TSMC

Phase A — Understand the Business

Lens 1 · Company Overview

The world's only pure-play foundry at the leading edge at scale — it manufactures chips it does not design, for customers who own no fabs. That separation is the thesis: TSMC sits one layer below every AI-chip brand and collects the manufacturing rent regardless of which designer wins. Q1 2026 revenue US$35.90B, +40.6% YoY, beat. Advanced nodes (≤7nm) 74% of wafer revenue, 3nm 25%. Engine: (1) leading-edge logic (N3 ramping, N2/GAA in production early 2026, Apple leading), and (2) CoWoS advanced packaging — the literal throughput governor on AI accelerators. Pricing power is now overt: 3–5% sub-5nm wafer price hikes in 2026, N2 wafers ~$30,000 (~20% premium to N3). Customers — Apple, NVIDIA, AMD, Broadcom, MediaTek — absorb it because there is no leading-edge alternative. Gross margin 66.2% (Q1'26) — a number that used to belong to fabless designers.

Lens 2 · Supply Chain

The chokepoint with its own upstream masters. ASML is the hardest dependency (EUV monopoly; High-NA for sub-2nm from ASML alone); deposition/etch from Applied Materials/Lam; metrology from KLA; resists/chemicals from Japan (JSR, Tokyo Ohka, Shin-Etsu) — a serial chain where TSMC pressures customers but not ASML. Downstream: the entire fabless universe + hyperscaler custom silicon (via Broadcom/Marvell). NVIDIA reportedly secured >60% of 2026 CoWoS capacity for Rubin; TSMC scaling CoWoS toward 130–150k wafers/mo by late 2026 (roughly doubling). Geographic diversification (Arizona N4/N3→N2, Kumamoto Japan, planned Dresden) is real but a margin headwind, not a tailwind — overseas fabs run structurally higher cost; leading-edge density/yield still live in the Taiwan cluster.

Lens 3 · Competitive Advantages

One of the widest moats in public markets, widening. Process leadership: a clear node+ ahead — 3nm at volume, N2/GAA on schedule, while Intel Foundry burns cash proving 18A/14A and Samsung Foundry fights yield. First 2nm customers: Apple, AMD, NVIDIA, MediaTek — Intel absent; the merchant leading edge is consolidating onto one vendor. Deeper moat: CoWoS near-monopoly + yield-at-scale + customer trust — yield compounds (more wafers → more learning → higher yield → lower cost → more customers). Capital scale ($52–56B 2026 capex) out-invests every rival on nodes only it can fill. Bargaining power over fabless customers now explicit (multi-year price hikes, pre-paid capacity booked to 2028). Weaker only upstream vs ASML and laterally vs the HBM makers who gate the other half of an accelerator.

Lens 4 · Segments

By platform (Q1'26): HPC 61% (was 58% FY25 — structurally overtook smartphone), smartphone 26%, rest IoT/auto/DCE. By node: ≤7nm 74% of wafer revenue, 3nm 25%, N2 entering. By geography: North America dominates end-customer revenue (US-HQ'd fabless), China/APAC shrinking under export controls. The cross-company takeaway: TSMC's mix is now more AI-levered than most "AI plays" — HPC is a growing majority — while the customer base is diversified across NVIDIA/AMD/Broadcom/hyperscaler-ASIC rather than any single brand.

Phase B — Measure Performance

Lens 5 · Earnings Result (Q1 2026)

Revenue US$35.90B, +40.6% YoY, beat. Gross margin 66.2%, above guidance — advanced-node mix + utilization, despite overseas-fab dilution. Net margin ~46.5%; ROE ~36%. Q2'26 guide US$39.0–40.2B, GM 65.5–67.5%. FY26 guide: revenue >30% USD growth, capex toward the high end of $52–56B (vs $40.9B 2025, ~+30%), >70% to advanced nodes. The signal: margins went up into a heavy overseas-buildout year — the opposite of the bear case — and capex was raised because AI demand pulls 2027–28 capacity forward. The cleanest evidence in the complex that AI demand is supply-constrained, not demand-questionable. TSM up ~150%+ over the run its chairman cited (June 2026).

Lens 6 · Earnings Calls (sentiment trend)

From "strong demand" (2024) to "insane" AI demand (2026), C.C. Wei framing AI as a multi-year structural pull, 2nm booked to 2028. Threads: (1) AI/HPC as primary capex driver; (2) CoWoS expansion (doubling, tied to Rubin); (3) overseas-fab cost/margin dilution acknowledged candidly — CFO Wendell Huang guides "high-20s% ROE through the cycle." Most confident in company history while still disciplined about the cost of globalization — selling a long demand curve, not a quarter.

Lens 7 · Comps

Lens 8 · Stock-Price Catalysts

Three families: AI demand/guidance (up) — capex raises + >30–40% growth guides re-rated it to ~$1.9T (chairman cited >150% run). Taiwan/China geopolitics & export controls (down) — the largest down-catalyst family; cross-strait/export headlines trigger 5%+ risk-off (Mar 2026 TSM −5.5% on geopolitics, not fundamentals). AI profit-taking (down) — high-beta to sector de-risking. Net: up-moves are company-specific (demand/margins); down-moves are mostly exogenous (geopolitics). That asymmetry is the "Taiwan discount" — the business keeps beating, the stock gets marked down on things TSMC doesn't control.

