TWD=X, 17 Aug 2026). Where a provider figure mixes a USD price with a TWD per-share denominator it has been recomputed by hand — see §15.All three horizons move to HOLD, from BUY / BUY / STRONG BUY. Two things drove it, and only one of them is about the market. The price rose 6.3%, from $405.27 to $430.97. And the valuation pillar was found to have been scored on a mixed basis — a forward next-twelve-months P/E divided by a warranted multiple built on trailing earnings, which credits a year of growth twice and, on this name, flatters by about 47%. Restated consistently on trailing earnings, the warranted ratio is 1.23, not 0.76, and the clean multiple of 34.0× is above the 28× semiconductor guardrail — which forces the Expensive band, fires the Valuation Ceiling gate, and caps every horizon at HOLD. This is a framework-application correction, not new information about the company. There has been no earnings print since the last report: Q2 2026 was released on 16 July and was already in it (margins reconciled to confirm).
Everything measuring the business got better. July revenue accelerated to +44.7% year on year. The tape turned — price reclaimed the 50-day average, and Timing rose 13 points. TSMC gained its own line in the macro report's watchlist at Outperform / Outperform / Strong Outperform, upgrading Economic Alignment from a mixed Headwind read at conviction 50 to a clean Tailwind at 72. The driver held at 88. None of it matters to the signal, because a HOLD never amplifies and a Full-or-Expensive band bars a STRONG BUY outright.
TSMC is a contract chip manufacturer — a foundry. It designs almost nothing of its own; it builds the chips that Nvidia, Apple, AMD, Broadcom, Qualcomm and the hyperscalers design, on wafers, in its fabs. Its distinction is that at the leading edge it is close to the only option: it took 73% of the pure-play foundry market in Q1 2026 (Counterpoint Research) and roughly 90% of production at the most advanced nodes, where the AI accelerators are made. Building a competing 2-nanometre fab costs tens of billions of dollars and takes years, and a customer moving a finished design to another foundry faces a re-tape-out measured in quarters — which is why TSMC can raise prices and still be sold out. The shares analysed here are the New York ADR: one ADR represents five Taiwan-listed ordinary shares, the company reports in New Taiwan dollars, and the ADR price is in US dollars.
totalDebt is lease-inclusive, so it overstates borrowings — but the company is net cash by a factor of three, so the trap cannot change the answer here. Nothing close to triggering.Four of the six checks are clear and one is a standing caution. The single triggered gate is the Valuation Ceiling, and it fires on the price, not on the business. Everything the Quality pillar measures — growth, margins, balance sheet, moat — got better this quarter. What changed is that the shares are now 23% above what today's earnings warrant on a disciplined growth assumption, and above the line the framework treats as "rich for this industry". That is a statement about entry price. It is not a statement that anything is wrong with TSMC.
Lifecycle & sector classification. Information Technology / Semiconductors — advanced-node foundry, classified Growth (revenue +36.0% year on year in Q2 2026, comfortably profitable). Because it is a semiconductor manufacturer the metric profile is gross margin, capacity utilisation, ROIC and cycle context, and the Valuation pillar uses a mid-cycle-aware P/E rather than a revenue multiple. It reports in New Taiwan dollars; every per-share figure below is translated to US dollars per ADR at the spot USD/TWD of 31.782.
| Sub-signal | Value | Comparison | Score | Rationale |
|---|---|---|---|---|
| Revenue trajectory | +36.0% YoY (Q2 2026) | Accelerating: Q4 25 +21.6% → Q1 26 +35.1% → Q2 26 +36.0%. July 2026 monthly revenue NT$467.58bn, +44.7% YoY; Jan–Jul +37.0% | 95 | Faster than the pure-play foundry market itself, which Counterpoint puts at +30% YoY in Q1 2026 — so TSMC is taking share while the market grows. The July release (10 Aug) is the freshest datapoint in this report and it accelerated. |
| Profitability vs peers | Gross 67.7%, operating 60.3% (Q2 2026) | TTM operating margin 56.0%; at the start of the eight-quarter series (Q3 2024) it was 47.5%. Q3 guided to gross 65.0–67.0% and operating 56.0–58.0% | 93 | A 60% operating margin on a capital-intensive manufacturing business is close to unheard of; it is the pricing power of a near-monopoly at the leading edge. Note the guide is a step down on both lines — operating margin from 60.3% to 56.0–58.0%, a 2.3–4.3 point give-back that the company itself is forecasting. That matters for the mid-cycle question in §4. |
| Cash generation | FCF US$23.0–35.2bn TTM | Against net income of US$70.5bn — conversion of 33–50%. Q2 alone: operating cash flow NT$783bn, capex NT$496bn | 62 | The weakest leg, and honestly so. Capex is budgeted at US$60–64bn for 2026 and TSMC has committed roughly US$265bn to the United States. That is demand-backed spending, but cash going into concrete is cash the shareholder does not get. The two providers disagree on the TTM figure — Yahoo US$23.0bn, FMP's operating-cash-flow-minus-capex US$35.2bn — so the range is quoted rather than a false precision. |
| Balance sheet health | Net cash US$77.1bn | Debt/equity 0.15; interest coverage 886× (recomputed); current ratio 2.46 | 96 | Cash and marketable securities were NT$3,518.0bn = US$110.7bn, confirmed on the Q2 call as "NT$3.5tn or US$110bn"; the unrounded figure is quoted because the rounded one does not reproduce net cash (110.7 − 33.6 = 77.1). The company could fund a full year of its record capex from the balance sheet. |
Gross margin 67.7% (threshold for the top band: >55%) · advanced-node utilisation effectively above 90% — N3 and N5 are sold out, which is why TSMC is able to plan price rises of as much as 10% for 2027.
Rating: STRONG — both legs in the top band. Benchmark score: 94/100.
Context: the framework's top band starts at gross margin >55% with utilisation >80%. TSMC clears both by a distance. The one caveat is that 67.7% is an all-time high and the company has guided the next quarter to 65–67%.
Moat score: 87/100 (average of the five). Up from 82 last report — not because the walls got taller, but because this run measured share and rival revenue directly instead of assuming.
The moat scores above are derived from this, not asserted alongside it. The headline: TSMC is gaining share, and the credible attack has a date on it — Intel's 14A 0.9 process design kit, targeted for external customers in October 2026. That is the decision point; 14A high-volume manufacturing is 2028, so it is not the share point.
| Rival | Threat type | Share trajectory | Moat-erosion vector |
|---|---|---|---|
| Intel Foundry | Direct merchant rival, state-backed | TSMC gaining — Q1 2026 foundry revenue US$35.9bn vs Intel Foundry US$5.4bn | The real one, and the dates matter. Intel has shipped the 0.5 process design kit for 14A and targets the mature 0.9 PDK for external customers in October 2026 — that is the point at which customers finalise volume and design commitments. But 14A risk production is H2 2027 at the earliest and high-volume manufacturing 2028 (Lip-Bu Tan pulled the schedule forward in July 2026 from a prior 2028/2029 guide). So the decision is weeks away; the share is years away. 18A is ramping now and 18A-P has drawn external inquiries after Panther Lake. Industry commentary (SemiWiki) holds Intel's ~10% share could double or triple if 14A lands cleanly, and that if it slips as 18A did, TSMC returns to 90%+ at the leading edge. |
| Samsung Foundry | Direct merchant rival | TSMC gaining — TSMC took 73% of the pure-play market in Q1 2026 (Counterpoint) | SF2 is in volume production at 2nm-class, so the technology gap is narrower than the share gap. Samsung's constraint is yield and ecosystem breadth, not lithography. Price is its lever, which pressures TSMC's pricing power at the margin rather than its volume. |
| Hyperscaler in-house silicon | Vertical substitution | Neutral — it does not leave TSMC | Google's TPUs, Amazon's Trainium, Microsoft's Maia and Broadcom's custom ASICs all displace Nvidia, not TSMC: they are fabricated at TSMC. This is the most misread threat in the sector. It changes TSMC's customer mix and could compress packaging pricing over time; it does not remove the wafer. |
| SMIC / Hua Hong (+ export controls) | Low-cost entrant, regulatory | TSMC losing — but only at mature nodes in China | US export controls stop TSMC expanding its Nanjing fab (16/28nm), which pushes that demand to Chinese domestic foundries. Small revenue share, strategically annoying, no leading-edge implication. |
Net effect on the moat: Switching Costs held at 85 rather than 90 (Intel 18A-P and Samsung SF2 are real, qualified alternatives at the margin); Cost Advantage held at 92 on the 6.6× revenue gap to the nearest merchant rival; Pricing Power raised to 95 on the realised margin expansion. Competitive threat level: MODERATE. Share trajectory: GAINING. It propagates to the §11 Bear (an Intel-14A design-win leg, correctly scoped — realised share loss cannot happen inside twelve months when 14A high-volume manufacturing is 2028) and to the §12 thesis-invalidation rule.
| Component | Value | Score | Note |
|---|---|---|---|
| ROIC (40%) | 52.5% | 95 | NOPAT of US$65.8bn (TTM operating income taxed at the 16.16% effective rate) over invested capital of US$125.3bn (equity US$202.4bn + debt US$33.6bn − cash US$110.7bn). Top decile against any semiconductor peer, and rising. |
| Capital allocation (30%) | Disciplined | 85 | Capex goes into capacity that is already sold; the dividend has been raised repeatedly; buybacks are minimal and the share count is flat. Marked down from higher only because the ~US$265bn US build is, at least initially, lower-margin capacity chosen partly for political reasons — and the CFO has named a construction-labour shortage in Arizona as a live execution risk. |
| Management skin in the game (30%) | Low insider ownership | 55 | Insider ownership is structurally low for a Taiwanese company with a large state-fund holder, and Section 16 insider-transaction data does not exist for a foreign private issuer. Scored neutral-to-weak with the data gap stated rather than guessed at. |
Composite: 80/100.
Quality confidence 75%. Base 80 for full metric coverage, less 5 because the two data providers disagree materially on trailing free cash flow and the figure had to be quoted as a range.
The 31 July report scored this pillar by dividing a forward next-twelve-months P/E of ~21× by a warranted multiple of 27.5×, giving a ratio of 0.76 and an Attractive band. That is a mixed basis and it flatters. The warranted multiple is Σt=1..5((1+g)/(1+r))t + terminal — an expression whose year-1 term is E0 grown by g. It is therefore a multiple of current earnings. Putting next year's earnings in the numerator credits a year of growth twice: once in the forward EPS, and again in g. On TSMC that is worth about 47%, because consensus has trailing EPS of US$13.60 becoming US$19.95 over the next twelve months.
Restated consistently, the answer is not close. On a trailing numerator the ratio is 1.23. On a genuinely consistent forward construction — deflating the warranted multiple by one year of growth, 27.70 ÷ 1.15 = 24.1× — the ratio is 0.90, not 0.76. Neither consistent basis reproduces the inherited number. This is a framework-application correction, not new information about the company.
