TSM › analyze
TSM · Analyze
Date: 2026-08-04 · Price: $417.17 · Market cap: $2.16T · Sector: Semiconductors
Verdict: WATCH — conviction 6.5. The business is the strongest on this entire watchlist and is getting stronger: 72.3% foundry share and rising, ROIC up 21.7% → 26.3%, gross margin 56% → 67.7%, and both challengers have failed to land a single committed external customer at the leading edge. The price is the problem. It trades at 26.2x 2026 consensus earnings — not the 19.3x the data feed reports — which is fair value, not a discount. There is no margin of safety in front of a non-diversifiable Taiwan binary.
🔴 CORRECTION TO THE 2026-08-04 TECH SWEEP. That screen called TSM "ACCUMULATE, 19.3x forward, FV $432–518" and named it the cheapest quality in the list. The 19.3x was wrong — it came from a vendor forward-EPS figure of ~$21.60 against an actual 2026 consensus of $15.91. At the correct multiple the fair value falls to $350–480 and the verdict moves from ACCUMULATE to WATCH. Details in §3.1. The screen's relative call — TSM over UMC — is unaffected and if anything strengthened.
Knowledge base: kb.py find returned pitfall-adr-book-value-corrupts-price-to-book
(written earlier today, from this same name) — applied throughout and extended by two new
findings in §5. Regime context from Knowledge/Market/regime.md.
1. Fundamentals
⚠️ 1.0 Data constraints — read first
Three vendor problems affect this name, and every figure below is built to route around them.
| problem | effect | handling |
|---|---|---|
| roic.ai per-share data uses a 4.669B share count | Matches neither the 25.93B ordinary count nor the 5.186B ADR count. Every roic.ai per-share metric on TSM is scaled by a wrong divisor | All per-ADR figures below recomputed by hand from aggregates |
| Yahoo P/B of 87.00 and BVPS of $4.80 | USD ADR price ÷ ordinary-share book value, unconverted | Rebuilt: see §1.5 |
| roic.ai free plan caps at 2 years | The framework asks for 5–8yr CAGRs; only a 4-year window is available | Stated as a limitation. CAGRs below are 3-year and peak-to-peak (2022 and 2025 are both cycle peaks), which is at least cycle-neutral |
FX assumption: ~31.6 TWD/USD, derived from the company's own Q2 print (NT$27.25 ordinary EPS × 5 = NT$136.25 per ADR ÷ US$4.31 reported = 31.61). Used consistently throughout.
1.1 Growth and cash flow (NT$)
| FY | Revenue | Op income | Net income | OCF | Capex | FCF |
|---|---|---|---|---|---|---|
| 2022 | 2.26T | 1.12T | 992.92B | 1.61T | −1.09T | 520.97B |
| 2023 | 2.16T | 921.43B | 851.74B | 1.24T | −955.40B | 286.57B |
| 2024 | 2.89T | 1.32T | 1.16T | 1.83T | −964.98B | 861.20B |
| 2025 | 3.81T | 1.94T | 1.70T | 2.27T | −1.28T | 992.38B |
| CAGR (2022→2025, peak-to-peak) | rate |
|---|---|
| Revenue | +19.0% |
| Net income | +19.6% |
| Free cash flow | +24.0% |
| Operating cash flow | +12.1% |
| Diluted share count | 0.0% |
2023 was the cycle trough (FCF fell 45% to NT$286.57B). Because 2022 and 2025 are both peak years, the CAGRs above are cycle-neutral rather than flattered.
