Financebotresearch desk研究台

AI & Infrastructure CapexAI 与基础设施资本开支

TSM

TSM 6.5

Price and target zones价格与目标区间

$417.17
below低于 $340–400 entry买入

What the company does公司做什么

No company description is on file. Run python .mcp/profiles.py to get one.

Investment thesis投资逻辑 from the watchlist来自关注列表

What's working

  • The business is the best on this entire watchlist and is still improving — foundry share 70.4% → 72.3% in one quarter (Samsung 6.5%), ROIC 21.7% → 26.3% while invested capital grew NT$5.26T → NT$6.39T (more capital at a higher return), gross margin 56.1% → 59.9% → 67.7% (Q2), incremental operating margin 67.1%. Q2 revenue $40.2B +36% YoY, EPS +77.4%. Net cash ~$54B, debt/assets 13.4%, interest coverage ~157x, share count dead flat four years
  • The adversarial stress-test drew no blood — the rarest outcome this framework issues: Intel has ZERO committed external customers on 18A and 14A (two prospects, decisions H2'26-H1'27; reports it may axe 18A for foundry entirely); Samsung hit a real 70% SF2P yield but converts to 6.5% share; no hyperscaler is building a leading-edge fab. N2 has 4x N3's tape-outs at the same stage (>20 received, >70 pipeline, already 3% of wafer revenue), and TSMC pre-announced price increases up to 10% for 2027

Red flags

  • The "19.3x fwd" was a wrong-fiscal-year vendor figure. Yahoo's forward EPS implied ~$21.60; the actual 2026 consensus is $15.91, so the true multiple is 26.2x — confirmed by the company's own prints (Q1 $3.49/ADR + Q2 US$4.31/ADR = ~$7.80 H1, which cannot annualise to $21.60) and by three independent vendors at 24.5-25.3x. Yahoo was the outlier, not the consensuspitfall-vendor-forward-eps-is-the-wrong-fiscal-year. FV $432-518 → $350-480, entry ≤$400 → ≤$360, ACCUMULATE → WATCH. Spot is AT fair value, not below it.
  • The exposure is concentration, not competition: Nvidia is now 19% of revenue (was 12%), Apple 17% (was 22%) — and Nvidia holds ~60% of 2026 CoWoS output with the top 3 at >85%. Growth is levered to one customer's capex plan
  • TSMC is building away its own scarcity — CoWoS 35K→75K→125-130K wpm target, supply gap narrowing 20% → 10% by end-2026
  • Three of four classical models say expensive — DYT $246 (0.92% yield vs 1.55% 5yr avg), DDM $175, Graham $95, P/FCF ~69x. All three are the wrong tools (27.6% payout, 56% of OCF reinvested at 26.3% ROIC), but "the models are wrong" ≠ "the stock is cheap" — the buyer here leans entirely on the earnings case
  • Taiwan is not a drawdown a fair price compensates — it is a binary to ~zero, and no entry price fixes it. The margin of safety can only ever be business quality → argues for a smaller position, not for paying any price.

Watch

  • ANALYZED 8/04 — first file ever. DOWNGRADED [7.0] → [6.5] within hours of the sweep that added it, and the reason is a data error in the sweep itself.
  • Dividend-grower overlay (the sweep missed this): 10yr div CAGR +15.7%, but the USD dividend went NOWHERE 2021-23 ($1.892→$1.856→$1.847) on TWD weakness — a US holder's income is a TWD dividend with an unhedged FX overlay.
  • Mean tgt $540 (+29.5%), every post-Q2 action a RAISE, zero downgrades in 8 months, most cautious live rating a Hold at $440 — treat unanimity as risk, not confirmation
  • Ignore reported P/B 87 and Graham $35.44 (ADR unit artifacts; real P/B ≈12.8x) → pitfall-adr-book-value-corrupts-price-to-book; roic.ai's per-share fields are ALSO wrong on this name (4.669B divisor, matching neither the 25.93B ordinary nor 5.186B ADR count) — the two vendors fail differently, so cross-checking them does not surface it
  • Sleeve placement is deliberate — it passes the "does a datacenter pause hurt this?" test, so it belongs here, which means TSM is NOT diversification against AVGO/KLAC/AMAT/MU/ANET; it is the largest and highest-quality expression of the same bet, sharing customers with AVGO. Entry ≤$360 (≈22.6x), strong <$320 (≈20x). Trim 32x fwd. Moat 9/10 — highest issued; fundamentals 9.5/10. At $360 this is a [7.5]; at $320 an [8.0]. At $417 it is hold-your-fire.