The trailing basis is also what the rest of this book uses: NVDA (clean TTM P/E 37.5×), AVGO, GOOGL, MSFT and AMZN are all anchored on a clean trailing multiple. An earlier version of this paragraph went on to say that scoring TSMC on a forward multiple made it the outlier. That was false, and a sweep of the whole book shows it was false in the opposite direction. Of the tracked names carrying a valuation anchor, roughly thirty are still anchored on a forward or next-twelve-month multiple — among them ICE, published in this same batch at this same timestamp and on the same pinned risk-free, whose actual multiple is an explicitly forward P/E measured against a warranted multiple built the E0 way. TSMC's 31 July forward anchor was the norm here, not an aberration. That does not weaken the correction: the E0 argument above stands on its own arithmetic, and the five mega-caps named are genuinely on a clean trailing basis. It does mean the inconsistency is book-wide rather than a single stale report, and it has been escalated as such rather than quietly used here as corroboration.
| Input | Value | Source / discipline |
|---|---|---|
| Risk-free rate | 4.63% | US 10-year Treasury, FRED series DGS10 as at 13 Aug 2026, pulled directly from FRED. The framework normally reads this from the macro report's market_snapshot.UST10Y — that key does not exist in the 12 August state file, so FRED is the fallback and the deviation is disclosed rather than silent. The macro report's prose carries 4.70% and the 14 August FRED print was 4.68%; the sensitivity line below covers the whole 4.63–4.72% span, including the report-date print of 4.72% on 17 August. All are below 5%, so the guardrail lines are used unmodified. |
| Equity risk premium | 4.50% | Fixed global constant — not a per-name knob. |
| Risk add-on | +0.00% | Business Quality is 87 (≥65) so no add-on. Beta 1.258, below the 1.6 threshold that would force one. |
| Discount rate r | 9.13% | 4.63% + 4.50% + 0.00%. |
| Near-term growth g_near | 15.0% | Consensus implies a 27.4% EPS CAGR from FY2026 to FY2029 (NT$532.7 → NT$1,100.3 per ADR); 0.75× that is 20.6%, which is then capped at the Information-Technology secular cap of 15%. The framework permits up to 20% for a proven, durable >20% grower if flagged — TSMC would arguably qualify, and we have declined to take it. See the sensitivity: it does not change the band. |
| Terminal growth g_term | 3.0% | Capped at long-run nominal GDP. |
| Two-stage warranted P/E (raw) | 27.70× | Σt=1..5((1+g)/(1+r))t = 5.867, plus a discounted terminal of 21.838. |
| Warranted P/E | 27.70× | Below the 28× Semiconductors guardrail, so the cap on the warranted multiple is inactive. Only the guardrail's floor-on-actual arm bites here — a distinction worth stating precisely. |
| Actual clean P/E | 34.0× | $430.97 ÷ clean trailing EPS of US$12.68 per ADR (step 7b below). |
| Warranted ratio | 1.23 | 34.0 ÷ 27.70. That is the Full band on the ratio alone (1.20–1.40) — not Expensive. |
| Band | EXPENSIVE | Forced by the guardrail, not by the ratio. The actual 34.0× is above the 28× Semiconductors line, and step 5(b) makes a name Expensive on that alone, regardless of the warranted ratio and with no growth exception. Both facts are stated because they disagree, and the stricter one governs. |
Plainly. A business that compounds earnings at 15% for five years and 3% thereafter, discounted at 9.13%, is worth about 27.7 times its current earnings. TSMC trades at 34 times those earnings on an operating basis, and 31.7 times as reported. The gap is 23%. It is not a factor of five, as it was on Palantir — it is a discipline call at the margin, and the report says so rather than dressing it up.
Which guardrail line. TSMC's GICS sector is Information Technology, whose line is 33×; its GICS industry is Semiconductors, whose line is 28× and is explicitly a mid-cycle P/E. The more specific line governs — and, more usefully, so does the strictest line, which is the rule the project's report linter now applies. Specificity and strictness happen to agree here; strictness is the better rule because it generalises (a name matching both "banks" and "financials" should get the tighter of the two, not the more specific-sounding one). It is also the line the previous report used. On either line the clean 34.0× breaches; on the Information-Technology line it breaches by 3%, on the Semiconductors line by 21%.
Sensitivity — four ways of being generous, none of which clears the guardrail:
| Measure | Reported | Clean (operating) | Note |
|---|---|---|---|
| TTM earnings per ADR | US$13.60 (NT$432.26) | US$12.68 (NT$403.12) | Clean = trailing operating income of NT$2,493,702m taxed at the 16.16% effective rate, over 5,186.4m ADS. |
| Trailing P/E | 31.7× | 34.0× | The clean multiple is 7% worse. Both are scored, the clean one governs. |
| Net margin | 50.4% | 47.0% | On TTM revenue of NT$4,450,379m. |
| ROE | 34.9% | 32.5% | Both over the same shareholders' equity of US$202.4bn (NT$1,240.37 per ADR × 5,186.4m ADS at 31.782): reported net income US$70.54bn → 34.9%, operating-after-tax US$65.78bn → 32.5%. The providers show 40.0%, which uses an average rather than a period-end equity base — on a fast-compounding balance sheet that flatters by five points, so it is not used here. |
| PEG (clean) | — | 1.12 | 34.0 ÷ 30.4% consensus FY26→FY27 EPS growth. On the forward multiple it is 0.71 — the honest counter-argument. |
| FX convention | Spot USD/TWD 31.782 | TSMC's own release translates at the period rate: Q2 2026 diluted EPS of NT$27.25 per ordinary share = US$4.31 per ADR at its assumed USD/NTD of 31.60 (2Q26 earnings release, 16 Jul 2026) — which reconciles to the NT$136.25 per ADR used here. Translating each of the four quarters at its own period rate instead of spot gives a trailing EPS near US$13.8 and a reported P/E near 31.3×, with the clean multiple near 33.6×. Both conventions sit above the 28× line, so the band does not turn on the choice. Spot is used because the numerator is today's price. | |
| Non-operating income | 8.0% of TTM net income (6.7% of pre-tax) | NT$179,858m TTM, of which NT$103,039m is interest on the cash pile and the rest is FX and investment income. Q2 alone carried NT$95,827m — roughly three times the prior quarters, on currency. | |
TSMC is not the earnings-quality problem the mega-cap AI names are. At 8% of net income it is well under the 15% threshold that forces a restatement and nowhere near the 30% that would fire Gate 4. The decomposition was run anyway, and the report scores P/E, PEG, margin and ROE on the clean figures — which makes the valuation worse, not better. That is the point of doing it before you know the answer.
| Lens | Weight | Reading | Score |
|---|---|---|---|
| The anchor | 40% | Ratio 1.23 (Full) but guardrail-forced Expensive at 34.0× vs the 28× line | 35 |
| Sector median | 20% | Mildly favourable, on a like-for-like basis. Comparing clean to clean, which is the whole point of this section: TSMC 34.0× against NVDA 37.5×, GOOGL 36.5×, AVGO 64.8×, MU 31.0× and MSFT 27.6× — a cohort median of about 35×. One caveat on that row: MU's 31.0× is its mid-cycle-normalised multiple, not its clean trailing one (18.9×), while TSMC's 34.0× is deliberately un-normalised — so that single cell is not strictly like-for-like in either direction. The median is 35× on either MU figure, so the score is unaffected. TSMC sits just below it, not comfortably beneath it. An earlier version of this row compared TSMC's reported 31.7× against those clean figures, which is exactly the mixed-basis error this section exists to correct. | 65 |
| Own-history decile | 15% | Reconstructed from 60 months of month-end ADR prices and trailing-four-quarter EPS translated at each month's spot FX. Clean P/E five-year range 11.8× – 43.0×, median ~28×. Today's 34.0× is the 87th percentile — decile 9 on an announcement-date availability convention (TSMC releases results ~16 days after quarter end). An independent reconstruction using the SEC filing date instead, which lags by roughly a further month, puts it at the 78th percentile — decile 8. Both matter for one thing only: whether the multiple is in the top decile, which is what Do-Not-Buy Trigger 2's relative arm requires. It is not, on either convention — and deciles 7 to 9 score identically on this lens. The range is quoted rather than a false precision. But the margin is thin and should be said out loud: on the announcement-date convention the 90th-percentile bar sits at 34.46× against an actual clean multiple of 33.98×. That is a gap of 1.4% — about $6 on the ADR. The second condition of that trigger, no growth acceleration, is already conceded (forward EPS growth 30.4% against trailing 55.2%). So a roughly 1.5% rise in the share price, with nothing else changing, would put this name into a hard Do-Not-Buy on the relative arm. It is negative today; it is not comfortably negative, and the next refresh should treat it as a live watch item rather than a settled question. | 25 |
| PEG | 10% | Clean PEG 1.12; forward PEG 0.71. Not demanding on either. | 60 |
| Analyst consensus | 15% | Price sits 21–27% below consensus — equivalently, consensus is 27–37% above the price ($430.97 against $547.09 Yahoo and $589 FMP). An earlier version stated the upside figures as the discount; §12 had it right. Grades 18 buy / 7 hold / 0 sell | 82 |
Blended, those weights give roughly 49 — which would be the top of Fair. It is not permitted to land there. The framework is explicit that the relative lenses order a name within the band the anchor sets and can never push an Expensive name into Fair. So the pillar is scored at 38: inside the Expensive band, but at the very top of it, because three of the four relative lenses genuinely support the name. A score of 38 rather than 20 is the difference between "rich" and "Palantir".
Run the anchor backwards. At $430.97 and a clean trailing EPS of US$12.68, the market is implying about 20.2% earnings growth a year for five years, then 3% forever, discounted at 9.13%. On the reported EPS it implies 18.4%.
Our disciplined estimate is 15%. Analyst consensus is 27.4%.
So the price sits above what our haircut growth supports and below what consensus growth would support. That is the whole argument in one line, and it is why this is a Hold rather than anything stronger in either direction. The framework is not claiming the market is deluded about TSMC; it is declining to underwrite 27% compounding for five years, and at 15% the price is 23% too high.
| Multiple / metric | Value | Note |
|---|---|---|
| Trailing P/E (reported) | 31.7× | US$430.97 ÷ US$13.60 |
| Trailing P/E (clean) | 34.0× | The scored figure |
| Forward P/E (NTM) | 21.6× | NTM EPS US$19.95 — 4.5/12 of FY2026 plus 7.5/12 of FY2027 consensus |
| P/E on FY2026 consensus | 25.7× | FY26 EPS NT$532.66 = US$16.76 (Zacks has US$16.45 — a 2% difference, within FX noise) |
| P/E on FY2027 consensus | 19.7× | FY27 EPS NT$694.73 = US$21.86 |
| FCF yield (on EV) | 1.1–1.6% | FCF US$23.0–35.2bn over an enterprise value of US$2,158bn (market cap US$2,235bn less US$77.1bn net cash). Thin — the capex cycle is eating it. Below the 3–5% "fair" band on any provider's number. |
| Price/book | 11.0× | Computed by hand. Book value NT$1,248.46 per ADR = US$39.28. Note two per-ADR equity figures are in play and they differ by 0.65%: NT$1,248.46 (total book value, used here) and NT$1,240.37 (shareholders' equity, used as the ROE and invested-capital denominator) — the gap is consistent with non-controlling interests. Both give P/B 11.0× (11.04 against 10.97). ROIC does differ at one decimal, and the earlier version of this sentence wrongly said it did not: 52.5% on the shareholders'-equity denominator used throughout this report, 51.9% on the total-book one — a 0.54-point gap. It changes nothing that matters: the ROIC sub-score is 95 on either, nothing else in the report uses the alternative denominator, and §3, §4 and §15 all run consistently off the US$125.3bn invested capital. The split is disclosed rather than silently averaged. Yahoo returns 89.9× because it divides a USD price by a TWD book value; FMP returns 9.6× on an inconsistent price. Neither is usable. |
| Price/sales | 16.0× | Also recomputed — Yahoo's 0.50× is the same cross-currency error inverted. |
| Dividend yield | 0.81% trailing / ~1.01% forward | Four ex-dividends in the last twelve months totalled US$3.4957. The forward figure is a range depending on construction: the four most recent declarations sum to US$4.0774 (0.95%), while annualising the latest declaration gives US$4.342 (1.01%). The 11 Aug 2026 declaration of US$1.085439 (ex 10 Dec, pay 7 Jan 2027) is lower in dollars than the US$1.113621 declared in May — that is the New Taiwan dollar moving, not a cut; the ADR amount floats with FX. The dollar payout is up 78.5% over two years on a like-for-like comparison — the December ex-dividend went from US$0.608106 in 2024 to US$1.085439 in 2026. |
Net framing: the warranted multiple on today's operating earnings justifies about US$351 of the US$431 price. The optionality above is genuine, and together with the supportive relative lenses it is why the pillar sits at the top of the Expensive band at 38 rather than deep inside it. It is a tilt of a few points, not a re-rating: it does not make an Expensive core cheap.
| Source | Low | Median | Consensus / mean | High | n | Upside to consensus |
|---|---|---|---|---|---|---|
| FMP | $500 | $577 | $589 | $700 | — | +36.7% |
| Yahoo | $431.50 | $536 | $547.09 | $700 | 18 | +27.0% |
| Prior report (31 Jul) | $500 | $600 | $596 | $700 | 25 | +47.1% |
Neither panel is degenerate, so no mandatory fallback applies — but they disagree by US$42 on the mean and the prior report's figures reconcile with neither, so the more conservative Yahoo panel is used for scoring and both are shown. One detail matters more than the average: Yahoo's low target is $431.50, which is today's price. The most bearish analyst on the panel thinks the shares are worth exactly what they cost. That is a fair description of where this report lands too, and it is a more useful number than the +27% mean.
