---
ticker: "TSM"
company: "Taiwan Semiconductor Manufacturing Company Limited"
date: "2026-10-15"
view: "Hold"
fair_value: 372.61
price: 421.14
price_source: "last trade"
price_time: "2026-10-15T19:01:47Z"
price_vs_fair_value: "+13.0%"
bear: 185.79
bull: 624.16
exit_price: 620.54
restated: "2026-10-26"
method: "case-spread-2026-10-24"
original_fair_value: 327
original_view: "Avoid"
page: "https://alpha.dming.si/research/2026-10-15-tsm"
---

# TSM · Taiwan Semiconductor Manufacturing Company Limited · Hold

Research report as of 2026-10-15, partly unblinded. Page: https://alpha.dming.si/research/2026-10-15-tsm

We value TSM at $373 a share. The report's $421.14 price (Oct 15, 2026) was 13% above fair value: Hold.

_Restated Oct 26, 2026 (method case-spread-2026-10-24); the report's original value was $327, Avoid._

| Lens | Range | Marks |
| --- | --- | --- |
| Our report | $186 to $624 | fair $373; original $327 |
| Street targets (7) | $500 to $700 | average $586, median $560 |
| Price | | $421.14 |

Bands run from our bear to bull (the hollow mark is the original $327) and the Street's low to high. Street targets forecast future prices, not value today. Market data, captured Oct 16, 2026.

| Figure | Value |
| --- | --- |
| Fair value | $373 |
| Price | $421.14 (market data, Oct 15, 12:01 PM PT) |
| Five-year return | 9.4% (a year, at the report's price) |
| Bear / bull | $186 / $624 |
| Exit price | $621 (on Dec 31, 2031) |
| Confidence | 30% |

_This text argues the original call. The current figures are at the top._

**Original call, Oct 15, 2026 (before the Oct 26, 2026 restatement)**

_As of 2026-10-15 · partly unblinded. We committed and pushed the model () at 12:01:38 PDT, before reading any TSM quote, multiple, estimate, target, headline or Alpha rank; our recall of the 2024-25 ADR range, $130-300, was not blinded (see assumptions.md). Price first read 12:01:48 PDT. Includes 3Q26 results, released by TSMC today (6-K filed 2026-10-15 06:40 UTC). Revised after one review round and three rule corrections (section 9): blind $313, now $327; a sourced beta took $18 off and a multiple judged from TSMC's own history added $31. All values US dollars per ADR (1 ADR = 5 common shares). NT$ items translated at USD/TWD 32.358 (market data FX_IDC:USDTWD, 2026-10-15 18:51 UTC)._

- **Fair value today:** $327 per ADR (cost of equity 12.34%, the yearly return investors require, including a 1.5% Taiwan premium); $328 weighting bear/central/bull 25/50/25 (our weights)
- **Price:** $421.14, last trade, 2026-10-15 19:01:47 UTC → downside 22%
- **Five-year exit price:** $545 at 2031-12-31 (median of three estimates: judged exit multiple 15x $545, long fade $426, outside view $953)
- **Five-year return:** 6.8% a year from $421.14, including dividends before the 21% Taiwan withholding
- **Purchase price:** $332 to earn the 12% hurdle (the yearly return we require). Since 2026-10-22 it no longer drives the call: we buy at or below the central fair value, with no second hurdle. It sits $5 above fair value because the hurdle is below the 12.34% cost of equity
- **Bear / bull today:** $164 / $494, each a coherent operating case
- **Confidence:** 30% (subjective, not calibrated): chance the 2031 value lands within ±20% of $545
- **Moat:** The only foundry at the leading edge at volume (N3, N2, A14 next), with CoWoS (advanced chip packaging) at 5x reticle that no one else ships; 36-47% return on new capital in 2016-20 and 2020-25 (28-36% with margins held at each window's start). Cost: Taiwan location and a capex wave
- **Market expects:** On our earnings, 20.4x 2032 NOPAT (after-tax operating profit) at exit (ours 15x, TSMC's own 2015-24 median 16.3x); or at our 15x, 2032 NOPAT of $224B (ours $165B, bull $227B); or our central case at a 9.37% cost of equity
- **Key driver:** Revenue growth 2027-32: ±3 points a year is −$38 / +$43. Next, the judged exit multiple: each turn (1x) is about $17 (14x $310, 16x $344). The Taiwan premium, charged once in the discount rate, costs $23 (0% gives $350)
- **Alpha model:** Capfolio's stock-ranking model: rank 241 of 709 (tier Above Average, score 66.1, as of 2026-10-14; rerate required): quality B, growth B, momentum C, valuation C (#479), expectations C. Integrity CLEAN
- **Street:** market data: 20 buy, 1 hold, 0 sell; mean target $552.42 (range $440-665), updated 2026-10-15 11:17 UTC, before or around today's results. Street mean EPS per ADR (NT$, translated at 32.358): 2027 $22.09, 2028 $28.17, 2029 $33.69 (4 analysts). Targets forecast a future price, so they show expectations, not validation or value today