1.2 Margins and returns — the moat, quantified
| metric | FY2024 | FY2025 | Q2 2026 |
|---|---|---|---|
| Gross margin | 56.12% | 59.89% | 67.70% |
| Operating margin | 45.68% | 50.83% | — |
| Net margin | 40.02% | 44.57% | — |
| ROIC | 21.65% | 26.26% | — |
| ROE | 30.19% | 35.37% | — |
| ROA | 18.95% | 23.22% | — |
| Incremental operating margin | 54.68% | 67.13% | — |
| Effective tax rate | 17.66% | 16.97% | — |
A 67% incremental operating margin means two-thirds of every new revenue dollar reaches operating income. ROIC rising from 21.7% to 26.3% while invested capital grew from NT$5.26T to NT$6.39T is the single best moat evidence in this report — TSMC is deploying more capital at a higher return, which is rare and very hard to fake.
1.3 Capital allocation — a reinvestor, not a returner
FY2025 disposition of NT$2.27T operating cash flow:
| use | amount | % of OCF |
|---|---|---|
| Capex (growth) | NT$1.28T | 56% |
| Dividends | NT$466.78B | 21% |
| Buybacks | NT$0 | 0% |
| Acquisitions | NT$0 | 0% |
| Net debt change | +NT$37.37B | — |
| To balance sheet | ~NT$640B | 28% |
At a 26.3% ROIC, reinvesting 56% of operating cash flow is the correct decision and creates more value than any buyback would. But it has a consequence the buyer must accept: see §3.2 — this stock does not pay you in cash.
R&D was NT$246.43B (6.5% of revenue), up from NT$163.26B in 2022.
1.4 Balance sheet — fortress
| metric | value |
|---|---|
| Cash | NT$2.77T |
| Debt | NT$1.06T |
| Net cash | +NT$1.71T (~US$54B) |
| Debt/Assets | 13.4% |
| Current ratio | 2.46 |
| Interest coverage | ~157x |
| Goodwill | NT$5.89B (0.07% of assets) |
Essentially no leverage, essentially no goodwill. Nothing to discuss.
1.5 Per-ADR figures — rebuilt by hand
(aggregate ÷ 5.186B ADRs ÷ 31.6 TWD/USD)
| metric | FY2025 | reported by vendor | verdict |
|---|---|---|---|
| Revenue/ADR | $23.26 | $734.47 (TWD, mislabelled) | rebuilt |
| EPS/ADR | $10.38 | — | ✓ consistent with $11.64 TTM |
| FCF/ADR | $6.06 | — | rebuilt |
| Book value/ADR | $32.71 | $4.80 | ❌ vendor wrong by ~6.8x |
| P/B | 12.8x | 87.00 | ❌ vendor wrong |
Fundamentals scorecard: 9.5/10. The only deduction is the absence of a 5–8yr window.
2. Moat
2.1 The position
| player | Q1 2026 foundry share |
|---|---|
| TSMC | 72.3% (up from 70.4% the prior quarter) |
| Samsung | 6.5% |
| Everyone else | ~21% |
TSMC is not defending share; it is taking it, from 70.4% to 72.3% in a single quarter, in a record $47.95B quarterly market. Moat sources: efficient scale (dominant), switching costs (a tape-out is a multi-year, multi-hundred-million-dollar commitment to one PDK), intangibles (process IP and yield learning), and cost advantage (volume across the largest installed leading-edge base).
2.2 Adversarial stress-test — the attack has been attempted and has failed so far
"You are a well-funded rival — how do you take this?" Two parties are actively trying, one with a nation-state behind it.
Intel. 18A is in production on Panther Lake — internal silicon only, with no major external customer secured. 14A has zero committed external customers; Intel says two prospects are engaged on PDK 0.5 with firm decisions expected H2 2026 into H1 2027. There are reports Intel may axe 18A for foundry customers entirely to concentrate on 14A. After years and enormous public subsidy, the leading Western challenger has not taken one committed leading-edge external customer.
Samsung. The more credible near-term threat. SF2P has hit a 70% yield — the recognised "golden threshold" for high-volume manufacturing and the first time any competitor has demonstrated 2nm stability at that level. It establishes a real dual-sourcing option for the largest designers. But it converts to 6.5% share today.