Grades: 18 buy, 7 hold, 0 sell — 72% bullish, consensus Buy across 25 firms. In the last 30 days there have been no upgrades and no downgrades: Bernstein maintained Outperform on 11 August, Needham maintained Buy on 27 July, and every other recent action is a maintain. The Street is unanimous in direction and inert in action.
FMP financial-health cross-reference: overall B+ (3/5). Return on equity 5/5 and return on assets 5/5 — it agrees with our Quality read. But price/earnings scores 2/5 and price/book 1/5. An independent framework, built differently, flags exactly the two things this pillar flags.
Valuation confidence 78%. Base 80, plus 5 for hard analyst targets, plus 5 for the grades distribution, plus 3 for the ratings snapshot — 93 — less 10 because every provider's ratio field for this ADR is cross-currency corrupt and the multiples had to be rebuilt by hand from the TWD statements, and less 5 for the FX sensitivity in that translation (a one-dollar move in USD/TWD shifts the trailing multiple by about 3%). Stated as 83% until the round-4 audit added the components up: they give 78, and §15's own "15 points of haircut" agrees. Corrected. Overall confidence is unchanged at 70, which is the Timing pillar's floor either way.
The driver. TSMC does not choose how many chips the world wants; the hyperscalers do, when they set their capital budgets. Every AI accelerator that matters — Nvidia's, AMD's, Broadcom's custom ASICs, Google's TPUs, Amazon's Trainium, Microsoft's Maia — is fabricated by TSMC. That is what makes this a volume story rather than a discount-rate story, and it is why the macro report holds the name above its own sector. The secondary driver is geopolitics: export controls and Taiwan Strait risk, which act on the driver from the downside only.
Step 2b does not apply. The commodity price-trend overlay is mandatory only where the primary driver is a commodity price. TSMC's driver is a capital-expenditure cycle, so there is no 50-day average to test. Stated explicitly so the omission cannot read as a skip.
| Horizon | Reading | Weight | Evidence & date | Score |
|---|---|---|---|---|
| Historical (12–24m) | Steeply improving | 25% | Combined 2026 capex guidance for Alphabet, Amazon, Microsoft and Meta sits around US$725bn, up 77% from roughly US$410bn in 2025. TSMC's own revenue has gone from +21.6% year-on-year in Q4 2025 to +36.0% in Q2 2026 and +44.7% in July 2026. | 90 |
| Current state | About as favourable as it gets | 50% | July 2026 revenue NT$467.58bn, +44.7% year on year and +5.6% on the month (TSMC release, 10 Aug 2026); the first seven months are +37.0%. Q3 revenue guided to US$44.6–45.8bn and full-year growth to "slightly above 40%" (Q2 call, 16 Jul 2026). Advanced nodes are sold out and price rises of up to ~10% are planned for 2027. | 92 |
| Forward (6–12m) | Strong, with the first visible cracks in visibility | 25% | Morgan Stanley raised its five-hyperscaler capex forecast to about US$805bn for 2026 and US$1.1tn for 2027; the CFO describes a "multi-year demand mega trend" and has committed a further US$100bn to Arizona (~US$265bn total US investment, 10+ fabs). Against that: sell-side commentary now flags "limited visibility on capex beyond 2027 amid greater investor demand for spending discipline", and the CFO has named an Arizona construction-labour shortage as an execution risk. Marked down for that, not for demand. | 78 |
Driver score = (90 × 0.25) + (92 × 0.50) + (78 × 0.25) = 88 — Strong Tailwind, unchanged from the 31 July report, and re-derived this run rather than carried.
The 12 August macro report cites 2026 hyperscaler capex consensus of about US$527bn; Morgan Stanley's figure is about US$805bn. The gap is basket composition, and the table above shows it: the four big cloud providers alone guide to about US$725bn for 2026, and Morgan Stanley's US$805bn adds Oracle plus leases and data-centre sites. The macro report's US$527bn is a narrower core-guidance cut again. Each is quoted with its source rather than averaged into a fake consensus; the direction is not in dispute on any of them.
Amplification needs three things: a base signal of BUY (or SELL), a driver at 65 or better, and Economic-Alignment pressure of Tailwind (or Headwind). Two of the three are satisfied and it still cannot fire, for two independent reasons:
| Test | Requirement | Actual | Result |
|---|---|---|---|
| Driver | ≥ 65 | 88 | PASS |
| Economic pressure | Tailwind | Tailwind on all three horizons (macro watchlist O / O / SO) | PASS |
| Medium horizon extra condition | driver ≥ 65 | 88 | PASS |
| Base signal | must be BUY or SELL | HOLD on all three horizons | FAIL — HOLD never amplifies |
| Valuation band | warranted ratio must be below 1.20 (Attractive or Fair) | 1.23 — the Full band; and Expensive on the guardrail | FAIL — a Full or Expensive band bars STRONG BUY on every horizon |
So the long horizon's STRONG BUY is withdrawn on the band alone, before the base signal is even consulted. Had valuation stayed where the 31 July report put it, the band test would have passed at 0.76 and the driver and pressure tests would have carried it. The amplification did not fail because the tailwind weakened — the tailwind got stronger. It failed because the valuation band moved — the basis correction in §4, with a 6.3% price rise on top — into territory the framework will not amplify.
Thesis-invalidation floor. The driver breaks if hyperscaler capital budgets are guided down — not trimmed, guided down — by more than 20% year on year, or if TSMC's monthly revenue growth decelerates below roughly 15% year on year for two consecutive months. July's +44.7% is nowhere near it, which is precisely why the bear case in §11 is a de-rating story and not an earnings story.
Driver confidence 72%. Base 70 for fresh current-state data; the driver's relationship to the share price is direct and uncontested (+2), less nothing for staleness — the newest input is eight days old.
The 12 August 2026 macro report reads the regime as "energy-shock stagflation: a supply-driven inflation impulse hitting a contracting labour market, with Fed cuts priced OUT and the live debate hike-versus-hold" — scenario weights Stagflation 40%, Soft Landing 27%, Deflationary Bust 20%, Reacceleration 13%.
This is an upgrade in the source, not just the score. The 31 July report had no watchlist line for TSM and fell back to mapping the Information-Technology sector (XLK), which gave Headwind on the short and medium horizons. The 12 August report gives TSM its own entry — Short Outperform, Medium Outperform, Long Strong Outperform — explicitly held above its sector, with the reasoning: "Hyperscaler capex consensus is ~$527bn for 2026 and TSM sits at the toll booth — this is a volume story, not a discount-rate story. Taiwan-Strait risk remains the tail we are not paid for." The sector itself is only N / N / O, so the macro analyst is deliberately separating TSMC from the rate-sensitive software complex.
Pressure: Tailwind on all three horizons (anchored on the medium, which is Outperform). Stance is therefore Trend-Following — a long position rides the economic trend rather than fighting it. Conviction 72: high, because the signal is name-specific and the long horizon is the strongest reading available, but short of the 80s because the same macro report carries a live September-hike debate at ~44% implied and a 4.63–4.72% ten-year (4.72% on the report's own date), and rising discount rates work directly against a long-duration multiple — which is the same arithmetic that pushed the warranted multiple down in §4.
Effect on the signal: none. The pressure is a Tailwind and it enables amplification, but the base signal is HOLD on every horizon and HOLD never amplifies; independently, the Full/Expensive band bars a STRONG BUY outright. Both bars are shown with their arithmetic in §5. Economic Alignment moved from a mixed Headwind/Headwind/Tailwind read at conviction 50 to a clean Tailwind at 72 — and it changed nothing, because it is a context pillar.
Source: watchlist-signal — TSM has its own line in the 12 Aug macro report's Economic Watchlist Forecast (Short O, Medium O, Long SO) · Macro report 2026-08-12
Price $430.97. The nearest meaningful support is the rising chain of daily lows built since the 29 July bottom — $372.72 (29 Jul) → $387.60 (30 Jul) → $398.80 (31 Jul) → $412.11 (4 Aug) → $418.87 (11 Aug) → $427.02 (12 Aug), with shallow interruptions on 3, 6 and 14 August. The shelf that matters for a stop is the pair of lows at $405.51 (9 Jun) and $407.99 (6 Aug), and the early-August low at $398.20 (3 Aug). The dates are given because these are levels from the tool's value-sorted array, not a chronology. A logical stop below the $405.51 shelf sits at $396, which is $34.97 away, or 2.50 ATR at an ATR of $14.00. That is the top of the "moderate" band and the bottom of "unfavourable" — you have to risk a lot to find out whether you are right. Position-risk score: 55.
Proximity: no weekly or monthly support within 3% below (the weekly cluster is far lower, at $319 and beneath), and no weekly or monthly resistance within 3% above (the next weekly level is the $479.00 all-time high, 11.1% away). So neither the +15 bonus nor the −15 penalty applies. Price is at the 81st percentile of its 52-week range of $223.70–$479.00 — near the highs, with momentum, and extended.
| To 14 Aug 2026 close | TSM | SPY | vs SPY | SOXX | vs SOXX |
|---|---|---|---|---|---|
| 1 month (14 Jul → 14 Aug) | +1.42% | +3.26% | −1.84pp | −3.08% | +4.50pp |
| 3 months (14 May → 14 Aug) | +2.07% | +3.77% | −1.70pp | +3.85% | −1.78pp |
Corrected after audit. The first version of this table read SOXX off weekly bars (bases of 17 Jul and 15 May) while reading TSM and SPY off daily bars dated 14 Jul and 14 May. SOXX fell 7% between 14 and 17 July, so the mismatched base inverted the one-month comparison — it was published as −4.06pp when the correct figure is +4.50pp. Every cell above is now on matched daily closes. The error is recorded rather than quietly repaired: it is exactly the kind of basis mismatch this report spends §4 correcting in someone else's work.
Middling, not poor — relative-strength score 48. TSMC lags the S&P on both windows, but only by 1.7–1.8 points. Against its own sector it is ahead over one month and 1.8 points behind over three. That is a name moving broadly with the market and the sector, not a leader and not a laggard — which is itself informative: the best-positioned company in the AI supply chain, reporting +44.7% monthly revenue growth, is not being rewarded for it. Some of that is the June spike to $479 and the July round-trip to $372.72 unwinding. Blended with the position-risk read at 60/40, risk-reward scores 52.
| Sub-signal | Reading | Score |
|---|---|---|
| Fed direction | Hawkish — cuts priced out, a September hike at roughly 44% implied, three hawkish dissents in July. Unfavourable for a long-duration multiple. | 25 |
| VIX | 15.19 (FRED VIXCLS, 17 Aug 2026 — the report's own date) — just above the 15 threshold, so neutral rather than risk-on. It was 14.25 on 14 August; an earlier version of this report scored the stale figure and read it as risk-on. | 60 |
| Yield curve | Normal and bear-steepening, 10s2s about +48bp | 70 |
| Sector regime | XLK is Neutral / Neutral / Outperform, but TSM's own macro line is Outperform / Outperform / Strong Outperform — rotation is into this name specifically | 75 |
Average 57.5, taken at 56 for the 19 August FOMC minutes sitting two days out. A hawkish read of those minutes is the single most likely near-term source of multiple compression.