**We rate TSMC Avoid at $421 per ADR: our central fair value is $327, 22%
below the price.** TSMC makes almost every leading-edge chip in the world, and the AI build is
running through its fabs: 3Q26 revenue was US$46.7B, up 41% in dollars, at
a 58.7% operating margin, and management guides 4Q26 to US$52.3-54.3B. Our
central case grows revenue to US$353B by 2031 (from US$176B in 2026) at a
54% operating margin, and is worth $327 per ADR at 15x 2032 operating
profit after tax, a multiple below TSMC's own 2015-24 median (16.3x)
because growth at exit is half what it was then. The price, $421, needs
about the Street's 2028-29 earnings, which are 19-30% above ours (ours are
16-23% below) and match our bull case, or a 20x exit multiple on ours. The
reviewers' best alternatives (Street revenue with held margins, a lower
discount rate) reach $395 and $406, still below the price (section 9). We
do not have an edge on AI demand; the gap is the Street's 2028-29
revenue. The call is Avoid at $421, Hold near $327-380, with the reasons
it could be wrong in section 6.

**Terms.** Fair value is our estimate of what an ADR (a US-traded share
certificate for 5 TSMC shares) is worth today. We value each company three
ways. The central case is our best estimate, the bear case is a severe but
plausible bad outcome, and the bull case is the matching good outcome. The
exit price at the end of 2031 is the median of three estimates: a judged
multiple of 2032 NOPAT (net operating profit after tax) that we set from
TSMC's own history, the long fade (which lets growth and the return on new
investment drift to long-run levels), and the outside view (the multiple
that peers trade at). The Taiwan premium is the extra 1.5% we add to the
cost of equity for Taiwan-specific risk. WACC is the weighted average cost
of capital, the rate at which the whole business's cash flows are
discounted. N2, N3 and A14 are TSMC's chip manufacturing generations, with N2
the newest in production. The Street means the analysts who follow the stock.
Two reviewers checked the report: a model auditor, who looks for errors, and a
Buy advocate, who argues for a higher value (section 9).

## 1. TSMC is a contract manufacturer with no second source at the leading edge

TSMC is a contract manufacturer. Chip designers (Nvidia, Apple, AMD,
Broadcom, Qualcomm, hyperscalers' own chips) pay per wafer, and the price of
a wafer rises with each node: 3Q26 wafer revenue was 8% N2, 33% N3, 28% N5,
10% N7, so 79% came from 7nm and below (6-K 2026-10-15). By platform, HPC
(AI accelerators, server CPUs, networking) was 68%, smartphones 21%.
Customers pay because there is no second source at the leading edge:
Samsung and Intel trail by a node or more (management claim, 3Q26 call;
our inference from the share of 7nm-and-below revenue). Advanced packaging
(CoWoS) is the other bottleneck TSMC owns: "We are in the real-world
production of 5x reticle size" (C.C. Wei, 3Q26 call).

What limits it is capital and geography. Machinery is depreciated over five
years and buildings over 10-20 (20-F Note 4); capex was 31-54% of revenue in
2016-2025 and is US$60-64B in 2026, raised from US$52-56B in July. Customers
now want capacity outside Taiwan: Arizona fab 2 starts volume production in
3Q27, Japan's second fab moves to N3 in 2028, and management expects
overseas fabs to cut gross margin by "closer to 2%" now and 3-4 points "in
the latter part of the five-year period" (3Q26 call). Customers also prepay
for capacity: NT$234B of "temporary receipts" at 2026-06-30, refundable or
offset against future sales (2Q26 statements Note 20). We treat them as a
claim in the bridge.

Per-ADR cash flow is simple: almost no stock pay (NT$0.2B in 1H26), no
buybacks, a quarterly dividend that management promises never to cut (NT$24
a share paid in 2026; NT$7 a quarter now), and a cash pile of NT$4.0T
(US$124B) at 2026-09-30 against NT$1.03T of debt.

## 2. Recent results: TSMC has met or beaten the top of its revenue guide

| Quarter | Guide (revenue, US$B) | Actual | Gross margin guide / actual | Operating margin guide / actual |
| --- | --- | --- | --- | --- |
| 1Q26 | 34.6-35.8 | 35.90 | 63-65% / 66.2% | 54-56% / 58.1% |
| 2Q26 | (high end reached, per call) | 40.20 | n/a / 67.7% | n/a / 60.3% |
| 3Q26 | 44.6-45.8 | 46.72 | 65-67% / 66.4% | 56-58% / 58.7% |
| 4Q26 | 52.3-54.3 | due mid-January 2027 | 65-67% | 56.5-58.5% |

Sources: 6-Ks of 2026-01-15, 2026-04-16, 2026-07-16 and 2026-10-15
(reported facts and management guidance). 3Q26 beat the top of the revenue
guide by US$0.9B; gross margin fell 1.3 points from 2Q on the N2 ramp. Full
year 2026 is guided "around 44%" in dollars; our 2026 is US$176.1B (+43.9%),
using the 4Q midpoint. 2Q26 net income included a NT$63.2B gain on selling
part of the Vanguard (VIS) stake (2Q26 statements, cash-flow statement), so
the 2Q net margin of 55.6% overstates the run rate. The valuation uses
operating profit only, and since the review 2026 net income (which sets
2027 dividends and the 2026 EPS of $17.01) excludes the gain too.