Phase C — Judge People & Books

Lens 9 · Management

C.C. Wei holds CEO + Chairman (succeeded Mark Liu; June 2026 AGM re-seated the board with Wei in command of US/Japan/Europe expansion while protecting the Taiwan core). Post-Morris-Chang professionalized governance — founder culture (manufacturing discipline) intact, run by career engineers, not a founder-owner: capital discipline is institutional, not personal. Cleanest capital allocation in the sector: pour into advanced-node + CoWoS capacity where it's pre-sold ($52–56B capex, >70% advanced), fund from operating cash (effectively net cash), return a rising dividend (+~28–33% to ≥NT$23–24/share, yield ~0.62%). Low yield is intentional — capex is the better use while demand is supply-constrained. Red flags modest: dual CEO/Chairman reduces board independence; capex committed years ahead of revenue (execution risk if AI softens); leadership/assets Taiwan-domiciled. Concentration risks, not scandals.

Lens 10 · Forensic Red Flags

• Capex/depreciation cycle — $52–56B/yr (from ~$41B) steps up depreciation 2026–28; today's mid-60s% GM is earned while the D&A wave builds. If utilization dips after capacity lands, margin math inverts fast — the single most important forensic variable.
• Overseas-fab margin dilution — management guides several points of multi-year GM dilution (Arizona/Japan/Europe), front-loaded as fabs ramp under-utilized. Real, disclosed, recurring.
• Customer concentration — Apple 48–50% of N2 allocation; NVIDIA the largest revenue customer ($33B 2026E vs Apple ~$27B); two customers plausibly ~35–40%+ of revenue, mitigated only because both are captive.
• Geopolitical contingency (20-F) — the Taiwan risk factor is the dominant disclosed contingency and the explicit reason the multiple sits ~22× vs customers at 38–55× (analysts frame a ~50% geopolitical discount). Export-control/China a secondary risk.
• Related-party/accounting — none material surfaced; n/a at filing-detail level (20-F not retrieved this pass).

Phase D — Project & Stress-Test

Lens 11 · Forward Projection (EPS, next 3 fiscal years)

Bottom-up; calendar-year; TSMC reports TWD, EPS here in USD-ADR terms (1 ADR = 5 shares), approximate.

Lens 12 · Bull vs Bear

Bull. TSMC is the one name in the entire AI complex that wins regardless of which accelerator brand wins — NVIDIA, AMD, Broadcom, Google's TPU, Apple, every hyperscaler ASIC is fabricated and CoWoS-packaged by TSMC. It owns the two literal chokepoints (leading-edge wafers + CoWoS), it's raising prices into a sold-out book, margins are rising into the buildout, and it trades at 22× forward — a discount to every customer it supplies. This is the rare case where the highest-quality, most-diversified, most-defensible business in a value chain is also the cheapest. Bear (impairment vectors). (1) Taiwan — a cross-strait conflict or blockade is a permanent impairment no DCF survives; this is the real reason for the discount and it is genuinely binary. (2) Customer concentration (Apple + NVIDIA ~35–40%) — captive today, but a single redesign or in-sourcing is a swing. (3) The capex/depreciation wave + overseas-fab dilution compress margins if AI utilization disappoints after the capacity lands. Pre-mortem (2027): not a business failure — a geopolitical event. A Taiwan blockade scenario, or even a credible escalation, re-rates the multiple to crisis levels regardless of the (excellent) fundamentals; alternatively, AI demand digests just as the $52–56B capex converts to depreciation, and 66% margins fall to high-50s on under-utilized overseas fabs. Contrarian view: the market obsesses over which AI brand wins (NVIDIA vs Broadcom vs custom) and pays $5T for the leader — while the supplier that monetizes all of them, with no competitor at the leading edge, trades at a discount to all of them. The variant perception: the safest, highest-quality AI exposure is the one nobody calls an "AI stock," priced cheap for a tail risk (Taiwan) that is real but binary and hedgeable — not for any business weakness.

Lens 13 · Devil's Advocate

The discount exists for a reason: the entire company — leadership, leading-edge fabs, yield knowledge, 90% of value — sits on one island that is the most likely flashpoint for great-power conflict this decade. No amount of Arizona/Japan/Dresden diversification changes that for a decade-plus; those fabs are higher-cost and lower-yield and are a margin drag, not a hedge. TSMC commits $52–56B/yr years ahead of revenue on the bet that AI demand compounds — if it digests, it's the most capital-intensive name in the complex holding the bag, with a depreciation wave landing into softening utilization. And two customers (Apple, NVIDIA) are ~35–40% of revenue — captive until they aren't. The "cheap" 22× forward is the market correctly pricing a real, binary, un-diversifiable geopolitical tail.

Lens 14 · Management Questions

1. What's the realistic gross-margin floor if AI utilization dips after the $52–56B capex lands and overseas fabs are ramping? 2. Apple + NVIDIA combined % of revenue, and the redesign/in-source risk? 3. CoWoS capacity vs committed demand through 2027 — sold out, or building ahead? 4. Overseas-fab margin dilution path (how many points, how many years)? 5. N2 pricing premium durability vs Samsung/Intel if they close the gap? 6. Depreciation step-up 2026–28 and its GM impact at flat utilization? 7. Concrete continuity plan for a Taiwan disruption (fab mirroring, IP, people)? 8. China/export-control revenue exposure trajectory? 9. Capital-return policy if AI capex needs ease? 10. High-NA EUV timing and its effect on sub-2nm cost/roadmap?