Customer in-sourcing. Apple, Google, Amazon and Tesla all design their own silicon — and all of them still need someone to fabricate it. No hyperscaler is building a leading-edge fab. Designing a chip and running a $30B fab at 70%+ yield are different businesses.
Verdict: the stress-test does not draw blood. This is the rarest outcome in this framework.
2.3 N2 — the next node is already won
- 4x the tape-outs of N3 at the same point in its life
- >20 customer tape-outs received, >70 in the pipeline
- Already 3% of wafer revenue
- Management calls it the strongest customer adoption in company history
TSMC has also announced price increases of up to 10% starting 2027. Companies without pricing power do not pre-announce price increases a year ahead.
2.4 🚩 Where the moat is genuinely exposed
Not to competition — to concentration.
| customer | 2024 | 2026 |
|---|---|---|
| Nvidia | 12% | 19% |
| Apple | 22% | 17% |
| Top two combined | 34% | 36% |
And at the packaging layer it is far more extreme: Nvidia has secured roughly 60% of TSMC's entire CoWoS output for 2026, and the top three customers take an estimated >85%.
TSMC's growth is now levered to one customer's AI capital-expenditure plan. Nvidia overtaking Apple is usually reported as a milestone; read as a risk disclosure it says the marginal dollar of TSMC's growth depends on the durability of a single buyer's order book.
🚩 And the shortage that creates the pricing power is ending. CoWoS capacity has grown
~80%/yr — ~35K wafers/month (end-2024) → ~75K (end-2025) → a target of 125–130K by end-2026 —
and the supply-demand gap is forecast to narrow from 20% to 10% by end-2026. TSMC is
deliberately building away its own scarcity. Related: pattern-shortage-pricing-masks-share-loss
— the inverse case here, since TSMC is gaining share and holding shortage pricing, but the
pricing half of that pair has a visible expiry.
2.5 Evergreen assessment
9/10 as a business — the highest rating this framework has issued. Leading-edge logic fabrication is a natural monopoly: the capital cost of a competing N2 fab exceeds what any private balance sheet will fund, the yield-learning curve compounds with volume, and the customer list has nowhere else to go at scale.
But the corporate rating is not the position rating. See §4.3.
3. Valuation
3.1 🔴 The forward multiple — the number that changes the verdict
| source | 2026 EPS | implied fwd P/E at $417.17 |
|---|---|---|
| Consensus (2026E) | $15.91 (+49.3% vs 2025) | 26.2x |
| GuruFocus (18 Jul 2026) | — | 25.3x |
| Other vendors | — | 24.5x |
Yahoo / fin.py |
~$21.60 implied | 19.31x ❌ |
Yahoo's forward EPS is roughly 36% too high, and every conclusion keyed to it is wrong. The company's own reported results settle it: Q1 2026 EPS was NT$22.08 (≈$3.49/ADR) and Q2 was US$4.31/ADR, so H1 2026 delivered ~$7.80. A full year near $16 is consistent; a full year near $21.60 is not.
This is the second time in one day that a vendor forward-EPS figure has been the wrong fiscal year — the QCOM analysis found the same failure. Captured as a note in §5.
3.2 ⚠️ Free cash flow is the wrong anchor here
| basis | value |
|---|---|
| FCF/ADR (FY2025) | $6.06 |
| P/FCF | ~69x |
Do not read that as expensive. Capex is 56% of operating cash flow and 33.6% of revenue, and it is growth capex deployed at a 26.3% ROIC — TSMC is converting cash into future earnings on purpose. Free cash flow understates earning power for as long as the build continues, and capex guidance was just raised from $52–56B to $60–64B, with $100B more committed to Arizona (US total now $265B).
The corollary matters for this investor: TSM pays you in earnings growth, not cash. 0.9% yield, no buybacks, and free cash flow consumed by expansion.