## 3. Our forecast and valuation: $327 at a judged 15x exit multiple

Model: `model.py`, output in `output.txt`, assumptions
and sources in `assumptions.md`; returns and the reverse
valuation (working back from the price) in `unblinded.py` and
`unblinded-output.txt`.

| Case | 2031 revenue (US$B) | 2031 op margin | 2032 NOPAT | Multiple / fade / outside (2031 exit) | Exit price | Fair value today | Five-year return |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Bear | 235.9 | 47.8% | 95.9 | 11x $259 / $228 / $567 | $259 | $164 | −7.2% |
| Central | 353.4 | 53.7% | 164.6 | 15x $545 / $426 / $953 | $545 | $327 | +6.8% |
| Bull | 443.0 | 58.0% | 227.1 | 17.5x $846 / $643 / $1,303 | $846 | $494 | +15.9% |

The central case (our assumption, labelled in `assumptions.md`): revenue
+30% in 2027 (about 3.5% a quarter from the 4Q26 guide), then +16%, +12%,
+10%, +8%; 2029 revenue of US$297B is a 27% CAGR from 2024, above
management's January framework of "approach 25%", because 2026 alone beat
that framework by 15 points. Margin before depreciation goes from 72.8% in
2026 to 70%: overseas dilution and N2/A14 ramps against pricing. Capex is
derived, not chosen: year-end plant is a set multiple of next year's
revenue (0.67 at end-2026, rising to 0.80 as overseas fabs need more capital
per wafer; it was 0.86-1.27 in 2016-24), and depreciation is 19% of opening
plant plus 5% of the year's capex, which fits 2025 within 0.3%. Central
capex is US$74-81B a year in 2027-32; bear US$40-64B; bull US$96-107B.

**Capex is charged once, in the margin.** (This is our check against
counting the capex wave twice, a mistake made in an earlier Meta report.)
Depreciation rises from
US$25B (2026) to US$58B (2031) in the central case, which is what takes the
operating margin from 58.6% to 53.7%; the capex wave is charged there, in
the margin. The exit multiple and the fade then price the business that
results, at a mature return on new capital taken from TSMC's own history,
not from inside the build window: rise in NOPAT over capex less
depreciation was 36% over 2016-2020 (one build cycle with five-year tools)
and 47% over 2020-2025, against 38.6% on our own 2026-32 path. Those
history figures include margin expansion; with each window's starting
margin held, they are about 30% and 35% (the model auditor's figures; our rough
check gives 28% and 36%). Our path figure includes margin decline, so it
understates the path's capital efficiency. The central 35% is therefore at
the top of the margin-held history, not inside it with room to spare. It
drives only the long fade, which is not the median: at 30% the value is
unchanged ($327).

**The exit multiple is judged from TSMC's own history.** The blind model
called its 14x judged, but it was the value-driver formula
(1 − g/RONIC)/(WACC − g) to two decimals (RONIC is return on new invested
capital), which our valuation method (`valuation.md`) allows only as a
cross-check. The evidence now is TSMC's own year-end enterprise value
over the next year's NOPAT, 2015-2024 (market data key-metrics, fetched
2026-10-15; next-year NOPAT from our history, so it is hindsight, not
consensus): 11.0x, 13.5x, 16.5x, 16.0x, 17.6x, 26.7x, 17.4x, 14.2x, 14.7x,
20.3x; median 16.3x, across the 2019 and 2023 downturns. TSMC at the end
of 2031, in the central case, grows about 7% (half its 2016-2025 revenue
growth of about 17% a year in NT$), still leads every node at a 54%
operating margin, and reinvests about 13% of 2032 NOPAT. We judge 15x:
below the median because growth is slower and the US 10-year is about 2.5
points above its 2015-2024 average (our recollection, not a fetched
series), inside the 14.2-17.6x band of the middle six years because the
franchise is stronger than it was. Bear 11x, the history's low (2015), for
a foundry one year past a digest; bull 17.5x, the 2019 and 2021 level, for
9% growth at a 58% margin. The formula is printed only as a cross-check:
13.2x central (9.6x bear, 15.9x bull) with the Taiwan premium in WACC,
17.2x central without it. The judged 15x implies 6.6% perpetual growth at
a 35% return, against 7.1% path growth in 2032 and 6.8% in 2033 (bear 11x
implies 5.4% against 4.9%; bull 17.5x 7.4% against 8.7%). The bear's
implied growth is above its path; at the formula's 9.6x the bear would be
$150 rather than $164. The long fade lands 12-24% below the judged
multiple in every case.

