彼得·林奇(Peter Lynch) 视角 — TSM

I'll tell you up front where this one lands, because honesty is the whole job. TSM is the foundry that makes the most advanced chips on earth. It's a wonderful company. But it sits right on the edge of my circle — the same edge where I looked at NVIDIA and said, plainly, "I can't explain it the way I can explain a donut chain." So read this knowing the man writing it passed on the chip business on purpose. Let me do the homework anyway and show you exactly where I get off the bus.

The Two-Minute Drill

Here's the crayon version. TSMC doesn't design chips and it doesn't sell them to you. It's a contract manufacturer — a foundry. Apple, NVIDIA, AMD, Google all draw up their chips and pay TSMC to actually build them, because TSMC is the only company on the planet that can reliably make the tiniest, fastest ones (2-nanometer, just now in production). Think of the best print shop in the world: everybody brings their design, nobody can print it as well or as cheaply, so everybody lines up.

Why it'll do well: AI needs an enormous number of these advanced chips, and almost all of them run through this one factory. Revenue grew 34% last year and management guides for more than 30% growth again this year. Margins are fat and getting fatter — gross margin hit 66% last quarter.

What could go wrong: Three things. One, the factory is in Taiwan, and that's "the world's most dangerous geopolitical fault line" — China's claim over the island is a low-probability, total-loss risk I cannot model and frankly nobody can. Two, the whole AI buildout could prove to be over-ordering by a handful of cloud giants; if they pull back, this company's $52–56 billion of annual capital spending becomes a millstone. Three, it's a brutally capital-hungry, technology-treadmill business — fall one node behind and you lose the crown.

Where the thesis breaks: Two straight quarters of decelerating growth would tell me the AI super-cycle is rolling over. And the price tag already assumes the good story continues — pay up here and you're paying ahead of the growth.

Category

Fast grower riding on top of a deeply cyclical, capital-intensive industry — a fast grower inside a cyclical box. That straddle is the whole story, and the danger, as I always warn, is "pretending you're in the safe one."

Right now the numbers scream fast grower: 30%+ revenue growth, expanding into AI the way a restaurant chain expands into new states. But semiconductors are a textbook cyclical — autos, steel, chips, they all swing with the macro cycle. The single most expensive mistake an amateur makes is "buying a peaking cyclical — earnings high, P/E low, everything looks great — under the belief it's a fast grower." TSM's earnings are at a record, margins at a record, and the multiple looks almost reasonable because earnings are peaking. That is exactly the configuration that has fooled people for a hundred years.

The honest read: TSMC has a genuinely better moat than a typical cyclical — it's near-monopoly on the leading edge, not a commodity steel mill. So it deserves more fast-grower credit than, say, a memory-chip maker. But I will not pretend the cycle has been repealed. The rules of this box are inverted: the time to love a cyclical is when the P/E is high on trough earnings and the business looks awful. This business looks magnificent. That's the warning light, not the green light.

The Crayon Test

Borderline — and that's a verdict. I can draw the print-shop picture and feel good about it. But the moment I have to judge whether 2-nanometer yields hold up against A16 against what Intel and Samsung do next, whether CoWoS packaging is the real bottleneck, whether AI demand is durable or a bubble of cloud-giant over-ordering — I'm no longer using a crayon. I'm guessing on physics and on a capital-spending arms race I can't independently verify.

This is precisely NVIDIA all over again for me. I told people: "Missing a great stock you don't understand is fine. It's the price of not owning disasters you don't understand." TSMC might well be a great stock. But the part that determines whether it's great — staying one physical step ahead on a treadmill that costs $55 billion a year to run — is the part I cannot explain to a 12-year-old, and therefore the part I don't truly own.

PEG / PEGY Check

• Trailing P/E: 31.71 (stockanalysis.com TTM, 2026-06-12; a second source shows 37.2 on a different basis — I'll use the lower, cleaner one).
• Earnings growth rate (and how sourced): I refuse to plug in the +58% one-quarter net-income spike — that's a number designed to make the PEG look like a gift, and "if the only way to get the PEG under 1 is to assume heroic growth, that's your answer." A defensible forward rate is management's own >30% near-term revenue guide, hauled down toward the mid-20s once you respect the cyclical fade. Call it ~25–30%.
• PEG = P/E ÷ growth: 31.71 ÷ ~28 ≈ 1.1 — fair, not cheap. (Use the heroic 30%+ and you get to ~1.0; use a sober 20% normalized rate and you're at ~1.6, edging toward expensive.) So: priced fairly if the growth is real and repeatable, paying ahead of yourself if it isn't.
• PEGY (it pays a dividend): yield is only 0.64%, so PEGY ≈ 31.71 ÷ 28.6 ≈ 1.1. The dividend is a rounding error here — this is not a stalwart you hold for income; the yield rescues nothing.
• The cyclical caveat: here's the thing PEG can't see. These are peak earnings. PEG "breaks on cyclicals — at the trough, earnings are depressed and the P/E is sky-high by design." TSM is the mirror image: earnings inflated by an AI boom, so the P/E understates how richly you're really paying against mid-cycle earnings. A PEG of 1.1 on peak earnings is more expensive than it looks.

The Story — Why It Will Do Well (the reasons, written down)

These are what I'd watch; when one breaks, the story has changed.