3.3 Dividend-grower overlay (the screen missed this — TSM is a dividend payer)
Dividends per ADR, USD, from actual payment history:
| year | dividend/ADR | note |
|---|---|---|
| 2015 | $0.728 | |
| 2020 | $1.706 | |
| 2021 | $1.892 | |
| 2022 | $1.856 | ⚠️ down |
| 2023 | $1.847 | ⚠️ down again |
| 2024 | $2.340 | |
| 2025 | $3.121 | |
| TTM | $3.568 | |
| Run-rate | $3.82 | last two quarters × 2 |
| metric | value |
|---|---|
| 10yr dividend CAGR (2015→2025) | +15.7% |
| 5yr CAGR (2020→2025) | +12.8% |
| Last 2 years | +30%/yr |
| Current yield (run-rate) | 0.92% |
| Earnings payout ratio | 27.6% |
| FCF payout ratio 2025 | 47.0% |
| FCF payout ratio 2023 (trough) | 101.8% ⚠️ |
🚩 The finding that matters for an income-oriented investor: the USD dividend went nowhere for three years. $1.892 (2021) → $1.856 → $1.847 (2023) — falling in dollars while rising in Taiwan dollars, because the TWD weakened. A US holder's income from TSM is not a smooth grower; it is a TWD dividend with an unhedged currency overlay on top. The 15.7% ten-year CAGR is real, and so are the three flat years inside it.
⚠️ In the 2023 trough the dividend consumed 102% of free cash flow. Covered comfortably by net cash and by earnings (27.6% payout), but it shows FCF coverage is cyclical — use the earnings payout for this name.
3.4 The four classical models — and why three of them say "expensive"
| model | fair value | applicability |
|---|---|---|
| Forward earnings | $350–480 | ✅ Primary |
| Bogle expected return | see below | ✅ Secondary |
| Dividend Yield Theory | $246 | ⚠️ Report, heavily downweight |
| DDM (two-stage) | $175 | ⚠️ Report, heavily downweight |
| Graham IV | $95 | ❌ Void for this business type |
Dividend Yield Theory: current yield 0.92% against a 1.55% five-year average → $3.82 ÷ 0.0155 = $246, i.e. 41% below spot. DYT says clearly expensive.
DDM, two-stage (12% growth 5yr → 6% terminal, 9% discount): ~$175.
Graham IV: √(22.5 × 11.64 × 32.71) = $95.
Why all three are the wrong tools — and why that is not a free pass. Each keys off the dividend or the book value, and TSMC deliberately pays out only 27.6% of earnings while reinvesting 56% of operating cash flow at a 26.3% return. Every retained dollar compounds at 26%; these models penalise precisely the behaviour that creates the value. Graham's 22.5 ceiling (15 P/E × 1.5 P/B) mathematically cannot pass a 35%-ROE business.
But three of four models pointing the same way is information, not noise. It says the buyer at $417 is relying entirely on the earnings-growth case, with no support from asset value, yield, or cash return. That is the honest framing of the risk.
3.5 Fair value
| scenario | basis | value |
|---|---|---|
| Bear | 20x 2026E $15.91 — AI digestion, CoWoS gap closes, multiple compresses | $318 |
| Base | 22–26x 2026E | $350–414 |
| Bull | 26x 2027E (~$20 at +26% growth) | $520 |
Fair value: $350–480, central ~$415. Spot $417.17 is at fair value.
Bogle expected return: 0.92% yield + ~17% sustainable earnings growth = ~18%/yr at a flat multiple; ~12%/yr if the multiple drifts from 26.2x to 22x over three years; ~7%/yr if it compresses to 18x. Note the 2026 consensus growth of +49% is a peak-cycle rate and should not be extrapolated.