The bear's 20% mature return is above the bear path's in-window 11.2%
only because that path contains a year of idle capacity (2028); 20% is
below every TSMC build cycle on record. The bear carries three levers
(lower growth, lower margin, lower multiple), all from the same AI
downturn: at the central 15x on bear earnings it would be $205 rather than
$164.

**Taiwan risk is charged once, in the discount rate.** Cost of equity
12.34% = US 10-year 5.081% + beta 1.28 x 4.5% + 1.5% for Taiwan; WACC
12.03%. Sources: the 10-year from market data (2026-10-15); beta 1.28 is
the median of three published five-year ADR betas (market data and StockCharts
1.23-1.28, cited by the auditor, not opened by us; market data 1.38, read
2026-10-15); the 4.5% equity premium is the house convention used in every
2026-10-15 report, with no dated source (Damodaran's 2026-01-05 mature
premium is 4.23%, which would be +$5). No case contains a blockade or a
war; the bear case is an AI-demand cycle.

The premium sits in the cost of equity for 2026-31 and in the long fade's
WACC. The judged multiple is not lowered for Taiwan: it rests on TSMC's own
multiples, which carry whatever discount the market applied in 2015-2024.
So after 2031, the judged multiple (method 1) carries the market's historical
view of Taiwan risk, not ours. The premium costs $23 per ADR ($350 without it, $327 with
it), all of it in the 2026-31 discount. The alternative placement, the
premium only in a lower multiple (15x x 13.2/17.2 = 11.6x, the formula's
four turns) and none in the rate, gives $326; the placement barely
matters, but the two must not be combined.

**Three methods and their disagreement.** The three methods are the judged
multiple (method 1), the long fade (method 2) and the outside view (method 3).
The judged multiple and the long
fade sit 22% apart in the central case ($545 and $426); the fade is lower partly
because it charges our 1.5% premium in every year after 2031. The outside
view now uses forward multiples: the same six peers at the median of their
price over consensus EPS for the fiscal year ending in 2027, 28.4x (market data,
fetched 2026-10-15 19:37 UTC; ASML 30.0x, KLAC 31.8x, AMAT 23.2x, LRCX
29.4x, TXN 27.3x, ADI 24.4x), applied to TSMC's 2032 net income. It gives
$953, 75% above the median (bear +119%, bull +54%). Two inputs drive the
gap. The first is the discount rate: a 28x forward P/E is a 3.5% earnings
yield, against our 12.03% WACC. The second is maturity: these peers'
consensus EPS still grows 17-33% in fiscal 2028, while TSMC grows 7% at
exit. We believe the judged multiple there; it uses TSMC's own record.

The bridge omits TSMC's net pension liability (NT$5.9B at 2026-06-30, 2Q26
statements) and guarantee deposits (NT$0.8B): about $0.04 per ADR,
disclosed rather than changed.

**Dividends.** The blind model paid 25% of the prior year's net income,
below every year on record: dividends paid over prior-year net income were
48%, 29%, 43%, 40% and 37% in 2022-2026 (market data cash-flow statements; 2026 is
NT$24 a share, 2Q26 call). Management promises "a sustainable and steadily
increasing cash dividend per share" (3Q26 call). The model now pays 40%
of prior-year net income (the 2022-26 median), never cut and raised by at
most 33% a year (the 2026 increase, the largest recent step), which keeps
2027 near NT$32 a share against NT$24 in 2026. Effect +$1.43: the cap
binds for years, and cash kept still sits in the exit value. The
auditor's point stands in part: $76 of the central exit price (was $89) is
surplus cash earning 2.4% after tax while equity is discounted at 12.34%;
paying all of it out as it arrives would be worth up to about $9 (our
estimate, not run). We do not assume it, because management's stated
policy is steady per-share increases, not distribution of surplus.

**Sensitivities** (from `output.txt`): judged multiple 12x / 14x / 16x /
18x at a 70% margin before depreciation, $275 / $310 / $344 / $379; a 64%
/ 74% margin at 14x, $277 / $331; revenue growth ±3 points a year, $289 /
$370; capital intensity +0.15 every year, $302; TWD 10% stronger (about
−4 points of margin), $302; Taiwan premium 0% / 3%, $350 / $306; mature
return on new capital 20%→12% in the fade, no change (the fade is not the
median).

**Dividend withholding.** Taiwan withholds 21% of every cash dividend paid
to an ADR holder; the US has no tax treaty with Taiwan (20-F Item 10). A
US taxpayer in a taxable account can usually credit it against US tax
(not tax advice; holders should check). An IRA or other untaxed account
cannot, and loses 21% of the dividends: $6.22 of fair value before 2031
(21% of the $29.63 present value of 2027-31 dividends) and about 0.5% a
year after, at a 2.5% yield on the exit price. Headline values are before withholding,
as all our reports are before shareholder taxes.