Analyst grade actions, last 30 days: zero upgrades, zero downgrades. Bernstein maintained Outperform on 11 August; Needham maintained Buy on 27 July; Barclays (Overweight), DA Davidson (Buy), TD Cowen (Hold) and Susquehanna (Positive) all maintained around the 16–17 July print. Every action in the window is a maintain, which the framework scores 52 — no change in sentiment, in either direction.
Estimate revisions: upward. The company itself raised full-year revenue growth guidance to "slightly above 40%" on the Q2 call, and July's +44.7% put it ahead of that guidance.
News tone: strongly positive and, frankly, uncritical. Of twelve Polygon-scored articles in the last nine days, seven carry a positive TSM sentiment and five neutral; none are negative. The recurring framing is "the clear winner of the capex boom". One 16 August piece states the case and the counter-case in the same breath — 31× earnings against a ten-year average of 23× — and concludes to buy anyway. Tone scores 80, and at that level of unanimity it is worth reading as a crowding signal rather than a confirmation. Sentiment blended: 65.
| Date | Event | Impact |
|---|---|---|
| 19 Aug 2026 | FOMC minutes — the macro report's named trigger for its live September-hike tail | Medium-high, macro |
| ~10 Sep 2026 | TSMC August monthly revenue | Medium — the tightest read on the driver anyone gets |
| 16 Sep 2026 | FOMC decision + projections | Medium-high, macro |
| 15 Oct 2026 | Q3 2026 results — guided to US$44.6–45.8bn revenue and 65–67% gross margin | High, company-specific |
| 10 Dec 2026 | Ex-dividend, US$1.085439 | Low |
No company-specific catalyst inside 30 days and no cluster: two macro events plus a monthly disclosure. Catalyst clustering score 65 — "focused" rather than noisy, so no position-size reduction on catalyst density.
Timing = (76.1 × 0.30) + (52 × 0.20) + (56 × 0.15) + (65 × 0.18) + (65 × 0.17) = 64. That is "Improving" — up 14 points from 50 on 31 July, and the only pillar that moved in the stock's favour. The tape genuinely turned; it just does not get a vote once the Valuation Ceiling fires.
Timing confidence 70%. Base 75, less 5 for a high-impact macro release two days out. No earnings-proximity penalty (59 days), no VIX penalty (VIX 15.19 — the −10 penalty needs VIX above 30), no catalyst-clustering penalty (65).
| Date | Event | Impact | Forecast | Previous | Relevant? | Why |
|---|---|---|---|---|---|---|
| 2026-08-18 | Housing Starts / Building Permits (Jul) | High | 1.35m / 1.37m | 1.427m / 1.374m | No | Housing is not a semiconductor input |
| 2026-08-19 | FOMC Minutes | High | — | — | Yes | The macro report's named trigger for its live September-hike tail. A hawkish read lifts the 10-year, which lowers the warranted multiple in §4 directly |
| 2026-08-26 | Core PCE Price Index MoM (Jul) | High | 0.3% | 0.1% | Yes | The Fed's preferred gauge; feeds the same hike debate |
| 2026-09-01 | ISM Manufacturing PMI (Aug) | High | 55.0 | 55.6 | Yes | The listed high-impact release for semiconductors — a manufacturing-cycle read |
| 2026-09-04 | Non-Farm Payrolls / Unemployment (Aug) | High | +12k / 4.2% | −23k / 4.1% | Yes | July printed −23k; a second negative month would shift the whole regime read |
| 2026-09-11 | CPI (Aug) | High | — | 3.4% YoY | Yes | Inflation path drives the discount rate |
| 2026-09-16 | Fed Interest Rate Decision + Projections | High | — | 3.75% | Yes | The event the hike debate resolves into |
| Date | Event | Actual | Forecast | Surprise | Impact |
|---|---|---|---|---|---|
| 2026-08-12 | CPI YoY (Jul) | 3.4% | 3.4% | In line | Neutral — inflation sticky above target but not accelerating |
| 2026-08-12 | Core CPI YoY (Jul) | 2.5% | 2.5% | In line | Neutral |
| 2026-08-13 | Producer Price Index MoM (Jul) | 0.0% | 0.2% | −100% | Mildly positive — pipeline pressure softer than feared |
| 2026-08-14 | Retail Sales MoM (Jul) | −0.6% | +0.1% | Large miss | Negative for the economy, ambiguous for TSMC — AI capex is not consumer-funded |
| 2026-08-14 | Michigan Consumer Sentiment (Aug) | 51.0 | 54.5 | −6.4% | Negative — the stagflation read in the macro report |
Semiconductors carry medium macro sensitivity, so nothing here overrides the signal — but two things are worth naming. First, the FOMC minutes on 19 August are two days away and are the macro report's explicit trigger for its live September-hike tail. A hawkish read pushes the ten-year up, and every basis point on the ten-year lowers the warranted multiple in §4 — this is the one macro event with a direct, mechanical line into this report's central number.
Second, the recent tape is a genuinely stagflationary set: inflation sticky at 3.4% with retail sales at −0.6% and consumer sentiment at 51.0, the weakest of the series shown. That combination is bad for most equities and close to irrelevant for TSMC's order book, which is funded by hyperscaler capital budgets rather than household spending. It is, however, exactly the combination that keeps the Fed hawkish — and a hawkish Fed is bad for a 34× multiple. The transmission runs through the discount rate, not through demand.
| Timeframe | Trend | Direction | RSI | MACD | Key S/R | Breakout | Vol |
|---|---|---|---|---|---|---|---|
| Monthly | Uptrend ↑ | Bullish | 69.1 | +71.33, hist +11.91 | S: $134 R: $479.00 | Resistance breakout | 0.4× |
| Weekly | Uptrend ↑ | Bullish | 62.5 | +21.21, hist −3.74 | S: $319.07 R: $479.00 | Resistance breakout | 0.2× |
| Daily | Strong uptrend ↑↑ | Bullish | 56.5 | +2.19 vs signal −1.13, hist +3.32 | S: $419.19 R: $442.58 | Resistance breakout | 0.5× |
| Hourly | Strong uptrend ↑↑ | Bullish | 52.1 | +0.63, hist flat | S: $424.10 R: $436.04 | Resistance breakout | — |
| 15-min | Weakening → | Neutral | 46.5 | −0.14, hist −0.02 | S: $426.80 R: $435.71 | — | — |
| Confluence: Mostly Bullish · MTF Score 76 | |||||||
All four timeframes that matter are in confirmed uptrends and only the 15-minute chart is soft, which is noise. The structurally important change since 31 July is on the daily: price at $430.97 is now above both the 50-day average ($424.88) and the 200-day ($364.14), and the 50-day sits above the 200-day — the textbook strong uptrend configuration. On 31 July the close was $404.25 against a 50-day of $425.33 — below it. (The prior report recorded its price at $405.27, an intraday quote; and $425.16 is the 50-day as at 14 August, not 31 July. Both were misattributed here in an earlier version.) The MACD confirms, and precisely: the histogram has been positive for ten consecutive sessions, since 4 August (+0.38, +1.15, +1.90, +2.45, +2.62, +2.86, +3.36, +3.62, +3.35, +3.32) — the MACD line therefore crossed above its signal line on 4 August. What happened on 13 August is a different and stronger event: the MACD line crossed above zero (−0.33 → +0.83). An earlier version of this report conflated the two and undercounted the streak by six sessions. RSI at 56.5 is constructive without being stretched.
All prices here are raw, unadjusted closes. TSMC's dividend is under 1%, so the dividend-adjustment trap that inverts moving-average reads on high-distribution names cannot bite here — but the rule was applied anyway rather than assumed away.
The pattern that is not present. This is not "buying the dip in an uptrend" — the dip was in late July, at $372.72, and it has been bought. Nor is it a clean all-timeframe breakout, because the volume is not there: the 20-session average is 12.74m measured to the 14 August close and 12.63m measured to the 17 August close; the last two completed sessions traded 6.32m on 14 August (0.50×) and 9.48m on 17 August (0.74× on the first window, 0.75× on the second). Both windows are given because the label, not the arithmetic, is what a reader has to trust. The tool flags "resistance breakout" on every timeframe, and at 0.50–0.74× the 20-session average that is a level being cleared, not a move being confirmed. That distinction is what keeps the Technical entry group closed in §12.
Levels that matter: the rising floor of daily lows since the 29 July bottom — $372.72 → $387.60 → $398.80 → $412.11 → $418.87 → $427.02 — and the 50-day at $424.88, which price must now hold. (The data tool's swing-low array is sorted by value, not by date; read as a chronology it produces a sequence that is neither rising nor in order, and an earlier version of this report did exactly that.) Below, the tool's other daily supports are $419.19 (26 Jun), $407.99 (6 Aug) and $405.51 (9 Jun) — dated, because they are value-sorted levels rather than a sequence. Above, the daily resistance shelf is $442.58 and then the all-time high at $479.00.
Six months of raw daily closes to 17 August 2026 ($430.97), with the 50-day simple moving average. The June spike to $479.00, the July round-trip to $372.72 and the August reclaim of the 50-day are all visible. The purple line is what the framework's warranted multiple says today's operating earnings justify — the gap between it and the price is this report.
+31.1% over 12 months. Hyperscaler capex lands near Morgan Stanley's US$1.1tn 2027 forecast, the ~10% 2027 price increase is realised in full, and Intel's 14A slips as 18A did — leaving TSMC above 90% at the leading edge. FY2027 EPS beats the US$21.86 consensus toward US$24, and the market still pays about 23.5× forward. Trigger: monthly revenue holding above +40% year on year through Q4, and a Q3 print at or above the top of the US$45.8bn guide. This brackets the analyst panel — FMP's consensus is $589, Yahoo's mean $547.
+9.1% over 12 months. The earnings arrive and the multiple gives some back. Consensus FY2027 EPS of about US$21.9 is broadly met, and the market pays roughly 21.5× forward — the middle of TSMC's own five-year forward range rather than the top. Note what this says: the base case is that the business grows into the price and the holder is paid single digits for a year of waiting. That is not a bad outcome; it is simply a poor risk-reward against a bear that costs a third. Trigger: revenue growth decelerating gradually from +44.7% toward the high 20s as comparisons harden, with gross margin settling in the guided 65–67% band.
−32.7% over 12 months. Two legs, and the report is required to carry the first explicitly.
(1) The AI-cohort de-rating. The 12 August macro report keeps "S&P 500 concentration / AI earnings-quality unwind" armed,
with the AI Big 10 at 41% of the index. Its named triggers are a hyperscaler capex guide-down of more than 20% year on year, an AI private-valuation
markdown, or non-operating gains turning negative. If one fires, TSMC's clean trailing multiple compresses from 34× toward 20× —
the 40–50% cohort compression the framework requires be modelled — partly offset by earnings still growing. The arithmetic: 20× a clean EPS of about US$14.50 gives $290. That EPS is up about 14% on today's US$12.68, not down — the bear here is a de-rating, not an earnings collapse, which is why it stops at −32.7% rather than the −41% a flat-earnings compression would give. Sell-side
commentary already flags "limited visibility on capex beyond 2027 amid greater investor demand for spending discipline", so this is not a distant abstraction.
Falsified if breadth keeps broadening — the equal-weight S&P continuing to lead the cap-weighted index and the Nasdaq. The 12 August macro report's own reading was RSP +3.6% against SPY +2.8% and QQQ +0.6% over the month; recomputed here on daily closes from 14 July to 14 August it is RSP +4.4%, SPY +3.3%, QQQ +1.6% (the SPY figure being the same +3.26% used in §7). Different windows, same ordering — the tail's trigger is still receding on either read.
