No matter — I'll do the arithmetic inline. NPM 2025 = 55.21/122.42 = 45.1%. Rev CAGR (NT$) 2021→2025 = (3809/1587)^0.25 − 1 ≈ 24.5%. Trailing P/E on $1.99T / $59.68B ≈ 33.3. All consistent with the sourced figures. Writing the report.

TSM — Research Report (2026-07-14)

TL;DR

• What it is: The world's largest contract chip manufacturer (a "foundry" — it makes chips designed by others, like Nvidia and Apple, and sells no products of its own).
• My take: A near-monopoly on the most advanced chips on Earth, printing cash — the debate is price and geopolitics, not business quality.
• Revenue $122.4B, +36% in 2025 (AI demand, not a fluke).
• Gross margin 59.9% (rare pricing power for a manufacturer).
• Net margin ~45% (most manufacturers earn 5–15% — this is software-like).
• ~70% of the global foundry market (structural dominance, not a lead).
• Strongest bull: Every AI chip in the world is physically made here — demand is booked out years ahead.
• Strongest bear: It sits 130 km from China; a Taiwan conflict is a permanent, un-hedgeable tail risk.
• Watch next: 2nm ("N2") ramp through 2026 and whether the Arizona/US fabs hold margins.

So what: TSMC is the physical chokepoint of the entire AI and electronics economy — it manufactures the chips almost every major tech company designs but cannot build.

1. Business overview

So what: TSMC doesn't design chips or sell gadgets — it rents out the world's most advanced factories to whoever has a chip design, and takes a manufacturing margin most factories can only dream of.

A foundry is a pure-play chip factory: customers (Nvidia, Apple, AMD, Qualcomm, Broadcom) send their designs, TSMC etches them onto silicon wafers and ships them back. TSMC keeps no products of its own, which is why it can serve competitors simultaneously without conflict.

The business runs on process nodes — "3nm," "5nm," "2nm" measure how small the transistors are; smaller = faster and more power-efficient = commands a premium price. TSMC is 1–2 years ahead of everyone at the leading edge.

The 2024→2025 story in one line: Nvidia overtook Apple as the #1 customer, as AI data-center chips replaced the iPhone as TSMC's growth engine. (CNBC, Manufacturing Dive)

2. Financials (5-year)

So what: Revenue nearly doubled in two years and margins went up while it doubled — the opposite of what usually happens when a manufacturer scales.

All figures in NT$ billion unless noted. (NT$ = New Taiwan Dollar; ~NT$31 ≈ US$1.)

Revenue CAGR 2021→2025 ≈ 24.5%/yr (NT$). Note 2023 was a real down year (−4.5%) — TSMC is still cyclical; the AI wave masked but didn't repeal that. (TSMC IR, SEC 6-K FY2025)

Cash flow reality (the part factsheets skip)

So what: TSMC gushes operating cash, but spends most of it building fabs — free cash flow is real but thin relative to profit, because staying ahead is brutally capital-hungry.

Read this carefully: net income was $55B but FCF was only $18B because **$41B went straight back into new factories.** That's the foundry bargain — you must out-spend rivals forever to stay one node ahead. Profit is opinion, cash is fact, and here the cash is heavily pre-committed to capex. (Motley Fool Q3 transcript, GuruFocus FCF)

3. Metrics that matter — with the why

So what: Every number below is high and rising — the rare combination that signals durable pricing power, not a temporary boom.

4. Unit economics — the one number the business runs on

So what: TSMC's whole business is revenue per wafer, and the mix is shifting to its most expensive wafers ever.

The unit is a wafer — a silicon disc that gets diced into hundreds of chips. A leading-edge (3nm) wafer sells for ~$18,000–20,000; a mature-node (28nm) wafer is a few thousand. So the single lever that drives everything is what % of wafers are leading-edge — and that number just hit 74% at 7nm-and-below.

On top of raw wafers, TSMC sells advanced packaging (CoWoS) — stitching multiple chips + memory into one AI accelerator. CoWoS capacity is growing >80%/yr and is fully booked; TSMC pushed CoWoS price increases of 15–20% because it is the bottleneck for every Nvidia GPU. This is the clearest single sign of pricing power in the whole company. (longyield/SemiAnalysis, Silicon Canals)

5. Value-chain position — bottleneck or commodity?

So what: TSMC is the bottleneck of the AI economy — but it in turn depends on a bottleneck above it (ASML), and its customers have nowhere else to go.

Chip designers        →   TSMC (fab)        →   Their customers
(Nvidia, Apple, AMD)      = THE chokepoint       (cloud, phones, cars)
       ↑                        ↑
  design tools            equipment: ASML holds
  (Synopsys/Cadence)      the ONLY EUV lithography
                          machines — TSMC's own bottleneck

• Who has pricing power over TSMC? Almost no one downstream. Nvidia/Apple cannot get leading-edge chips elsewhere at volume — Samsung is a distant #2 with yield problems, Intel's foundry is unproven. TSMC raised prices in 2025 and customers paid.
• Who does have leverage over TSMC? ASML — the sole maker of EUV (extreme-ultraviolet) machines that print the smallest transistors. Every advanced foundry, TSMC included, must buy from ASML. That's TSMC's bottleneck above it (a value-chain lesson worth internalizing: even the chokepoint has a chokepoint).
• Commodity risk? Only at mature nodes (28nm+), where Chinese foundries (SMIC) compete on price. TSMC's frontier business is the opposite of a commodity.