## 4. What the market expects: a 20.4x exit multiple or the Street's earnings

At $421.14 the market values TSMC at US$2.18T, or US$2.10T after US$84B of
cash and investments net of debt, leases, prepayments and minorities: 19.4x
our 2027 NOPAT, 24.8x our 2026 EPS ($17.01, without the VIS gain) and
19.7x our 2027 EPS ($21.39). Working back from the price (`unblinded-output.txt`):
on our central earnings the price needs a 20.4x exit multiple, a level
TSMC's own record reached only at the ends of 2020 and 2024; at our 15x
it needs 2032 NOPAT of US$224B, 36% above our central US$165B and about
our bull US$227B; or, on our earnings and multiple, a 9.37% cost of
equity, below our rate even with no Taiwan premium (10.84%).

The Street (Street analyst estimates, translated at spot; Street estimates)
has 2027 revenue US$228B, at our central (+1%); but 2028 US$289B and 2029
US$357B with EBIT margins of 57-59%, against our US$266B and US$297B at
54-55%. Its 2029 EPS of $33.69 per ADR equals our bull case ($33.00). The
2028-29 Street figures come from five and four analysts.

The price fell 2.1% today, from $430.00 (official close 2026-10-14) to
$421.14 (19:01 UTC 2026-10-15), with SOXX down 1.8% ($519.42 to $510.01)
and QQQ down 0.4% ($728.95 to $726.06) over the same span. Headlines today
are a sector sell-off on AI-financing worries ("Chip selloff intensifies
as investors question sustainability of AI spending", Seeking Alpha) as
well as the 4Q margin guide. We read the move as mostly the sector, and
did not measure day 5 or day 20.

Alpha, Capfolio's stock-ranking model, ranks TSM 241st of 709 (as of 2026-10-14), Above
Average: good quality and growth, middling valuation (price-to-sales at
the 19th percentile of peers is its top drag). It does not price country
risk at all; it is a cross-sectional ranker, so it neither confirms nor
contradicts our level.

## 5. Our view against the market's: the gap is mostly the Street's 2028-29 revenue

The market's thesis: TSMC compounds near 20% a year to 2029 with margins
near today's, and Taiwan risk is small enough to ignore. Ours: the
earnings path is a little lower than the Street's from 2028, because
overseas fabs and the capex wave raise depreciation from 14% to 16% of
revenue and growth decelerates as the AI build matures; and Taiwan risk is
worth $23 per ADR in the five-year discount. The first point is now the
bigger one: Street revenue for 2028-29 with held margins is worth $395,
our bull case $494; no premium at all, $350.

What settles it: for earnings, the 2027 capex guide and the 2027
gross-margin path (January 2027) and whether 2028 revenue tracks above
US$280B. For the premium, nothing settles it until something happens;
it is a judgment about a tail, and the report gives its price so Dong can
substitute his own.

## 6. Bear case, and why the other side could be right

The bear case ($164): the AI build digests in 2028, revenue falls 8%,
margin before depreciation drops to 64%, and capacity built in 2027 for
2028 demand sits idle. TSMC's revenue has fallen in a year before (2023,
−4.5% in NT$) while capex stayed near 44% of revenue.

Why the buyers at $421 could be right:
- **No Taiwan premium is the right premium for an investor who holds
  through.** TSMC has run without disruption for 39 years, its importance
  to the US and China is itself a deterrent (an argument, not a measured
  fact), and a growing share of new capacity is in Arizona and Japan.
  Damodaran's country premium for Taiwan is 0.78% (advocate's source,
  2026-01-05). With his rates and the blind beta the value is $348.
- **The Street's 2028-29 may be the most likely path.** Management says AI
  accelerator growth "probably will be higher" than its mid-to-high-50s
  CAGR, N2 capacity will "more than double in 2027", three new N3 fabs
  start in 1H27, 3Q27 and 1H28, and "the demand is higher than the supply"
  (3Q26 call). A14 is drawing more customer activity than N2 did, and an
  analyst on the call cited reports of price rises across nodes. Street
  revenue for 2028-29 on our margins is worth $381; on a held 72% margin,
  $395; our bull case, which matches the Street, $494.
- **Pricing.** The CEO says "there is no ceiling for TSMC's profitability";
  gross margin has held 66-68% through two quarters of N2 dilution.
- **The multiple.** TSMC traded at 20.3x next-year NOPAT at the end of
  2024 and 26.7x at the end of 2020; the price needs 20.4x on our 2032
  earnings.

Together, the advocate's discount rate and Street revenue give $406,
still below the price; the bull case ($494), Street revenue with bull
margins, is above it. The flip is the bull's earnings path: 2027 guided
at +35% or more and 2028 tracking US$289-299B with margins held near 58%.

The assumption that matters most is revenue growth after 2027 (±3
points a year is −$38 / +$43). The second is the judged exit multiple
(about $17 a turn).