📋 🔴 8/04 分析 —— 史上第一份报告。信心度从 [7.0] 下调至 [6.5],而且是在把它加入自选的同一天之内,原因是那次扫描本身的一处数据错误。 🚩 所谓"19.3 倍前瞻市盈率"是一个财年错配的供应商数据。 雅虎的前瞻每股收益隐含约 $21.60;而 2026 年的实际市场一致预期是 $15.91,因此真实倍数是 26.2 倍 —— 这一点已由公司自己的财报证实(第一季度 $3.49/ADR + 第二季度 US$4.31/ADR = 上半年约 $7.80,不可能年化到 $21.60),并有三家独立数据商给出的 24.5-25.3 倍佐证。雅虎才是异类,不是市场一致预期pitfall-vendor-forward-eps-is-the-wrong-fiscal-year。合理价值 $432-518 → $350-480,入场 ≤$400 → ≤$360增持 → 观望。现价正处于合理价值,而非低于它。这门生意是整个自选清单里最好的,而且仍在变得更好 —— 代工市占率单季度从 70.4% 升至 72.3%(三星 6.5%),投入资本回报率从 21.7% 升至 26.3%,同期投入资本从 NT$5.26T 增至 NT$6.39T(更多的资本、更高的回报率),毛利率 56.1% → 59.9% → 67.7%(第二季度),增量营业利润率 67.1%。第二季度营收 $40.2B,同比 +36%,每股收益 +77.4%。净现金约 $54B,负债/资产 13.4%,利息保障倍数约 157 倍股本连续四年纹丝不动。✅ 对抗性压力测试没有见血 —— 这是本框架给出过最罕见的结论:英特尔在 18A 和 14A 上都是零个已承诺的外部客户(两家潜在客户,决策在 2026 下半年至 2027 上半年;有报道称它可能干脆放弃面向代工客户的 18A);三星确实做到了 70% 的 SF2P 良率,但只转化为 6.5% 的市占率;没有任何一家超大规模云厂商在建先进制程晶圆厂。N2 在同一阶段的流片量是 N3 的 4 倍(已收到 >20 个,管线中 >70 个,已占晶圆收入的 3%),而台积电已提前宣布 2027 年最高提价 10%。🚩 风险敞口在客户集中度,而不在竞争:英伟达现已占营收 19%(此前 12%),苹果 17%(此前 22%)—— 且英伟达锁定了 2026 年 CoWoS 产出的约 60%,前三大客户合计 >85%。增长已经绑定在单一客户的资本开支计划上。🚩 台积电正在亲手拆掉自己的稀缺性 —— CoWoS 产能 35K→75K→目标 125-130K 片/月,供需缺口在 2026 年底从 20% 收窄至 10%。🚩 四个经典模型里有三个说贵 —— 股息率理论 $246(0.92% 股息率 vs 五年均值 1.55%)、股息贴现模型 $175、格雷厄姆 $95、市现率约 69 倍。这三个都是错误的工具(派息率 27.6%,经营现金流的 56% 以 26.3% 的投入资本回报率再投资),但"模型是错的"不等于"股票便宜" —— 在这个价位,买入者完全依赖盈利增长这一条腿。⚠️ 股息增长型叠加分析(那次扫描漏掉了):十年股息年复合增长 +15.7%,但美元股息在 2021-23 年原地踏步($1.892→$1.856→$1.847),原因是新台币走弱 —— 美国持有人拿到的是一份带着未对冲汇率敞口的新台币股息。⚠️ 平均目标价 $540(+29.5%),第二季度财报后每一次评级动作都是上调,八个月内零下调,最谨慎的在评级是 $440 的持有 —— 把这种一致性当作风险,而不是确认。❌ 忽略报告中的 87 倍市净率和 $35.44 的格雷厄姆值(ADR 单位换算的伪影;真实市净率约 12.8 倍)→ pitfall-adr-book-value-corrupts-price-to-bookroic.ai 在这只票上的每股数据同样是错的(除数为 4.669B,既不等于 25.93B 的普通股数,也不等于 5.186B 的 ADR 数)—— 两家数据商以不同的方式出错,所以互相交叉验证并不能发现问题。🚩 台湾不是一个合理价格可以补偿的回撤 —— 它是一个归零的二元风险,任何入场价都无法解决。这里的安全边际只可能是企业质量 → 这意味着应该降低仓位,而不是意味着可以在任何价格买入。⚠️ 归入本篮子是刻意为之 —— 它通过了"数据中心建设暂停会不会伤到它"的测试,所以它属于这里,这也意味着 TSM 并不能分散 AVGO/KLAC/AMAT/MU/ANET 的风险;它是同一笔押注中规模最大、质量最高的表达形式,并且与 AVGO 共享客户。入场 ≤$360(约 22.6 倍),强力买入 <$320(约 20 倍)。减持 32 倍前瞻。护城河 9/10 —— 迄今给出的最高分;基本面 9.5/10。在 $360 它是 [7.5];在 $320 是 [8.0]。在 $417,按兵不动。

Competitive advantages竞争优势 from the thesis and the reports来自投资逻辑和报告

Sources来源