and profit dropped more than a fifth before AI revived it. The owner earnings are real and large, but they ride a cycle, and their durability at this level has not yet been demonstrated.

The risk that isn’t in the business

TSMC was founded in 1987 by Morris Chang, who essentially invented the pure-play foundry model, and it is run today by Chairman and CEO C.C. Wei with the same engineering discipline: thirty-five years of conservative margin guidance, quarter after quarter modestly beaten. The most consequential decision of this era, though, is geographic. Under pressure from customers and governments, TSMC is building fabs outside Taiwan — in Arizona (where it added a further $100 billion commitment this year), in Kumamoto, Japan, and in Dresden, Germany.

Be honest about what those plants are. They are higher-cost, a step behind the Taiwanese leading edge, and they dilute margins. They are not an economic upgrade; they are an insurance premium — a deliberate, margin-reducing hedge against something happening to the island. That management is willing to pay it tells you they take the tail risk seriously; so should you.

But the build-out is more than a one-time premium: it is the visible edge of a second cost that needs no conflict at all. The same indispensability that makes TSMC a target makes it a prize every government now wants a domestic copy of — the U.S. and European Chips Acts, Japan’s subsidies, China’s push for self-sufficiency. The result is sovereign redundancy: leading-edge capacity duplicated across borders, each new fab higher-cost and built because politics, not economics, demanded it — a slow, high-probability compression of the very margins that define the moat. The rupture may never come. This tax is already being levied.

The price already assumes the best

Which brings us to the number, and it settles the question before geography even enters. The test this method leans on is the earnings yield — a company’s earnings divided by its price, read against the risk-free 10-year Treasury, which today yields about 4.56%. Put simply: a stock’s earnings yield is what the business earns for you each year per dollar you pay; if it sits below what a government bond pays for no risk, you are being asked to accept less to own more risk.

Bar chart of TSMC's earnings yield at $398.37 on four bases — 1.7% (5-year average), 2.7% (FY2025), 3.4% (trailing), 4.0% (FY2026 forward) — every bar below the 10-year Treasury at 4.56%.

At $398, TSMC trades near 30 times trailing earnings and about 25 times the forward figure — an earnings yield of 3.4% trailing and 4.0% forward, both below the Treasury. And compute the value on every defensible basis and the spread is the analysis:

Valuation matrix, US$ per ADR, across earnings bases (5-yr avg $6.63, FY2025 $10.64, trailing $13.44, FY2026 forward $15.91) and multiples (15x/22x/30x); today's $398 is justified only in the two shaded top-right cells.

The price is defensible only in the top-right corner — on AI-elevated forward or trailing earnings, at a peak-of-range multiple. On mid-cycle earnings at a normal multiple, the same share is worth a fraction of today’s quote.

There is a wrinkle unique to TSMC. For most of its history the finest manufacturer on earth traded at a discount to its own American customers, precisely because the market demanded compensation for the Taiwan risk. In the 2026 AI euphoria that discount evaporated: a judgment that stood for a decade — that TSMC should be cheaper than the customers it supplies, because one of them could not be bombed and TSMC could — was quietly repealed while the risk stayed exactly as large. Nothing about the strait changed; only the market’s willingness to charge for it did.

The honest counter-case

The bull has a real answer, and it deserves stating at its strongest. If AI demand is structural rather than cyclical, TSMC’s earnings do not merely stay high — they keep climbing, and this year’s “expensive” multiple is next year’s ordinary one. Management is guiding to 40%-plus growth and raising capital spending into that demand, which is not how a company behaves when it sees a peak; on the forward figure the yield is already 4.0%, a hair under the Treasury rather than a chasm below. That case may well be right. But it requires the AI cycle not to breathe even once, and it still pays you nothing for the strait.

The verdict: revere the business, refuse the price

Wonderful? Without question — the widest, deepest moat here, and this week’s record only widened it.

At a fair price? No — but closer than it was. A month ago the shares were $453 and the earnings yield sat far below the Treasury; today, 12% cheaper and on a fatter earnings base, the forward gap has narrowed to almost nothing. That is real progress. It is still not enough, for two reasons that compound: an earnings yield below the risk-free rate is no margin of safety — the test that keeps Apple, Microsoft, and Nvidia on the watch list — and on top of that full price you are asked to bear a catastrophic, unrecoverable tail risk, and the margin erosion already arriving, for no extra compensation. Graham’s point was never that catastrophe is likely; it is that a risk whose severity is total deserves a discount however low its probability — and here that discount is absent.

So here is the trigger I will watch, because a verdict without one is just a mood. On today’s consensus earnings the forward yield only reaches parity with the Treasury around $350 — and parity is not a margin of safety, and pays nothing for the island. To be paid for the strait — a forward earnings yield of roughly 5.5% to 6% — you are looking at the high-$200s. At $398 the stock is about 14% above the first line and 40% above the second. Closer than it was. Not there.

I should name my bias: I revere this company, and reverence is exactly the emotion that talks investors into paying up. It does matter. I lean toward “cycles are always cycles,” and I notice I lean that way partly because it lets me keep my discipline — the reasoning I would question in someone else. And a disclosure that is the whole point: I have no edge here. My only relationship with TSMC is yours — every phone and laptop I have used this decade runs on silicon it etched, and for most of that time I could not have named it. The most indispensable company in your life is one you never think about; the market only recently started to, and when it did, it stopped charging for the danger.

What I’m watching

Tracking table of falsifiable thresholds and one early-warning trigger for the TSMC verdict.

What this means for you

If you own it, you own a genuinely wonderful business at a price that leaves no cushion and pays nothing for its one uninsurable risk — not a reason to panic-sell a great company, but a reason not to add, and to know exactly what you hold. If you are waiting to buy, this week made the wait more interesting, not over: the finest manufacturer on earth is cheaper than a month ago, and a patient owner who has done the work now has a price in mind and a reason to act if it comes.

Questions people actually ask

Why did the stock fall on a record quarter? Because the record was expected and then some. The whole chip sector had its worst week in over a year on AI-rally profit-taking, and TSMC’s raised capital-spending guidance ($60–64 billion) stoked worries about free cash flow. When a stock is priced for perfection, “excellent” can still disappoint.

What is a foundry, in plain terms? A factory-for-hire. Fabless designers like Nvidia and Apple draw the chip; the foundry etches it into silicon. TSMC stock is the foundry almost everyone uses, especially for the most advanced chips.

Isn’t a below-Treasury earnings yield true of most big tech? Yes — and the method keeps Apple, Microsoft, and Nvidia on the watch list for the same reason. TSMC stock is distinct because it carries, on top of the full price, a risk that cannot be diversified or hedged, only priced.

This is general commentary for educational purposes, not financial advice, and not a recommendation to buy, sell, or hold any security. Figures are as of July 17, 2026 and will have moved by the time you read this. The author holds no position in any security mentioned.

Keep reading: https://www.wonderfulatafairprice.com/post/spacex-stock

Subscribe to the free Wonderful at a Fair Price newsletter for the weekly read on markets through a value-investing lens.