(2) Competition and geopolitics. Intel converts the October 2026 release of its 14A 0.9 process design kit into named leading-edge design wins, or the Taiwan Strait deteriorates from jurisdictional pressure to actual interdiction. Note the scoping, because the first version of this leg overstated it: Intel's 14A does not reach high-volume manufacturing until 2028, so it cannot take measurable share inside twelve months. What it can do inside twelve months is change the story — a marquee design win announced against a 34× multiple is a de-rating event well before it is a revenue event. On the second: the Institute for the Study of War's 14 August 2026 update records a PRC–Indonesian naval exercise inside Taiwan's exclusive economic zone on 12 August, and a Guangdong Maritime Safety Administration traffic-control scheme over northbound vessels in the Taiwan Strait from 6 to 9 August, which Taipei condemned as having no legal basis in international waters. The CFR tracker rates the confrontation "Developing" with critical US impact as of 2 August. The specific falsification condition for the Taiwan leg is a maritime quarantine, a customs-inspection regime, or an air-defence-identification-zone enforcement action that actually stops or diverts commercial traffic — drills and rhetoric are not it. The August traffic-control notice is the first item to sit near that line rather than comfortably below it: it asserted control over commercial shipping, but it ran four days, carried a typhoon pretext, and no stoppage or diversion of commercial traffic has been reported — the source is silent on whether any occurred, which is a weaker statement than none did, and is deliberately the one made here. A repeat without a weather pretext, or one extending beyond a week, would move this leg from a tail to a live risk. A genuine blockade is a far deeper outcome than $290, is not modelled here, and this report does not price it.
Forecast:
| Rule | Forecast | Basis | Confidence |
|---|---|---|---|
| ENTRY — Fundamental: price below $351 | Unlikely in 4–6 months without a de-rating or automatic by mid-2027 through earnings growth | Two routes, and they are very different. Route one: an 18.6% drawdown, which needs the §11 bear leg — possible, 20% weighted. Route two: the fair value rises. Fair value is 27.70× clean trailing EPS; that EPS compounds toward the FY2027 consensus, so on consensus delivery the $351 line reaches today's $431 around mid-2027 with no price fall at all. The most likely resolution of this Hold is not a crash — it is time. | Moderate |
| ENTRY — Technical: a volume-confirmed 50-day reclaim | Catalyst-dependent — days, if the right event lands | Volume conditions cannot be time-projected; they need an event. Three are dated: the FOMC minutes on 19 August, August monthly revenue around 10 September, and the Q3 print on 15 October. Price is already above the 50-day and RSI and MACD both qualify, so this group is one high-volume session away from opening. It is the group most likely to open next — but see the note below on what that would and would not mean. | Moderate–High |
| ENTRY — Catalyst: a >+5% post-earnings move | 15 October 2026, or not at all before then | Structurally dated. TSMC has guided Q3 revenue to US$44.6–45.8bn and gross margin to 65–67%; July revenue is already running ahead of the full-year guide. A beat is likely; a >5% single-day move on a US$2.2tn company is a higher bar, and the last print went the other way. | Moderate |
| EXIT — Stop-loss: two closes below $396 | Unlikely in the next 4–6 weeks | $396 is 8.1% below spot and below both the 50-day ($424.88) and the rising swing-low sequence. It would take a hawkish FOMC-minutes surprise or the start of the cohort de-rating. It sits above the late-July low of $372.72, so it is a stop that respects the structure without being loose. | Moderate |
| EXIT — Profit target: price at the median analyst target with RSI > 70 | Unlikely inside 12 months | The median target is $536 (Yahoo) to $577 (FMP), 24–34% above spot — roughly the bull case. Both legs would have to happen together. | Low |
An important caveat on the Technical group, so nobody misreads the ladder. If a high-volume session opens that group, the conviction ladder moves from Wait to Half-Size — but the signal stays HOLD. The ladder answers "how much, given you are permitted to buy"; the Valuation Ceiling gate withholds the permission. The gate clears when the clean multiple falls back below 28×, which at today's clean EPS means a price of about $355, or the same level reached from the other direction as earnings grow. Volume does not clear a valuation gate.
Forecast:
No exit trigger is live, so the action is Hold. For an existing holder nothing here says sell: the stop is 8.1% away, the thesis is intact on every leg, and the profit target is a quarter above spot. The Hold signal is a statement about new money at $430.97, not an instruction to existing holders. Those are different questions and the framework answers both — see the box below.
You are risking about 8% to a mechanical stop and a third to the bear case, to gain a single-digit expected return over twelve months.
What you are risking. The hard stop at $396 is −$34.97, or −8.1%. The bear case is $290, −32.7%, and it is weighted at 20% — not a remote tail. Two of the three entry rules are not merely unmet but structurally closed: you would be buying 23% above what the framework's warranted multiple says today's earnings justify, and on sessions that cleared the 50-day average at 0.50–0.74× normal volume rather than the 1.5× the rule requires. You would also be buying a name that has lagged the S&P on both the one- and three-month windows and its own sector ETF over three months, while everything about the business improved — a leader being paid like an average stock. And the FOMC minutes land in two days, with the macro report's September-hike tail live at ~44%; a hawkish read lifts the ten-year and lowers the warranted multiple mechanically.
What you are gaining. Immediate exposure to the base case of $470, +9.1%, and the bull at $565, +31.1%; the probability-weighted value is $457.75, +6.2%. You collect a forward dividend yield of about 1.0% while you wait, and you own the optionality: advanced packaging priced inside a foundry multiple, a ~10% 2027 price increase that consensus has not obviously banked, US$77bn of net cash, and a US Arizona premium currently valued at nothing. You also own a business compounding earnings at 30%+ with 60% operating margins and 73% of its market — which is why the Quality pillar reads 87 and the driver 88.
The read. Risk-reward from here is 1.12 to 1 on the base case ($39.03 of upside to $470 against $34.97 to the stop) and only 0.77 to 1 on the probability-weighted value ($26.78 against the same $34.97) — and that is before noting the bear case is four times the size of the stop. Waiting improves the deal in two ways, and one of them does not require the shares to fall: below about $355 the Valuation Ceiling clears outright, and on consensus earnings delivery the warranted fair value climbs to today's price by roughly mid-2027 on its own. Acting now is not reckless. It is paying a full price for a great business, on a tape that is not confirming, two days before a Fed event. That is an assessment, not a verdict.
You are giving up a compounding position in the best-placed company in the AI supply chain to protect against a de-rating that has not started.
What you are giving up. The base case to $470 (+9.1%) and the bull to $565 (+31.1%). The ~1.0% forward dividend, whose December instalment has risen 78.5% in dollar terms over two years. The optionality listed in §4. And a real chance of being wrong in the most expensive direction — TSMC has beaten its own guidance every month this year, July revenue accelerated to +44.7%, and the Street is 18 buys to 7 holds with no downgrades in 30 days. Selling here also crystallises a decision the framework itself is not making: the warranted fair value of $351 is what today's earnings justify, and those earnings are growing at 30%+, so the fair value is a moving target moving toward you.
What you are protecting. A third of the position if the AI-cohort de-rating fires — the macro report keeps that tail armed, and the compression from 34× to 20× is the framework's own required magnitude. You would sidestep the 19 August FOMC minutes and the 16 September decision. And you would be selling into a multiple at the 87th percentile of its own five-year range, with FMP's independent health screen marking price/earnings 2 out of 5 and price/book 1 out of 5.
The read. No exit rule is triggered. The stop is 8.1% below, no thesis-invalidation condition holds, and the profit target is 24–34% above. There is no mechanical reason to sell, and the Valuation Ceiling gate deliberately does not create one — it caps new buying, it does not force liquidation. For an existing holder this is a hold-and-do-nothing zone. If the position has grown large on an 87% twelve-month move (the 29 Aug 2025 close was $230.87), trimming size back to a normal weight is a portfolio decision this framework supports and does not require.
Position sizing not computed — no portfolio allocation or role was specified for this batch run. For context rather than sizing: the conviction ladder reads Wait (0 of 3 entry groups met), which is a ladder_factor of 0× — the framework's guidance at this reading is not a percentage but a level. The levels to watch are a volume-confirmed hold of the 50-day at $424.88, the support shelf at $405.51, and the price at which the Valuation Ceiling gate actually clears, about $355.
Volatility context. ATR is $14.00, or 3.2% of price — a typical daily range of roughly $14. Beta is 1.258, so a 5% position behaves like about 6.3% in market-risk terms. The twelve-month drawdown from the June high of $479.00 to the July low of $372.72 was −22.2%, inside eight weeks, with no change in the fundamentals — a useful calibration of what this share does to a portfolio in a bad month.
{
"ticker": "TSM",
"exchange": "NYSE",
"exchange_ticker": "NYSE:TSM",
"api_ticker": "TSM",
"storage_ticker": "TSM",
"isin": "US8740391003",
"date": "2026-08-17",
"version": "v6",
"brand": "TSMC",
"company": "Taiwan Semiconductor Manufacturing Company Limited",
"currency": "USD",
"reported_currency": "TWD",
"usd_twd_spot": 31.782,
"security_note": "NYSE ADR (1 ADR = 5 TWSE ordinary shares); analysed as the USD ADR. All per-share figures are USD per ADR, translated from TWD at spot USD/TWD 31.782 (17 Aug 2026). Shares x price (5,186.4m ADS x $430.97 = $2.235tn) reconciles to the reported market cap.",
"analysis_status": "on-going",
"finder_ticker": "TSM",
"finder_exchange": "\ud83c\uddfa\ud83c\uddf8 NYSE",
"section": "EM Equities",
"mode": "batch",
"user_horizon": null,
"user_allocation_pct": null,
"portfolio_role": null,
"lifecycle_stage": "growth",
"sector": "Semiconductors",
"sub_industry": "Semiconductors \u2014 Advanced-Node Foundry",
"gics_sector": "Information Technology",
"country": "Taiwan",
"price_at_rating": 430.97,
"price_asof": "2026-08-17",
"market_cap_usd": 2235214594048,
"beta": 1.258,
"fifty_two_week_range": "223.70-479.00",
"signal_short": "HOLD",
"signal_medium": "HOLD",
"signal_long": "HOLD",
"primary_signal": "HOLD",
"base_signals_pre_amplification": {
"short": "HOLD",
"medium": "HOLD",
"long": "HOLD"
},
"amplification_applied": {
"short": "none \u2014 base HOLD, and HOLD never amplifies; independently the Full/Expensive band bars STRONG BUY",
"medium": "none \u2014 driver 88 (>=65) and pressure Tailwind both PASS, but the base signal is HOLD and the warranted ratio 1.23 is above the 1.20 amplification bar",
"long": "none \u2014 prior report's BUY->STRONG BUY is withdrawn: the band test (ratio must be <1.20) fails at 1.23 before the base signal is consulted"
