沃伦·巴菲特(Warren Buffett) 视角 — TSM

(Taiwan Semiconductor Manufacturing — NYSE ADR, 价 USD 421.07，市值约 $1.92T，2026-06-13)

I ran this through my own loop in the order I actually decide in: understand the business, judge the moat and its durability, look at the people, then — and only then — price it and sleep on it. Two things stopped me cold. One I can name in a sentence; the other took the arithmetic. Let me walk you through both.

Verdict

Watchlist (lean Pass). A genuinely wonderful business — one of the widest moats in the world — but at this price Mr. Market is asking me to pay a quality multiple for heroic, decade-long continuation, and the whole enterprise sits on a geopolitical fault line I cannot underwrite. Wonderful company; not a fair price; and a tail risk that fails my Sleep Test. I'd admire it from the bench, not buy it here.

Circle of Competence — Boundary

How it makes money, in one breath: TSMC is the world's contract chip kitchen. Companies that design chips but don't want to spend $20 billion building a fab — Apple, NVIDIA, AMD — hand TSMC their blueprints, and TSMC manufactures the silicon at a yield and a node nobody else on Earth can match. It gets paid per wafer; the customer leaves only if a rival can match its process and yield, which presently none can.

That part I understand as a business — it sells a critical service, has obvious pricing power, and a customer base that is glued in. But mark the word Boundary honestly. Two things sit at the edge of, or past, my circle. First, the technology cadence — whether N2 yields hold, whether A16 lands ahead of Intel, whether the next node after that is even physically achievable — is an engineering judgment I am not equipped to make. I missed Google when GEICO was buying its ads and I held IBM too long; I know what it feels like to be a tourist in the semiconductor business. Second, and decisively, the value of this enterprise rests on a geopolitical question — the Taiwan Strait — and I do not forecast geopolitics any more than I forecast the stock market. When the core risk of an investment is a macro/political tail I cannot price, the honest answer is that my framework gives only a weak signal. I'll give you the business read; I'll tell you plainly where I'm out of my depth.

Key Assumptions (the case hangs on these — check them later)

1. The process-technology lead is durable for 10 years, not just today. TSMC stays one to two nodes ahead of Samsung and Intel and converts that into yield and cost advantage the customer can't get elsewhere.
2. AI demand is a durable secular wave, not a capex bubble. The data says AI silicon goes from ~5% of revenue (2020) toward ~28% (2026). If hyperscaler capex rolls over, the whole growth story re-rates hard.
3. The Taiwan Strait stays quiet. A conflict isn't a drawdown — it's a permanent loss of the asset. The entire valuation assumes this never happens.
4. The market's implied ~20%+ owner-earnings growth for a decade actually shows up. My reverse-DCF says today's price requires it; history (26% NI CAGR off a smaller base) made it look plausible, but a $1.9T company compounding at 20% for ten years is a rare thing.
5. The U.S./Japan/Germany fab build-out doesn't permanently crush the margin. Arizona reportedly runs 50%+ costlier than Taiwan; geographic diversification trades moat-durability for margin.

Business Quality & Moat

This is where TSMC shines, and I won't pretend otherwise.

• Moat mechanism — name it, not adjective it: It is low-cost scale plus a physical/technological process barrier, reinforced by switching costs. Mechanically: TSMC is the only foundry mass-producing 2nm-class chips at 70–80% initial yield. Yield is cost — a customer who moves to a rival with worse yield pays more per working chip. On top of that, the customer's design IP and yield-tuning data are co-developed with TSMC, so leaving means re-qualifying silicon at a less capable shop. That's a mechanism, not a slogan.
• Strength & trend: Wide and, for now, widening at the leading edge — >90% share of advanced nodes, ~38% of foundry overall. Gross margin has risen (Q1 2026 at 66.2% vs. mid-50s historically), which is the fingerprint of a moat that's holding and pricing up, not leaking. That is the opposite of the newspaper business draining out from under me.
• Pricing power — Yes, with evidence: Reported price increases into 2024–2026 with rising margins and rising volume. Pricing power is the single best one-question moat test, and TSMC passes it cleanly. Few businesses on the planet pass it this well.
• $1B-competitor test: Hand a rival $1 billion? They get nowhere — a single leading-edge fab costs $20B+, and the yield know-how took decades. Intel has spent vastly more than $1B and still trails by two to three nodes. No, a funded competitor cannot take this. That's a real moat.