## 7. Dates to watch, and what would prove us wrong

Dates to watch:
- 2026-11-10 (approximate; TSMC files monthly revenue around the 10th):
  October revenue against a 4Q run rate of about US$17.8B a month.
- 2026-12-10: ADS ex-dividend for NT$7.0 (paid 2027-01-07).
- About 2027-01-14 (expected, not confirmed by TSMC): 4Q26 results, 2027
  revenue outlook and 2027 capex budget. Our 2027: revenue US$229B,
  capex about US$76B, operating margin 57.6%.

Forecasts in `record.json` for 4Q26 (graded on the results day): revenue
US$54.3B (range 52.8-55.8), gross margin 66.5% (65.5-67.5%), operating
margin 58.0% (56.8-59.3%). Revenue sits at the top of the guide because
TSMC met or beat the top in each of the last three quarters; the margins
sit half a point above the guide's midpoints because the guide assumes
NT$31.8 per dollar and the TWD is 1.7% weaker today (about +0.7 points on
our 4-points-per-10% rule). The model itself uses the guide midpoint
(US$53.3B); a US$1B higher base lifts every later year by 0.6%, about $2
of fair value (our estimate; not run).

Falsifiers:
- 2027 revenue guided at or above +35% in dollars with a gross margin at or
  above 64%: our central earnings are too low; move toward the bull case
  ($494), and with 2028 tracking US$289B, to $381-395 on our discount rate.
- 2027 capex guided above US$90B without a matching revenue guide (above
  +35%): capital intensity is rising faster than modelled; toward $302
  (intensity +0.15).
- Any quarterly revenue decline in 2027 or 2028, or gross margin below 60%:
  bear case.
- A change in cross-strait risk (sanctions, blockade drills that halt
  shipping, or a durable agreement) moves the premium: $350 at 0%, $306
  at 3%.

## 8. Problems with our tools for a foreign filer

Our tools were built for US companies that file 10-Ks and 10-Qs. TSM is our
first foreign filer (it files 20-F annual reports and 6-K interim reports). Each item was checked by this analyst
unless marked otherwise.

1. **market data statements mix currencies and share units.** `reportedCurrency` is
   TWD, but EPS is per ADR in NT$ (2025: 333.05 = 5 x 66.61) and share
   counts are in ADR units (5.19B). The EPS matches neither the company's
   TIFRS EPS (NT$66.25) nor the 20-F's IFRS EPS (NT$65.47). Confirmed
   against the 6-Ks and the 20-F.
2. **market data figures are wrong for TSM in places.** 2025 revenue NT$3,848.5B
   against 3,809.05B (20-F); 4Q25 revenue 1,056.0 against 1,046.09 (6-K);
   1Q26 operating income 657.3 against 658.97; quarterly D&A is erratic
   (2Q25 188 against about 165 in the quarters around it). The vendor's 2025 annual
   operating income (1,955.7) is 1% above the 20-F's 1,936.1. Earlier years
   (2023-24) match. Not used for 2025-26.
3. **Street analyst estimates are NT$ with per-ADR EPS in NT$.** A US-dollar
   comparison needs an exchange rate the estimates do not carry; at spot,
   Street 2026 revenue is US$167.8B against a company guide that implies
   about US$176B.
4. **The SEC second source skips foreign filers.**
   `src/alpha/data/sec_statements/filer_class.py` classifies 20-F, 40-F and
   6-K filers as `FOREIGN_OR_CANADIAN`, and the statement fetch and the
   income-statement reconciliation accept only 10-Q and 10-K forms. So
   TSM's market data fundamentals, including the errors in items 1-2, reach
   Alpha, Capfolio's stock-ranking model, unchecked. Read in code; the effect on TSM's rank was not
   measured. (The model's valuation features do convert currency through
   an FX table, `src/alpha/services/model/features/valuation.py`; that part
   handles foreign filers.)
5. **The record has no currency, ADR ratio or exchange rate.** `record.json`
   (loaded by `record.py`) holds bare numbers. This record gives US dollars
   per ADR by convention and states the rate only in prose. It cannot
   express: the reporting currency (NT$), the ADR ratio (5), the exchange
   rate and its date, or that dividends are withheld at 21%.
6. **Two accounting standards for one company.** Quarterly 6-Ks are TIFRS
   (Taiwan's IFRS), the 20-F is IFRS as issued by the IASB; they differ on
   when the 5% surtax on undistributed earnings is booked (2025 tax rate
   16% against 17.0%; EPS NT$66.25 against 65.47). Our analyst
   instructions and research-eval checks assume one set of statements.
7. **Our analyst instructions, research-eval checks and review procedure
   name 10-Q and 10-K.** The analyst skill's step 1 ("the latest 10-Q or 10-K
   notes"), research-eval check S2 ("checked against the latest 10-Q or
   10-K") and `review.md` all do. A 20-F filer's
   quarterly notes arrive in a separate 6-K weeks after the release (2Q26
   statements on 2026-08-14); there are no 3Q26 notes yet, so the bridge
   uses 2026-06-30 notes for leases, prepayments and minorities.
8. **No convention for country risk, ADR ratio or withholding.**
   `valuation.md` builds the cost of equity from the US 10-year and a beta;
   it says nothing about country risk, depositary ratios or dividend
   withholding. This report added a 1.5% premium and charged it once.
9. **Street fiscal labels again.** the vendor's "next year" EPS (17.07)
   and revenue (US$171.8B) are fiscal 2026, the current year, as with Broadcom (AVGO).