1. Near-monopoly on the leading edge. ~90%+ share of advanced-node foundry, the only one shipping 2nm at 70–80% yield. This is a real moat — "a niche, a local monopoly nobody can easily attack."
2. AI demand funneling through one factory. NVIDIA, Google TPUs, custom ASICs all build here. HPC/AI went from ~5% of revenue (2020) to a projected ~28% (2026).
3. A user-adjacent position done right. Lynch attribute: "a user of technology, not a maker" — well, TSMC is the maker's maker. It sells picks and shovels to every chip designer, so it wins no matter which designer wins. That's the closest thing to my "people have to keep buying it" in this industry.
4. Margins expanding, not just revenue. 66% gross margin and rising — pricing power, not a price war.
5. Customers locked in by yield and IP. Top-10 customers are 76% of revenue, glued by switching costs.

Perfect-stock attributes present: a genuine niche / local monopoly; a product people must keep buying (every advanced chip needs it). Attributes MISSING — and this matters: it is not dull, not under-owned (a ~$1.9 trillion company that everyone on Wall Street follows), not under-covered, and it is the opposite of "a business any idiot can run" — it requires the most demanding engineering on earth, run flat-out, forever. "The ones I love are usually the ones nobody wants to talk about at a cocktail party." This is the cocktail-party stock of 2026. That alone makes my skin itch.

Balance Sheet & Survival

• Debt level: Low and clean. ~USD 34B total debt against ~USD 106B cash and securities — net cash of roughly USD 72B. Long-term borrowing jumped 7x in 2024 (overseas fab financing), worth watching, but this is a fortress balance sheet by any measure.
• Can it survive a bad stretch in its category? Yes, financially — emphatically. "It's very hard to go bankrupt when you don't have any debt." TSMC throws off USD 27B of free cash flow and has net cash; it can ride out a demand air-pocket that would bury a leveraged competitor. The survival risk here is not the balance sheet. It's the one thing a balance sheet can't fix: a missile, a blockade, an export-control wall. No amount of net cash survives the factory being in the wrong place at the wrong time, and that is a risk I have no tool to price.

Category-Specific Verdict & Sizing

• Buy rule for this category met? No — not cleanly. As a fast grower, the rule is "PEG reasonable and the runway long." PEG is ~1.1 (fair, not the under-1 bargain I hunt for), and the runway is real but the growth sits on top of a cycle near a euphoric high. As a cyclical, the rule is inverted — buy at a high P/E on trough earnings when the business looks awful. This business has never looked better, which by the inverted logic is a caution flag, not a buy signal. Neither box gives me a green light.
• Position posture: Observation post, not a build. If this somehow fit my circle, it earns a small research position you grow only as the story proves out — never a swung-for-the-fences buy at a fair price on peak earnings.

Sell / Hold Check — the only two reasons

1. Has the STORY changed? Not yet — growth is accelerating, margins expanding, the moat intact. For an existing holder, nothing in the fundamentals says sell. Do not sell this because "it went up a lot" — "the stock doesn't know you own it," and "most of the gain in a big winner comes in the middle innings." If you own it and understand it, the AI cycle still turning is a reason to hold, not bail. The tripwire is two straight quarters of decelerating growth.
2. Is there a clearly BETTER opportunity? For me, yes — a business I can actually explain with a crayon, bought at a PEG under 1. That's not a knock on TSMC; it's the circle of competence doing its job. I'd rather rotate into something dull and understandable than hold something brilliant and unexplainable.

What I'm Watching (the story-change tripwires)

• Each quarter, watch:
  • Revenue growth rate — the engine. Two straight quarters of deceleration = the AI super-cycle is cresting (the cyclical roll-over).
  • Gross margin — if 66% starts sliding, pricing power is eroding or the high-cost Arizona fabs (reportedly 50%+ costlier than Taiwan) are diluting the mix.
  • CapEx vs. utilization — $52–56B is the largest bet in the company's history. If demand softens while that spend is locked in, fixed costs crush earnings fast. That's the cyclical's classic trap.
  • Customer concentration / hyperscaler capex commentary — top-10 are 76% of revenue; if the cloud giants signal a spending pause, that's the demand domino.
• Sell signal: two consecutive quarters of decelerating growth (the fast-grower exit), OR a low P/E on record peak earnings with the cycle visibly cresting (the cyclical exit). And the un-modelable one: any material escalation in the Taiwan Strait — that's not a tripwire I can manage, it's a reason the whole position is uninsurable in my framework.

Lynch's Take

Pass — for me, not against the company. Call it "Not my lens."

Let me be straight, the way I am about Apple and Kaiser and the ones I blew. TSMC is a genuinely great business — fortress balance sheet, a real monopoly on the leading edge, growth a restaurant chain would envy. If you handed it to Buffett or to a chip engineer who can truly judge the 2nm-versus-A16 race, they might have a wonderful case. But I'm the fellow who looked at NVIDIA and said it was outside my circle on purpose, and the same wall stands here: the part of this story that decides whether it's great — staying one physical step ahead on a $55-billion-a-year treadmill, in a factory sitting on the world's most dangerous fault line — is the part I cannot draw with a crayon. "Missing a great stock you don't understand is fine. It's the price of not owning disasters you don't understand."

And even setting my circle aside, the Lynch math doesn't beg me in: a PEG around 1.1 is fair, not the gift I hunt for, and it's a fair price on peak earnings in a cyclical that has never looked better — which is exactly the configuration that fools amateurs into buying the top. I want under-owned and dull; this is the most-owned, most-talked-about stock in the world. None of my favorite attributes are flashing buy.

So: if you already own it and you genuinely understand the chip race, don't let me scare you out — "the real key to making money in stocks is not to get scared out of them," and the story is intact. Hold while the AI cycle runs, and watch those two quarters of growth like a hawk. But for me to buy it new, at this price, in a business I can't explain? No. I'll turn over the next rock instead.

基于 2026-06-13 共享数据；本分析为单一大师框架的演绎,非投资建议。