If the analysis stopped at moat quality, this would be a fat pitch. It doesn't stop there.

Management & Capital Allocation

• Integrity: Pass. No dual-class shenanigans, one-share-one-vote, a long record of straight dealing and operational honesty. C.C. Wei carries Morris Chang's culture. I see no integrity flag — and integrity is my automatic veto, so passing it matters.
• Capital allocation: This is not the Berkshire style, and I'll be honest about the friction. TSMC pours nearly all its operating cash back into capex — $52–56B planned for 2026, the highest in its history — and pays a modest, growing dividend (~20% of OCF) with essentially no buybacks. Now: I don't fault reinvestment when the returns are there, and TSMC's ROIC (reported ~48–52% TTM) says the returns have been spectacular. A business earning 50% on incremental capital should reinvest every dollar it can. So I won't call this empire-building — it's rational while the moat holds. But understand what it means for me: this company hands its owners very little actual cash; the value is all in reinvestment compounding inside the business. That raises the bar on durability, because if the moat ever cracks, all that retained capex was buried in fabs that depreciate.
• Institutional imperative warning: Mild, present. Building in Arizona, Japan, and Germany at a 50%+ cost premium is partly customer-and-government-driven — "everyone wants supply outside Taiwan" — and there's a whiff of doing the politically expected thing rather than the purely economic one. It's defensible as risk diversification, but it is margin-dilutive empire-geography, and I'd watch it.

Owner Earnings & Financials

Here's the trouble, and it's structural to my method. TSMC looks far cash-poorer on owner earnings than its headline profit, because its capex is enormous — but most of that capex is growth, not maintenance. My discipline subtracts only maintenance capex. Maintenance capex isn't disclosed, so I estimate conservatively (depreciation, ~$18–19B, as a floor), which lands owner earnings roughly at or slightly above reported net income (~$36.5B for 2024). That's the generous read.

• Owner earnings (2024, est.): ~$36–38B (≈ net income; maintenance capex ≈ D&A floor). I am estimating maintenance capex — say so.
• Owner-earnings yield at $1.92T market cap: ~1.9%. My gate is ≥ 8%. It fails by a mile.
• Plain FCF (total capex): 2024 ~$28B → ~1.5% yield; 2025 est. ~$30B (capex jumped to $40.9B); 2026 FCF gets compressed by the $52–56B capex plan. EV/FCF in the data is 54.6×. Fails.
• ROIC (TTM, reported): ~48–52% — passes the ≥12% gate with enormous room. This is the strongest number on the page and the reason the business is wonderful.
• FCF conversion: Lumpy and well below 80% of net income in heavy-capex years (2023 FCF/revenue 13.5%) — a growth-capex artifact, not an accruals problem, but by the letter of the gate it fails.
• Net Debt/EBITDA: Net cash (~$71.6B net cash 2024). Fortress balance sheet — passes with no leverage worry. I like this a lot.
• Gross-margin stability: Within range and improving — passes; the moat is holding.
• Data sourced? USD 2024 figures official (SEC/Yahoo); TWD from SEC EDGAR XBRL; owner-earnings split and maintenance capex are my conservative estimates — flagged.

So: a wonderful return-on-capital business that, at this price, hands its owner a ~2% real-cash yield. That's the crux.

Intrinsic Value & Margin of Safety

I'd rather be approximately right than precisely wrong, so take these as a range.

• Conservative owner-earnings DCF (base ~$36.5B, 10% discount, 12× terminal):
  • Growth capped at 8% (my standing cap for a non-cyclical): IV ≈ $767B, or ~$148/ADR.
  • At an optimistic 10%: IV ≈ $875B, ~$169/ADR.
  • At 12%: ~$193/ADR. At 15%: ~$236/ADR.
  • All well below the $421 price. Semis are cyclical; my discipline says value off normalized, not peak, earnings — which pushes IV lower, not higher.
• Reverse-DCF — the honesty check: Hold $421 fixed and solve for the growth it implies. The price requires owner earnings to compound at roughly 20%+ for a full decade (my model only reaches today's $1.92T at ~20% growth). Against a 5-year net-income CAGR of ~26% off a much smaller base, that's not impossible — but a $1.9T enterprise sustaining 20% for ten years is heroic, well past my "1.2–1.5× historical" fair zone. The market is paying for continued acceleration, which is hoping, not investing.
• Margin of safety at $421: Negative. On my conservative range I'm being asked to pay ~2–3× a defensible value. Required for even a wide-moat compounder: 10–20% discount below value. There is no discount here — there's a premium.