## 9. Review and revisions: blind $313 became $327

One review round (a model auditor and a Buy advocate, started by the
research coordinator), then three corrections from the coordinator that
enforce rules in `valuation.md` the draft broke (L1-L3). Finding IDs below: A
is the auditor's, B the advocate's, AQ the auditor's dividend question and L
the coordinator's rule corrections. Every change is in `model.py` and
priced in the "Review walk" and "Reviewer alternatives" blocks of
`output.txt`; the walk applies each change on top of the one before. The
blind central value was $313.34; the revised one is $326.89.

| # | Finding | Disposition | Effect on central fair value |
| --- | --- | --- | --- |
| A1a | The 1.5% premium sat mostly in the formula exit multiple; the prose ("about 7% over five years") described only the five-year part | Fixed, then superseded by L1-L2: the premium now sits only in the discount rate (2026-31 and the fade). None $350, discount only $327, exit only $350, both $327 | $0 |
| A1b | The "judged" 14x equalled the value-driver formula to two decimals | Fixed by L1: judged from TSMC's own multiples; formula printed only as a cross-check | see L1 |
| A1c | With methods 1 and 2 both on our WACC, the median never tests the discount rate | Fixed in part by L1: the judged multiple no longer moves with our WACC | $0 |
| A2 | Beta 1.2 and ERP 4.5% have no source | Fixed beta: 1.28, median of three published five-year betas (1.23 and 1.28 cited by the auditor, 1.38 read on market data 2026-10-15). ERP 4.5% kept as the house convention used in every report today and disclosed as unsourced; Damodaran's 4.23% would add $5 | −$18.39 |
| A3 | Historical return on new capital includes margin expansion | Disclosed: about 30% and 35% margin-held; the central 35% is the top of that range. It now drives only the fade; 30% leaves the value at $327 | $0 |
| A4 | Dividend timing measured from day 318, not 287 | Fixed | −$0.21 |
| A5 | 2026 net income keeps the NT$63.2B VIS gain | Fixed: 2026 EPS $17.01, P/E 24.8x. Lower 2026 income lowers the 2027 dividend, so the sign is negative here, not the auditor's +$0.10 | −$0.03 |
| AQ | 25% payout against "steadily increasing" dividends; surplus cash earns 2.4% | Fixed on evidence: 40% of prior-year income (2022-26 median), never cut, at most +33% a year. Full distribution of surplus not assumed (up to ~$9, estimate) | +$1.43 |
| B1 | Damodaran rates (CRP 0.78%, ERP 4.23%) | Rejected. Damodaran's country premium comes from Taiwan's sovereign default spread and prices default, not a blockade or war. The advocate's own counter-evidence (prediction markets 7-29% for an invasion by 2030; Metaculus 40% for a blockade; cited by the advocate, not opened by us) implies roughly 1.6-7% a year for an invasion; at a loss of 70-100% of value, about 1.1-7% a year, so our 1.5% sits at the low end. On the revised model these rates give $348 | $0 ($348 if adopted) |
| B2 | Street revenue 2028-29 (+26%, +24%) | Rejected for the central case, kept as the bull. TSMC's only multi-year figure, "approach 25%" CAGR 2024-29 (January 2026), puts 2029 near US$275B and was not raised on the 2Q26 or 3Q26 calls, though the CEO called AI demand "stronger and stronger" and the AI part "probably will be higher". Our central US$297B is already above it. On our margins, Street revenue gives $381 | $0 ($381 if adopted) |
| B3 | Margin fade to 70% before depreciation is ours; 3-10% price rises reported | Rejected: the CFO guided overseas dilution widening from about 2 to 3-4 points and 4Q gross margin down 40bp on N2 (3Q26 call); the price rises are press reports TSMC has not confirmed. Held margin with Street revenue: $395 | $0 |
| B4 | Payout 70% (+$8), cash yield (+$1), operating cash (+$2) | Payout: see AQ (40%, on the record). Cash yield 3% kept: 2025 interest income NT$94B (market data) on NT$2.5-3.1T of cash and securities is 3-3.7%. Operating cash 8% of revenue kept as our assumption | $0 |
| — | Advocate's combined case (B1 and B2) | On the revised model, $406, below the price | — |
| L1 | Method 1 must be a judged multiple; the formula is only a cross-check | Fixed: 15x central, 11x bear, 17.5x bull, judged from TSMC's own year-end EV over next-year NOPAT, 2015-24 (market data key-metrics, median 16.3x, range 11.0-26.7x, two downturns), lowered for half the growth and higher long rates. The formula gives 13.2x (17.2x without the premium) and is printed as a cross-check. This change moves the value toward the price; its evidence is TSMC's own record, not the price | +$30.75 |
| L2 | Charge Taiwan once and say where | Fixed: in the discount rate only (cost of equity 2026-31, WACC in the fade); the judged multiple is not lowered for Taiwan, since it rests on TSMC's own multiples. Cost $23 per ADR. The alternative (premium only in an 11.6x multiple, none in the rate) gives $326 | $0 (included in L1) |
| L3 | Method 3 used trailing P/E on depressed earnings | Fixed: the same six peers at forward P/E on fiscal-2027 consensus EPS (median 28.4x, market data, 2026-10-15), applied to TSMC's 2032 net income. Outside view $953 instead of $1,383; it is still not the median | $0.00 |