},
"short_entry_confirmed": false,
"short_hold_reason": "expensive",
"short_cap_reason": "Base matrix row is High Quality (87) / Expensive valuation (38) / any timing -> HOLD. The Valuation Ceiling gate (Gate 3) independently caps every horizon. The half-size quality-starter override was tested and does NOT fire: it requires Valuation >= 40 and a non-Expensive band, and requires medium and long to be BUY-or-better; none of those hold. Technical and Catalyst entry groups are also unmet (the 50-DMA was reclaimed on 0.50x (14 Aug) and 0.74x (17 Aug) volume against a 12.74 million-share 20-session average, versus the 1.5x the rule requires; no qualifying post-earnings event).",
"quality_score": 87,
"valuation_score": 38,
"timing_score": 64,
"driver_score": 88,
"quality_detail": {
"revenue_growth_yoy_q2_2026": 0.36,
"revenue_growth_yoy_july_2026": 0.447,
"gross_margin_latest_q": 0.677,
"gross_margin_guide_q3": "0.65-0.67",
"operating_margin_q": 0.603,
"operating_margin_ttm": 0.56,
"net_margin_ttm_reported": 0.504,
"net_margin_ttm_clean": 0.47,
"roe_reported": 0.349,
"roe_reported_provider_field": 0.4,
"roe_clean": 0.325,
"roic": 0.525,
"interest_coverage": 886,
"interest_coverage_provider_field": 285.4,
"current_ratio": 2.46,
"debt_to_equity": 0.153,
"net_cash_usd_bn": 77.1,
"fcf_ttm_usd_bn_range": "23.0-35.2",
"industry_benchmark_name": "Gross Margin + Capacity Utilisation",
"industry_benchmark_value": "GM 67.7% + advanced-node utilisation >90%",
"industry_benchmark_score": 94,
"moat_score": 87,
"roic_percentile_vs_peers": 95,
"capital_allocation": 85,
"management_skin_in_game": 55
},
"eps_trailing": 13.6,
"trailing_pe": 31.7,
"clean_pe": 34.0,
"clean_peg": 1.12,
"nonop_pct_of_net_income": 8.0,
"valuation_detail": {
"eps_trailing_reported_usd": 13.6,
"eps_trailing_clean_usd": 12.68,
"pe_trailing_reported": 31.7,
"pe_trailing_clean": 34.0,
"pe_trailing_op_plus_interest": 32.6,
"pe_forward_ntm": 21.6,
"pe_fy2026": 25.7,
"pe_fy2027": 19.7,
"peg_forward": 0.71,
"price_to_book": 11.0,
"price_to_sales": 16.0,
"fcf_yield_on_ev_pct_range": "1.1-1.6",
"enterprise_value_usd_bn": 2158,
"dividend_yield_trailing": 0.81,
"dividend_yield_forward": 1.01,
"implied_growth_rate": 20.2,
"consensus_growth_rate": 27.4,
"historical_valuation_decile": 9,
"historical_pe_percentile_5y": 87,
"historical_pe_percentile_5y_alt_convention": 78,
"dnb_relative_arm_margin_pct": 1.4,
"dnb_relative_arm_watch": "WATCH ITEM FOR THE NEXT RUN. On the announcement-date convention the 90th-percentile (top-decile) bar is 34.46x against an actual clean multiple of 33.98x - negative by only 1.4%, about $6 on the ADR. The second condition of DNB Trigger 2s relative arm (no growth acceleration: forward EPS growth 30.4% vs trailing 55.2%) is ALREADY MET, so the decile test is the only thing holding the trigger off. A ~1.5% rise in the share price with earnings unchanged would fire a hard DO NOT BUY on the relative arm. Correct as negative today; re-test explicitly next run rather than assuming headroom.",
"historical_pe_percentile_note": "87th percentile / decile 9 on an announcement-date earnings-availability convention (TSMC releases ~16 days after quarter end); 78th / decile 8 on an SEC filing-date convention, which lags a further month. Independently reconstructed both ways; the 5-year maximum agrees at ~43.0x either way. Only the top-decile question matters for DNB Trigger 2, and it is negative on both.",
"historical_clean_pe_5y_range": "11.8-43.0, median ~28",
"fair_value_estimate": 351
},
"timing_detail": {
"mtf_confluence": 76,
"risk_reward_score": 52,
"relative_strength_vs_spy": -1.84,
"relative_strength_vs_sector": 4.5,
"relative_strength_vs_spy_3mo": -1.7,
"relative_strength_vs_sector_3mo": -1.78,
"relative_strength_benchmark": "SOXX",
"catalyst_clustering_score": 65,
"dynamic_macro_weight": 0.15,
"macro_overlay_score": 56,
"sentiment_score": 65,
"rsi_daily": 56.5,
"atr_daily": 14.0,
"sma50_daily": 424.88,
"sma200_daily": 364.14,
"vix": 15.19,
"vix_asof": "2026-08-17"
},
"relative_strength_vs_spy": -1.84,
"relative_strength_vs_sector": 4.5,
"val_band": "expensive",
"val_band_basis": "guardrail-forced: the warranted RATIO of 1.23 is the Full band (1.20-1.40), but the actual clean multiple 34.0x is at or above the 28x Semiconductors guardrail line, which makes the name Expensive regardless of the ratio (SKILL Pillar 2, THE ANCHOR step 5(b)). Both facts are recorded; the stricter governs.",
"warranted_multiple": 27.7,
"actual_multiple": 34.0,
"warranted_ratio": 1.227,
"val_multiple_basis": "clean trailing P/E (operating income after tax, per step 7b) \u2014 34.0x. Trailing, not forward: the two-stage warranted multiple is an expression in E0, so a forward numerator credits a year of growth twice. The 2026-07-31 report used a forward NTM P/E against the same E0-based denominator; that mixed basis produced 0.76 and is corrected here.",
"discount_rate_r": 0.0913,
"risk_free_10y": 0.0463,
"risk_free_source": "FRED DGS10, 2026-08-13 = 4.63%, retained as a DISCLOSED BATCH-WIDE PIN for cross-name comparability. Full fortnight span 4.63-4.72%: 4.68% on 14 Aug, 4.70% in the macro report of record, and 4.72% on 17 Aug - THE REPORT OWN DATE. Sensitivities: 4.70% gives warranted 27.37x and ratio 1.24; 4.72% gives warranted 27.28x, ratio 1.25 and fair value about 346. Band, gate and signal unchanged at every value in the span; 4.63% is the fortnight LOW and therefore the most generous input, so the pin does not flatter the conclusion - the report-date rate makes the name dearer",
"g_near": 0.15,
"g_term": 0.03,
"sector_guardrail": 28.0,
"sector_guardrail_note": "Semiconductors line (28x, defined as a mid-cycle P/E) applied \u2014 the more specific of the two applicable lines. The Information Technology line is 33x and the clean 34.0x breaches that too.",
"competitive_share_trajectory": "gaining",
"competitive_threat_level": "moderate",
"economic_alignment_stance": "Trend-Following",
"economic_alignment_conviction": 72,
"economic_alignment_pressure": "Tailwind",
"economic_alignment_pressure_short": "Tailwind",
"economic_alignment_pressure_medium": "Tailwind",
"economic_alignment_pressure_long": "Tailwind",
"economic_alignment_source": "watchlist-signal",
"macro_report_date": "2026-08-12",
"macro_regime": "Energy-shock stagflation",
"macro_watchlist_signal_tsm": {
"s": "O",
"m": "O",
"l": "SO"
},
"sector_signal_xlk": {
"s": "N",
"m": "N",
"l": "O"
},
"em_equity_signal": {
"s": "N",
"m": "N",
"l": "O"
},
"analyst_consensus_target": 547.09,
"analyst_target_high": 700,
"analyst_target_low": 431.5,
"analyst_target_median": 536,
"analyst_target_upside_pct": 27.0,
"analyst_target_source": "Yahoo (n=18) \u2014 the more conservative of two panels; FMP shows 500/577/589/700",
"analyst_grades_consensus": "Buy",
"analyst_bullish_pct": 72,
"analyst_coverage_count": 25,
"fmp_rating": "B+",
"fmp_overall_score": 3,
"recent_upgrades_30d": 0,
"recent_downgrades_30d": 0,
"overall_confidence": 70,
"quality_confidence": 75,
"valuation_confidence": 78,
"timing_confidence": 70,
"fair_value_est": 351,
"stop_loss": 396,
"target_price": 470,
"scenario_base_target": 470,
"scenario_bull_target": 565,
"scenario_bear_target": 290,
"scenario_probabilities": {
"bull": 25,
"base": 55,
"bear": 20
},
"scenario_weighted_value": 457.75,
"entry_groups_met": 0,
"entry_conviction": "Wait",
"exit_groups_live": 0,
"exit_action": "Hold",
"hard_gate_state": "caution",
"gates_triggered": [
"Gate 3 \u2014 Valuation Ceiling: clean trailing P/E 34.0x is above the 28x Semiconductors guardrail line, forcing the Expensive band and capping every horizon at HOLD"
],
"gates_caution": [
"Gate 5 \u2014 Taiwan Strait geopolitical risk: structural rather than a dated binary ruling, so caution not triggered. Re-verified this run (ISW China-Taiwan Update 14 Aug 2026: PRC exercises inside Taiwan's EEZ, called a 'military provocation' by Taiwan's MAC; CFR tracker 'Developing', 2 Aug 2026). Position-sizing caution."
],
"do_not_buy_triggers": [],
"dnb_note": "DNB Trigger 2 was tested on both arms and fires on neither. RELATIVE ARM: needs the primary multiple in the top decile (90th+ percentile) of its own 5-year range AND no growth acceleration. The second condition holds (forward EPS growth 30.4% vs trailing 55.2%) but the FIRST DOES NOT \u2014 the clean multiple sits at the 87th percentile (decile 9) of a 5-year range reconstructed from 60 months of month-end prices and trailing-four-quarter EPS at each month's spot FX (range 11.8x-43.0x, median 27.8x). ABSOLUTE ARM (a): needs >=2.0x warranted (it is 1.23) or >=1.5x the guardrail line (42.0x; it is 34.0x) \u2014 neither. ABSOLUTE ARM (b): needs the Expensive band AND a LIVE de-rating catalyst. The band condition holds. The catalyst condition does not: the macro tail 'S&P 500 concentration / AI earnings-quality unwind' is ARMED but its trigger is RECEDING (breadth broadening \u2014 RSP +3.6% vs SPY +2.8% vs QQQ +0.6% over the month), and an armed-but-not-triggering tail is not a live de-rating catalyst; it caps via the Valuation Ceiling and belongs in the bear case, which is where it is carried. THIS IS A DETERMINISTIC-GATE OPT-OUT AND IS FLAGGED AS ONE: the SKILL wording for arm (b) is that the tail must be currently ARMED, and the 2026-08-12 macro state does record it as armed (status: armed but trigger RECEDING). The reading applied here rests on the project standing precedent armed-not-triggering-tail-hold-not-dnb, which holds that an armed tail whose trigger is receding caps via the Valuation Ceiling rather than firing DNB 2(b). Step 1 of the same SKILL section is separately satisfied - the cohort de-rating IS inherited into the section-11 bear with the required 34x-to-20x magnitude. Recorded explicitly so the opt-out is auditable rather than silent. On cohort membership: TSM is in the AI cohort on the CAPEX-LEVER limb (its earnings are materially levered to hyperscaler capital spending) but NOT on the index-concentration / non-operating-gains limb \u2014 it is a foreign private issuer ADR and not an S&P 500 constituent, and its non-operating income is 8.0% of net income and a positive contributor, i.e. the opposite of the earnings-quality distortion the tail describes. Trigger 5 (structural business-model threat) does not fire: Intel 14A is a dated competitive risk (0.9 PDK October 2026, risk production H2 2027, high-volume manufacturing 2028), not an existential one. Gate 4's earnings-quality backstop (>=30% non-operating) also does not fire at 8.0%.",
"tails_ruled": {
"1_ai_concentration_earnings_quality": "armed, trigger RECEDING \u2014 carried as the primary leg of the 11 Bear scenario with the required 34x->20x cohort compression; does NOT feed DNB 2(b) because it is not triggering",
"2_private_credit_crack": "building, NOT armed \u2014 no inheritance (inheritance requires an armed tail); no direct transmission to a net-cash foundry in any case",
"3_hormuz_closure": "LIVE but not triggering (Brent 88.52 on the 14 Aug 2026 settle, against the >$100 sustained trigger; the 12 Aug macro report of record quotes 88.58 on its own earlier date). The level is not the whole story and the earlier framing understated it: Brent rose about 8% that week, from roughly 81.99 to 88.52, on Reuters reporting of a US threat of an indefinite naval blockade of Iran - so the move is geopolitical, not demand, and it is closer to the trigger than a flat reading suggests. No material transmission to TSMC beyond generic energy-cost and freight inflation; noted, not scored",
"4_fed_hikes_september": "LIVE, ~44% implied \u2014 its trigger, the 19 Aug FOMC minutes, sits inside the next-update window and is the single macro event with a mechanical line into this report (a higher 10-year lowers the warranted multiple). Semiconductors are MEDIUM macro-sensitivity so it does not re-schedule the next update."