A wonderful business justifies a thinner margin of safety. It does not justify a negative one.

Sell / Hold Check

(I don't own it, so read this as "would I buy/hold at $421?")

1. Moat breaking? No — it's widening at the leading edge. This is the bull's strongest ground.
2. Management integrity issue? No — clean. Not a sell reason.
3. Price wildly above intrinsic value? Yes — on my conservative owner-earnings math the price embeds ~20%+ decade growth. This is the binding reason to wait.
4. Clearly better use of capital? For me, yes by default — when I can't price the tail risk and the cash yield is ~2%, cash and patience beat reaching. There's no called strike for letting this pitch go by.

The Sleep Test — No (final override)

If the market shut for 10 years and I could not sell, would I happily own all of TSMC at $1.92T? No. Not because of the moat — the moat would let me sleep. Because the single asset sits in the one place on the map where a political event wouldn't be a drawdown, it would be a permanent loss of the enterprise. Risk, to me, is the permanent loss of capital, not a wiggle in the quote. I cannot put a probability on the Taiwan Strait, and a business whose worst case is "zero, and I couldn't have seen it coming" is one I can't own at a full price with the market closed. The Sleep Test overrides the spreadsheet, and here the spreadsheet already said wait — so they agree.

Key Risks (the few ways this becomes a permanent loss)

1. Taiwan Strait conflict. Low probability, total severity. Not hedgeable, not estimable. The whole valuation quietly assumes it never happens.
2. AI capex proving to be a bubble. ~28% of revenue is now tied to AI silicon and hyperscaler spending. If that demand was pulled forward, peak earnings re-rate and the "20% for a decade" the price needs evaporates.
3. Customer concentration + technology slip. Top 10 customers ≈ 76% of revenue; Apple + NVIDIA ≈ 34%. A single missed node (N2/A16 yield failure) against a concentrated book is how a wide moat starts draining — exactly the trend I'd sell on.

What I'd Watch

• Each year: (1) Leading-edge gross margin and yield — the moat's vital signs; the day they roll over, the durability thesis is failing. (2) AI revenue mix and hyperscaler capex commentary — the demand's durability. (3) Arizona/overseas margin drag — is geographic diversification quietly eating the cost advantage?
• Buy trigger (for me): A Mr. Market panic — a Strait scare or an AI-capex air-pocket — that drops the ADR toward my conservative range (~$170–240) with the moat and margins still intact. That's when a wonderful business goes on sale.
• Sell trigger (if I owned it): Sustained gross-margin compression / yield loss at the leading edge (moat narrowing), or any sign Samsung/Intel closes the node gap. The price going down, by itself, would never be my reason to sell.

Buffett's Judgment

Let me give you the conclusion first, then the reason. This is one of the great businesses of the age, and I'm passing on it here — that's not a contradiction, it's the whole discipline. TSMC has a moat I'd happily own: hand a rival a billion dollars and they get nowhere, customers pay up and don't leave, margins are rising. If I could buy it at a fair price with the world a calmer place, I'd back up the truck.

But two things stop me. The first is the price. My job is to buy a stream of cash for less than it's worth, and at $421 the stock hands its owner about a 2% real-cash yield while asking me to bet that a near-$2-trillion company compounds at twenty percent for a decade. I've watched a 26% grower look unstoppable right up until it wasn't; paying for the extrapolation is how smart people overpay for wonderful businesses. The second is the one I can't model at all — the Strait. I don't forecast geopolitics, and when the worst case isn't "I lose half" but "I lose all of it and never saw it coming," I can't own the whole thing with the lights off for ten years. That's my Sleep Test, and it says no.

So this goes on the watchlist with a note: wonderful company, wrong price, untouchable tail. There are no called strikes in this game. I'll let this pitch go by and wait for the day Mr. Market gets frightened about Taiwan or AI and offers it to me near what it's worth — moat intact. If that day comes, I'll be interested. Today, the answer is patience and cash. And I'll say plainly: the part of this that's about the chip-node race and the part that's about the Strait are both at or past the edge of my circle — so weight my voice on the price discipline, and route the technology and geopolitics to people who actually live there.

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。