Net: $313 to $327. The two material changes went opposite ways: the
sourced beta −$18 (away from the price), the judged multiple +$31 (toward
it, on TSMC's own record). View unchanged: Avoid, with the central value
22% below the price and a 6.8% five-year return against the 12% hurdle.
Purchase price $332 sits above fair value because the 12% hurdle is
below the 12.34% cost of equity.

Checks after revision: our automated report check
(`scripts/research-check.py`) gave six PASS and one WARN (purchase price above
fair value, explained above). Research-eval, our evaluation of the basics,
ran five coherence checks, run by this analyst: (a) the judged multiples imply 6.6% / 5.4% / 7.4% growth
against path growth in 2033 of 6.8% / 4.9% / 8.7%; the bear's 0.5-point
excess is small and stated in section 3, PASS; (b) mature returns from
TSMC's history across two build cycles, margin-held caveat disclosed,
PASS; (c) the bear's three levers are one AI downturn, explained, PASS;
(d) the outside view is 54-119% from the median and both inputs behind
the gap (discount rate, peers' faster growth) are named, PASS; (e) the
bear's 20% mature return is above its 11.2% path and explained, PASS.
Warning signs the evaluation looks for: one was raised (the bear lowers
growth, margin and multiple together; at the central 15x it is $205). The
others cleared: the material changes went both ways, the value moved 4.3%
net, three methods are printed with the multiple judged, and five more
checks cleared.

## 10. Tooling problems met while writing this report

- The file-writing tool refused `report.md` from a helper agent; written
  through a `printf` stub and Edit, as the lesson from an earlier Amazon
  (AMZN) report says.
- The pre-commit secret scanner flagged base64 images embedded in the EDGAR
  release HTML; the releases are saved as text with images removed.
- The pre-push lint refused blank lines at the end of the saved
  transcripts; stripped and amended before the push.
- The coding environment's guard refused compound shell commands and inline
  programs; every helper went into `.tmp/*.py`.
- The vendor's TSM data cost the most time: every figure had to be checked against
  the 6-K or the 20-F (section 8).
- EDGAR worked first time with `EDGAR_IDENTITY`; the vendor's Quartr
  transcripts for 4Q25, 2Q26 and 3Q26 were available the same day as the
  call and carry no price levels.
- Found late: nothing material. The 20-F's restricted-stock and treasury
  notes and the many Form 4 filings were deliberately not opened before the
  blind commit, because they carry share prices.

## 11. Sources (all read 2026-10-15)

- SEC EDGAR, CIK 1046179: 6-K 0001046179-26-000726 (3Q26 release and
  slides), -000451 (2Q26), -000199 (1Q26), -000008 (4Q25), -000541 (2Q26
  reviewed statements), -000274 and -000536 (board resolutions,
  dividends); 20-F 0001628280-26-025362 (FY2025, filed 2026-04-16).
- Quartr call transcripts via market data: 4Q25 (2026-01-15), 2Q26
  (2026-07-16), 3Q26 (2026-10-15); saved in `data/`.
- Market data statements and analyst estimates (`uv run alpha market-data fetch`), peer
  `ratios-ttm`; saved in `data/`. After review: TSM `key-metrics` and
  `ratios` annual (12:35 PDT) and peer `quote` and `analyst-estimates`
  (19:37 UTC), saved in `data/` and `data/peers-forward/`. The vendor's TSM
  enterprise values are in NT$; whether they use the Taiwan share price or
  the ADR was not checked.
- Market quotes and analyst ratings, Street forecasts and news,
  Market data FX and yields: `data/quotes-2026-10-15.json`,
  `data/market-inputs.txt`.
- Alpha, Capfolio's stock-ranking model: `scripts/readonly-data-run.sh uv run alpha rank TSM --raw --json
  --allow-stale-model-cache` and `alpha ticker analyze TSM --signals --json`.