},
"next_update_date": "2026-08-31",
"next_check_date": "2026-08-31",
"next_update_basis": "default +14d from 2026-08-17 (Mon 31 Aug is a trading day; US markets closed 7 Sep for Labor Day, not relevant here). No dated company catalyst inside the window \u2014 Q3 2026 results are expected 15 Oct 2026 (sourced externally after get_earnings_calendar returned nothing, from a search citing TSMC's financial calendar; TSMC's own calendar page lists only monthly-sales dates and third-party calendars mark it unconfirmed, so treat as high-confidence not confirmed - immaterial either way, the +14d default governs) and the August monthly-revenue release is ~10 Sep. Semiconductors are MEDIUM macro-sensitivity, so the 19 Aug FOMC minutes do not qualify as a scheduling trigger under the 3-day high-sensitivity exception.",
"last_updated_human": "Aug 17, 2026",
"prior_report": "calibration-TSM-20260731-1200.json",
"prior_primary": "STRONG BUY",
"delta_vs_prior": "All three horizons cut to HOLD from BUY / BUY / STRONG BUY. TWO causes, and only one is market news. (1) Price +6.3%, 405.27 -> 430.97. (2) A FRAMEWORK-APPLICATION CORRECTION: the 31 July report scored the anchor by dividing a FORWARD NTM P/E (~21x) by a warranted multiple built on TRAILING earnings, a mixed basis that credits ~47% of growth twice on this name. Restated on a consistent trailing basis the ratio is 1.23, not 0.76 (a consistent FORWARD construction, deflating warranted by one year of growth to 24.1x, gives 0.90 \u2014 so neither consistent basis reproduces 0.76). The clean trailing multiple 34.0x then breaches the 28x Semiconductors guardrail, forcing val_band attractive -> expensive, firing Gate 3 (Valuation Ceiling, newly TRIGGERED) and capping all three horizons. NO NEW EARNINGS PRINT since the last report: Q2 2026 was released 16 Jul and was already in it (margins reconciled to confirm; the 14 Aug date on the filing is the 6-K). Everything measuring the business improved: July revenue accelerated to +44.7% YoY, Timing rose 50 -> 64 as price reclaimed the 50-DMA, Economic Alignment rose 50 -> 72 and its pressure flipped Headwind/Headwind/Tailwind -> Tailwind on all three horizons because TSM gained its own macro watchlist line (O/O/SO) in place of the XLK sector map, and the driver held at 88 (re-derived, not carried). None of that reaches the signal: HOLD never amplifies, and a Full-or-Expensive band bars STRONG BUY outright. Scores: Quality 88 -> 87 (-1, the FCF-conversion sub-signal was cut after reconciling two providers' conflicting FCF figures; separately the moat ROSE 82 -> 87 on re-measurement), Valuation 64 -> 38 (-26), Timing 50 -> 64 (+14), Driver 88 flat. Ladder cut: entry_conviction Half-Size -> Wait and entry_groups_met 1 -> 0, as the Fundamental entry group closed on fair value 475 -> 351. Scenarios: base 475 -> 470, bull 560 -> 565, bear 325 -> 290 (a fuller cohort de-rating leg), stop 380 -> 396. Own-history decile 5 -> 9 \u2014 basis change, not a market move: the prior figure was on a forward multiple, and on the trailing basis now used 34.0x is the 87th percentile of a 5-year range rebuilt from 60 months of prices and earnings. PORTFOLIO WATCHLIST: the short-horizon BUY is gone, so TSM VACATES its EM Equities x US slot-2 tile.",
"post_audit_fixes_round1": "Independent audit (fresh agent, not a fork) returned FAIL on 3 MAJOR + 13 MINOR; all corrected, signal unchanged at HOLD/HOLD/HOLD. MAJOR-1: relative strength vs SOXX was computed on WEEKLY sector bars (17 Jul / 15 May) against DAILY stock and index bars (14 Jul / 14 May); SOXX fell 7% between 14 and 17 Jul, so the 1-month cell was published as -4.06pp when it is +4.50pp. Corrected on matched daily closes; relative_strength_vs_sector -4.06 -> +4.50 and _3mo -6.17 -> -1.78; RS rescored 20 -> 48, risk_reward 41 -> 52, timing_score 63 -> 65 (later 64 in round 3 on a report-dated VIX), short composite 62 -> 63 (later 62). MAJOR-2: Intel 14A dates were wrong - the 0.9 PDK is targeted for OCTOBER 2026 (not Q1 2027), with risk production H2 2027 and HVM 2028, so the section-11 bear leg claiming 14A would take leading-edge share inside 12 months was mechanically impossible and is rescoped to design-win announcements. MAJOR-3: the ISW China-Taiwan update of 14 Aug 2026, which this report cites by name, also records a Guangdong MSA traffic-control scheme over northbound vessels in the Taiwan Strait 6-9 Aug that Taipei condemned - the exact category named in this report own falsification condition, omitted while asserting a bare no-interdiction. Now carried in Gate 5 and section 11 and tested explicitly; Gate 5 stays CAUTION because the notice ran four days under a typhoon pretext, had already expired by the report date (so there is no pending binary outcome for the gate to key on), and no stoppage or diversion of commercial traffic has been reported - ISW does not address whether any occurred, so the claim is the absence of a report, not the absence of the event. MINORs fixed: risk-reward ratio restated (1.12:1 base / 0.77:1 weighted), reported ROE 40.0 -> 34.9 percent on a period-end equity base, interest coverage 285.4x -> 886x recomputed, swing-low sequence corrected (the tool array is sorted by value not date), volume ratios restated against a 12.74m 20-session average - THIS CHANGE WAS WRONG AND ROUND 3 REVERSED IT: the original 0.74x (9.48m, 17 Aug, a COMPLETED session) was correct, and this round replaced it with 0.57x (7.25m) taken from a PARTIAL-session feed while labelling it in-progress, which contradicted the report own use of the 17 Aug close. See post_audit_fixes_round3, 16 Jul volume 1.68x not 1.8x, own-history percentile presented as 78th-87th across two availability conventions (this finding was WITHDRAWN by the auditor in round 2 - its own announcement-date reconstruction reproduced the original 11.8x/43.0x/median-27.8x/86.7th-percentile triple exactly; both conventions are retained as strictly more informative than either alone), sector-median lens restated clean-vs-clean (score 70 -> 65), hyperscaler-capex reconciliation corrected to name Oracle, breadth figures attributed to the macro report and cross-checked, Q3 earnings-date provenance downgraded to high-confidence-not-confirmed, risk-free deviation from the SKILL disclosed (macro state has no market_snapshot.UST10Y key), exercise date corrected to 12 Aug, forward dividend yield given as a 0.95-1.01 percent range. NOTE: this log originally also claimed Brent had been date-stamped; it had not, and round 2 caught that. It is stamped now.",
"post_audit_fixes_round2": "Round 2 returned FAIL on 1 MAJOR + 4 MINOR, every one the same failure mode - a round-1 correction that landed on one surface and left a duplicate elsewhere. MAJOR: the section-3 Competitive Environment callout still read Intel 14A PDK targeted for Q1 2027, one line above the corrected table saying October 2026 - fixed. MINOR: the Q3-earnings-date caveat existed in 1 of 4 places (Gate 2, the section-15 self-audit bullet, the WebSearch provenance note and next_update_basis all still asserted the strong version) - all four now aligned to high-confidence-not-confirmed. MINOR: the swing-low chronology was fixed in section 12 only; section 7 and section 9 still carried the tool value-sorted array read as a rising sequence - both replaced with the real chain off the 29 Jul low. MINOR: Brent was still undated in tails_ruled while post_audit_fixes_round1 claimed it had been date-stamped - now stamped (88.52 on the 14 Aug settle) and reconciled against the macro report 88.58 on its own earlier date. MINOR: the phrase diverted no traffic asserted an affirmative negative that ISW does not address; restated on all five surfaces as no REPORTED stoppage or diversion, with the source silence stated. The auditor WITHDREW two round-1 findings (Intel risk-production/HVM dates - its 2028/2029 came from Feb 2026 remarks superseded by the Q2 2026 call, so the report H2-2027/2028 dates are current; and the own-history percentile, which its announcement-date reconstruction reproduced exactly). It confirmed Gate 5 CAUTION is correct because the traffic-control notice had ALREADY EXPIRED, leaving no pending binary outcome for the gate to key on, and confirmed Medium 63 and Long 69 are genuinely unchanged by the Timing move (rounding absorbs it).",
"post_audit_fixes_round3": "Round 3 (FRESH auditor, not the round-1/2 agent - chosen because section 15 now contains long prose about the audit itself, which the previous auditor could not read impartially) returned FAIL on 4 MAJOR + 9 MINOR. Signal unchanged HOLD/HOLD/HOLD. MAJOR: (1) section 15 Confidence-impact still said the get_earnings_calendar gap was closed from the company own calendar - a SIXTH surface of a fix declared complete at five, and falsified three times in the same section; corrected. (2) The section-15 data-source row still recorded SOXX as WEEKLY bars - the provenance of the round-1 basis error; numbers were fixed in round 1, the provenance line was not, and it survived two sweeps because it contains no figures; corrected to daily. (3) NEW, missed by both earlier rounds: sections 9 and 12 said the MACD histogram was positive for four consecutive sessions and that the MACD line crossed above its SIGNAL on 13 Aug. Both wrong - the histogram has been positive for TEN sessions since 4 Aug (which is when the signal line was crossed) and 13 Aug was a cross above ZERO. (4) Two inputs stale against the report own date: the risk-free sensitivity claimed the whole plausible span while omitting the 17 Aug print of 4.72 percent (4.63 percent is the fortnight lowest), and VIX was scored at 14.25 from 14 Aug when 17 Aug is 15.19, above the report own risk-on threshold. VIX now report-dated: macro overlay 62 -> 56, timing_score 65 -> 64, short composite 63 -> 62. Risk-free stays 4.63 percent because the batch pinned one rate across all names, disclosed, with the full 4.63-4.72 span and its effect (warranted 27.70x -> 27.28x, ratio 1.23 -> 1.25, fair value 351 -> 346) now shown; band, gate and signal unchanged on every value. MINORS: 17 Aug volume restored to 9.48x-million = 0.74x after a round-1 fix wrongly replaced the correct figure using a PARTIAL-session feed while the report headline price is that session close; shelf levels dated (405.51 = 9 Jun, 407.99 = 6 Aug, 419.19 = 26 Jun) since they come from a value-sorted array; Gate 5 typhoon rationale hedged AT THE POINT OF RULING and the US State Department objection of 11 Aug 2026 added; Gate 3 now states its own-history arm (top 5 percent) is NOT met at the 87th percentile so the gate fires on the guardrail arm alone; cash quoted unrounded at 110.7 so the stated inputs reproduce net cash 77.1 and invested capital 125.3; section 9 corrected to the real 31 Jul close 404.25 and 50-day 425.33 (425.16 was the 14 Aug 50-day); Brent given its move and cause (about +8 percent that week to 88.53 on Reuters reporting of a US threat of an indefinite naval blockade of Iran); dnb_note now cites the standing precedent armed-not-triggering-tail-hold-not-dnb rather than asserting the distinction, and flags itself as a deterministic-gate opt-out. Also folded in: the guardrail-line rationale restated as the STRICTEST matching line (the rule the linter now applies) rather than the more specific one. The auditor OVERTURNED no earlier finding and independently reproduced the anchor, TTM roll, composites, relative-strength cells, swing-low chain, scenario arithmetic, Intel 14A dates and the 34.46x top-decile bar.",
"post_audit_fixes_round4": "Round 4 (a SECOND fresh auditor) returned FAIL on 4 MAJOR + 5 MINOR and moved NO number the signal depends on - all scores, band, gates, ladder and the three signals were reproduced independently and stood. MAJOR: (1) valuation_confidence 83 -> 78 had reached only 2 of 5 surfaces; still 83 on the section-1 scorecard card, the section-4 pillar header and the section-15 confidence paragraph, which published the correct arithmetic (15 points of haircut on a base of 93) beside the wrong total, while section 4 cited that very line as corroboration. Now 78 everywhere; overall confidence unchanged at 70 on the Timing floor. (2) The section-4 claim that scoring TSMC on a forward multiple made it THE OUTLIER was false and false in the opposite direction - a sweep of every tracked calibration finds about thirty names still anchored on a forward or NTM multiple, including ICE published in this same batch at this same stamp on the same pinned risk-free (actual 19.13 described as clean FORWARD P/E against warranted 20.37). TSMC prior forward anchor was the house norm. The E0 argument stands on its own arithmetic; the corroborating sentence was replaced by an escalation. (3) The risk-free span fix had reached 1 of 4 surfaces - the section-4 anchor-table row, section 6 and calibration risk_free_source all still published 4.63-4.70 percent. (4) A seventh VIX surface: section 7 confidence line still read no VIX penalty (14.25). MINORS: the 4.72 percent sensitivity arithmetic was the 4.70 percent result mis-transcribed - correct is warranted 27.28x, ratio 1.25, fair value about 346; the section-15 lesson block miscounted the defects and over-generalised, containing the very defect it described; the MU cell in the sector-median row is mid-cycle-normalised not clean and is now disclosed; the analyst-lens sentence stated the UPSIDES as the discount (price is 21-27 percent BELOW consensus; consensus is 27-37 percent above price); two per-ADR equity figures (1248.46 total book vs 1240.37 shareholders equity, 0.65 percent apart, consistent with non-controlling interests) are now reconciled in text; and this tails_ruled field had a duplicated clause and quoted Brent as both 88.52 and 88.53. ONE FINDING WAS NOT ACCEPTED: the auditor read the 20-session volume average of 12.74m as matching no window and offered 13.20m to 14 Aug. Recomputed from raw daily volumes the 20 sessions ending 14 Aug average 12.7377m - the label is correct, and 13.20m is a 21-session window. The figure was left alone and both windows are now stated (12.74m to 14 Aug, 12.63m to 17 Aug), because in round 1 a CORRECT volume figure was changed on an auditor prompt and had to be restored two rounds later.",
"changes_note": "Signal cut to HOLD on all three horizons. Driver of the change: a +6.3% price move plus the correction of a mixed forward/trailing valuation basis inherited from 31 July. No new earnings print. Business metrics improved across the board; the gate is on the price, not the company.",
"grid_cell": "EM Equities \u00b7 US",
"grid_status": "vacates \u2014 no live Short BUY (Short = HOLD, entry_conviction = Wait)"
}
get_stock_prices bar labels run a session early.auto_adjust=False. Used for the chart, the 60-month P/E reconstruction and the relative-strength table. SOXX is read off DAILY bars. This row previously said weekly — which was the provenance of the basis error the round-1 audit called MAJOR. The numbers were corrected in round 1; this line documenting where they came from was not, and survived two sweeps because it contains none of the figures a number-driven grep looks for.search_financial_news instead, which reached primary sources (TSMC's own press release for July revenue, TSMC's 2Q26 earnings release for the results and guidance, the Institute for the Study of War and the CFR conflict tracker for Taiwan). The Q3 earnings date is the one item that did not resolve to a primary source — see the get_earnings_calendar row above. WebFetch was also attempted and is restricted for the two news domains tried.Self-audit — provenance and live-status, evidence-cited.