3.6 Sentiment — uniformly, almost suspiciously, positive
| firm | rating | target | action |
|---|---|---|---|
| Barclays | Overweight | $650 | raised from $625 |
| Susquehanna | Positive | $600 | raised from $575 |
| B of A | Buy | $590 | raised from $490 |
| Needham | Buy | $530 | raised from $480 |
| DA Davidson | Buy | $500 | raised from $450 |
| TD Cowen | Hold | $440 | raised from $400 |
| Bernstein | Outperform | $330 | (Dec 2025, stale) |
| Freedom Broker | Buy | $210 | (Apr 2026, outlier) |
Mean target $540.20 (+29.5%). Every single post-Q2 action was a target raise. There has not been a downgrade in eight months, and the most cautious live rating is a Hold with a target 5.5% above spot.
⚠️ Treat unanimity as a risk flag, not a confirmation. When no covering analyst is willing to be negative, the marginal buyer has already bought. This is the "0 sells" condition the watchlist flagged on MU. It does not make the bull case wrong; it means the surprise is asymmetric to the downside.
4. Debate Round and Verdict
4.1 The tension
Fundamentals + Moat say: 72.3% share and rising, ROIC 21.7% → 26.3%, gross margin to 67.7%, N2 with 4x N3's tape-outs, both challengers with zero committed external customers, net cash of $54B, and pre-announced 2027 price increases. This is a 9.5/10 business and it is improving.
Valuation says: 26.2x forward — not 19.3x — plus DYT at $246, DDM at $175, Graham at $95, and a P/FCF near 69x. Spot is at fair value with no margin of safety.
Sentiment says: mean target +29.5%, zero downgrades in eight months, every post-print action a raise.
4.2 Resolution
Per framework §5, business quality outweighs static asset-based models for a high-ROIC compounder — so Graham, DDM and DYT are correctly downweighted. But "the models are wrong" is not the same as "the stock is cheap." On the one model that fits — forward earnings — the stock is at 26.2x against a fair range of 22–26x. That is fair value, reached after the correct multiple replaced the wrong one.
The framework's own first principle decides this: great company + fairly priced = wait, not buy. This is the "Great / Expensive → wait, watch" cell of the §0 table, not the "Great / Cheap → value" cell that the morning screen placed it in.
4.3 🚩 The risk that is not in any model
Taiwan. Every number in this report assumes the fabs keep operating. This is not a drawdown risk that a fair price compensates — it is a binary that takes the position to approximately zero, and no entry price fixes it. The Arizona build ($265B committed) reduces it over a decade; it does not address it now, because the leading edge and essentially all advanced packaging remain on the island.
The corollary for sizing: the margin of safety on TSM can only ever be business quality, not price. That is an argument for a smaller position, not for paying any price.
4.4 Verdict
WATCH — conviction 6.5
The best business on this watchlist at a price that already reflects it. Buy the drawdown, not the headline. At $360 this is a [7.5]; at $320 it is an [8.0]. At $417 it is a hold-your-fire.
| Fair value | $350–480 (central ~$415) |
| Entry | ≤ $360 (≈22.6x 2026E) |
| Strong buy | < $320 (≈20x 2026E) |
| Trim | 32x fwd |
| Moat rating | 9/10 — highest issued |
| Fundamentals | 9.5/10 |
| Conviction | 6.5 — capped by price and jurisdiction, not by business |
Break triggers: - Intel 14A lands a tier-1 external customer at volume (decisions land H2 2026 – H1 2027) - N2 gross-margin dilution exceeds the guided 3–4 points - Nvidia revenue concentration passes ~25%, or Nvidia publicly cuts its CoWoS allocation - CoWoS supply-demand gap closes below 10% — the pricing-power tailwind ends - Samsung SF2P converts its 70% yield into a tier-1 design win (share above ~10%) - Any hard change in Taiwan Strait posture - 2027 price increases are walked back or not implemented
Upgrade triggers: price ≤$360 with the thesis intact · Intel formally exits external foundry · 2027 price increases confirmed in guidance.
A portfolio-specific passage was removed from the public build.