6. Competition & market

So what: TSMC's lead at the leading edge is widening, not narrowing — the two rivals are years and margins behind.

The market itself (AI/data-center silicon) is growing double-digits and TSMC is gaining share within it. This is the §0 ideal: a bottleneck with monopoly structure inside a macro-underwritten trend. (Semiwiki 2nm, Semiecosystem)

7. Ownership & management

So what: No controlling owner, ~74% foreign-institution-held, with the Taiwan state as the largest single (but minority) shareholder — a proxy for how strategic this company is to a nation.

The Taiwan government's 6.4% and its role as national crown jewel is a feature (the "silicon shield" thesis — Taiwan's chip dominance deters conflict) and a risk (TSMC is a geopolitical hostage) at the same time. Dividend rising steadily (NT$14→18) signals confident, shareholder-friendly management — no signs of empire-building beyond the (necessary) capex. (Wikipedia/TSMC, KamilFranek)

8. Valuation

So what: You're paying a premium price (~33× earnings) for the best business in tech — reasonable only if AI capex keeps compounding; expensive if 2026 disappoints.

What a reasonable buyer pays: A pure quality investor could justify ~25–30× forward earnings given ~30% growth, ~45% net margins, ~70% share and net cash — that's PEG ≈ 1 for a monopoly, cheap on quality. A value investor notes the trailing P/E is at a 10-year high, ~70% above its own historical average, and that 2023 proved the cycle isn't dead. Both are right: great company, priced for continued perfection. (stockanalysis.com, Macrotrends P/E, GuruFocus fwd P/E)

9. Bull / Base / Bear

Bull: AI capex is a multi-year supercycle; TSMC captures it all regardless of which chip designer wins; N2 + CoWoS pricing pushes margins past 60%; earnings compound 25%+/yr and the multiple holds. → stock doubles again.

Base: Revenue grows ~30% in 2026 as guided, then normalizes to mid-teens; margins ~58–60%; capex stays heavy so FCF lags profit; stock tracks earnings, ~+15–20%/yr, no re-rating.

Bear: AI capex digestion (a 2023-style air-pocket) hits in 2026–27; a customer over-ordered (Nvidia is 19% — concentration cuts both ways); US fabs dilute margins; and the multiple compresses from 33× toward its 20× average → stock falls 30–40% even with a fine business.

Catalysts: N2 volume ramp (H2 2026), CoWoS capacity doublings, 2026 capex guide ($52–56B) signaling demand confidence, Arizona fab margins.

What would prove the thesis wrong: (1) a quarter where TSMC cuts capex guidance — that's the tell that end-demand cracked; (2) gross margin rolling below ~55% (pricing power lost); (3) any credible sign a rival matched TSMC's leading-edge yields at volume; (4) a Taiwan-Strait escalation — the un-modelable one.

📚 What this company teaches

1. The bottleneck captures the margin, not the brand. Nvidia gets the headlines, but the only factory that can build its chips quietly raises prices 15–20% and everyone pays. When studying any value chain (§0), find who can't be routed around — that's where durable profit lives. And note the recursion: TSMC is Nvidia's bottleneck, but ASML is TSMC's. Chokepoints stack.

2. High margins + high capex ≠ high free cash flow. TSMC earns a ~45% net margin yet converts only ~a third of profit to FCF, because staying ahead demands ~$40B/yr of factories. "Profit is opinion, cash is fact" — always walk the cash flow statement down to FCF, never stop at net income.

3. A great business and a great investment are different questions. TSMC's quality is not in doubt; its price (33× earnings, a decade high) is the entire debate. The best company in a sector can still be a mediocre investment if you overpay at a cyclical/valuation peak — separate "is this good?" from "is this cheap?"

Self-test question: TSMC has ~$90B in cash and a net-cash balance sheet, yet its free cash flow is only ~$18B on $55B of net income. Where did the missing ~$37B of profit go — and why is that spending actually the source of the company's moat rather than a drain on it?

Sources: TSMC 2025 Annual Report/IR · SEC 6-K FY2025 results · Manufacturing Dive FY2025 · Q3 2025 call · CNBC Nvidia #1 customer · stockanalysis.com valuation. Full-year OCF/FCF figures marked "est." are aggregated from quarterly disclosures, not a single reported annual line — verify against the 20-F cash-flow statement when filed. Not financial advice.

Report delivered above per §3/§3b/§4. Two source caveats: annual OCF/FCF is aggregated from quarterly filings (labeled est.), and P/E feeds ranged 33–38× (I used the reconciled ~33× on TTM net income). Want me to save this to investing/TSM-2026-07.md and add a watchlist line?