get_income_statement decomposed; non-operating income is 8.0% of TTM net income (6.7% of pre-tax); P/E, PEG, net margin and ROE are all scored on operating earnings, which makes the multiple worse (34.0× vs 31.7× reported). Split stated in §4.totalDebt lease-inclusive: checked. It overstates borrowings, but TSMC is net cash by roughly US$77bn, so Gate 1 is unaffected either way. Stated rather than assumed.eps ≠ reported EPS: the TTM figure was rebuilt from four quarters of net income over the ADS count rather than taken from the field.get_stock_dividends, not a search summary. No raise is claimed; the sequential USD decline is identified as FX.auto_adjust=False). Immaterial at a sub-1% yield, applied anyway.get_earnings_calendar failed, so 15 Oct 2026 was sourced externally rather than carried forward — from a search citing TSMC's financial calendar, though TSMC's own calendar page currently lists only monthly-sales dates and third-party calendars mark the date unconfirmed. High-confidence, not confirmed; nothing in this report turns on a ±1 week error.short_entry_confirmed=false and signal_short=HOLD — consistent. The half-size quality-starter override was tested and correctly does not fire: it requires a non-Expensive valuation, and this name is Expensive, so short_hold_reason is expensive rather than neutral_timing_starter. Per the audit brief, an Expensive short HOLD is the correct output, not a mis-applied cap.Independent audit — round 1, and what it changed. A separate agent that did not write this report re-ran the live checks and the arithmetic from scratch. It reproduced every load-bearing number exactly — the warranted multiple to 27.7017×, the TTM roll to the last digit, the 16.157% effective tax rate, the clean EPS of US$12.684, the ratio of 1.2274, the fair value of $351.24, the three horizon composites, the implied-growth solve at 20.2%, the dividend history and the share-count reconciliation — and independently confirmed that the trailing valuation basis is the correct one under the framework and that the 31 July forward basis was the only one of its kind in the book. It returned FAIL on three material defects, all now corrected in this document:
Thirteen minor defects were also raised and fixed, including a risk-reward ratio that did not match its own two figures, a swing-low "rising sequence" that was the tool's value-sorted array rather than a chronology, a volume ratio — which this round got backwards, and round 3 reversed: the original figure was correct and the "fix" substituted a partial-session one (see the round-3 block below) — and a sector-median comparison that put this report's reported multiple against its peers' clean ones — the very mixed-basis error §4 exists to correct. Each is noted in place rather than silently repaired.
Independent audit — round 2. The same auditor re-checked the fix round and returned FAIL again, on one material defect and four minor ones — every one of them the same failure mode: a correction that landed on one surface and left a duplicate standing somewhere else. The material one was the worst kind: §3's Competitive Environment callout still carried the wrong Intel date one line above the corrected table that contradicted it. The others were the Q3-earnings-date caveat present in one of four places, the swing-low chain fixed in §12 but not in §7 or §9, an undated Brent figure that the fix log had already claimed was date-stamped, and — the most interesting — the phrase "diverted no traffic", which asserts something the ISW source is simply silent on. That last one is now stated as the absence of a report rather than the absence of the event, which is the weaker and truthful claim. All five are fixed on every surface, and this run's sweep was done by grepping each changed claim across the prose, the calibration and this block rather than checking only where the edit was made.
The auditor also withdrew two of its own round-1 findings after re-deriving them — the Intel schedule (noted above) and the own-history percentile, where its announcement-date reconstruction reproduced this report's original figures exactly (11.8× / 43.0× / median 27.8×, 86.7th percentile, decile 9). Both conventions are carried here because they answer the same question different ways, which is more informative than either alone. It confirmed Gate 5 at CAUTION rather than TRIGGERED for a reason worth recording: the traffic-control notice had already expired by the report date, so there is no pending outcome for a binary-event gate to key on, however serious the precedent.
Independent audit — round 3 (a fresh agent, not the round-1/2 auditor). A third pass returned FAIL on four material defects and nine minor ones. Three of the four were the same failure mode round 2 named — a correction landing on some surfaces and not others — and two of those were inside this very block, which is precisely why a fresh reader was used: the previous auditor would have been reading its own findings back.
Nine minor defects were also fixed, of which two are worth naming. The 17 August volume had been "corrected" in round 1 from 9.48m to 7.25m using a partial-session feed and labelled in-progress — while this report's headline price is that same session's close. The original figure was right: 9.48m, 0.74×. And Gate 3's own-history arm requires the top 5% of the five-year range; at the 87th percentile that arm is not met, so the gate fires on the sector-guardrail arm alone, which is now said rather than left ambiguous. The auditor confirmed the anchor, the TTM roll, the horizon composites, the relative-strength cells, the swing-low chain, the scenario arithmetic, the macro quotations, the Intel 14A dates and the DNB watch-item distribution — reproducing the 34.46× top-decile bar and the 86.7th percentile independently — and overturned no earlier finding — though it did reverse an earlier fix, restoring the 17 August volume that round 1 had replaced with a partial-session figure. The distinction is worth keeping: the round-1 finding (restate the ratio against a proper 20-session average) was right; its substituted value was not.
Independent audit — round 4 (a second fresh agent). A fourth pass returned FAIL on four material defects and five minor ones, and moved no number the signal depends on — every score, the band, both gates, the ladder and all three signals were reproduced independently and stood. Three findings are worth recording:
One round-4 finding was not accepted, and the reason matters more than the figure. The auditor read the 20-session volume average of 12.74m as matching neither window, offering 13.20m to 14 August. Recomputed from raw daily volumes, the 20 sessions ending 14 August average 12.7377m — the label is right; 13.20m corresponds to a 21-session window. The figure was left alone and both windows are now stated (12.74m to 14 August, 12.63m to 17 August), because in round 1 a correct volume figure was changed on an auditor's prompt and had to be restored two rounds later. An auditor's finding is evidence, not instruction.
What four failed audit rounds taught this report about sweeping — recorded because it is the actual lesson. Most of the material defects found after the first round were the same shape — a correction that landed on one surface and left a duplicate standing on another. An earlier version of this sentence said "every one of the six", which was wrong twice over: the count was five at the time (it is nine after a fourth round), and it was not universal — the paragraph nine lines above says three of the four, and the MACD error was a fresh factual mistake, not a propagation miss. The lesson block contained the very defect it was describing, which is the most honest illustration of the point available and is left on the record rather than tidied away. The obvious response is to grep for the value you changed, and that is what was done after rounds 1 and 2. It is not enough, and round 3 showed exactly why. The stalest line in the document said this report's relative-strength series was built off SOXX weekly bars — the precise provenance of the error that caused the first FAIL. It survived two sweeps because it contains no numbers at all. A value-driven grep is structurally blind to a sentence that describes where a figure came from, what units it is in, or which basis it uses, and those are exactly the sentences a reader leans on when deciding whether to trust a number. The generalisation, which is now the rule here: grep the claim's SUBJECT — "SOXX", "calendar", "MACD", "typhoon" — not only its value, and run the sweep in both directions between the prose and the calibration, because the same miss has appeared in both. The three loudest classes to sweep are the ones with no numeral in them: provenance ("sourced from…", "…weekly bars"), status ("closed", "verified", "armed", "triggered"), and completeness ("the whole span", "every ratio", "all four places").
A fourth class, added last and the subtlest: where a stated total has its components printed beside it, add them up. A confidence figure of 83% survived four rounds and a clean linter next to its own arithmetic — base 93, haircuts of 10 and 5 — because every individual number in the sentence was correct and only the sum was not. The same axis then caught a claim, written into this report while fixing that very defect, that two equity denominators "neither change ROIC to one decimal" — they give 52.5% and 51.9%, and the inputs were printed two lines above. A sentence asserting an invariant over figures it has just published is the highest-yield thing to check, and the author of a fix is the least likely person to check it.
Confidence impact. Overall confidence is the weakest pillar: 70% (Timing). Quality 75% carries a 5-point haircut for the free-cash-flow disagreement between providers; Valuation 78% carries 15 points of haircut (93 less 10 less 5) for the cross-currency ratio corruption and FX translation sensitivity; Timing 70% carries 5 for the FOMC minutes landing two days out. That is higher than the 58% recorded on 31 July, and the reason is that this run rebuilt the multiples, the five-year P/E distribution and the relative-strength series from primary series rather than accepting provider fields — more work, less inherited, narrower error bars. The one genuine gap is get_earnings_calendar, and it is NOT closed. The 15 October 2026 date rests on secondary reporting that cites TSMC's financial calendar; TSMC's own calendar page lists only monthly-sales dates and third-party calendars mark it unconfirmed. It is high-confidence, not confirmed, and nothing in this report turns on a ±1 week error. An earlier version of this sentence said the gap had been "closed from the company's own calendar", which three other rows in this very section falsify — a sixth surface of the same one-place fix, found by the round-3 auditor.