Thesis. Every major customer — NVIDIA, Broadcom, AMD, Arista, Alphabet — said in August that TSMC capacity is their binding constraint. CoWoS sold out through 2026, COUPE silicon photonics in production, a roadmap to 50x system compute by 2029. When every customer says the constraint is you, you own the constraint.
Reviewed 19 Sep 2026, after Q2 2026 results on 16 Jul and the customer commentary through early September. Next review after Q3 2026 earnings on 15 Oct 2026.
What it does
TSMC is a pure-play foundry: it manufactures chips that other companies design. It is neutral infrastructure — it builds every AI accelerator regardless of which designer wins. Roughly 72% of all foundry, 90%+ of leading edge, and ~95% of AI accelerators are made here. CEO C.C. Wei, Hsinchu, Taiwan.
Two terms carry the AI story. Nodes (N3, N2) are transistor generations; smaller means faster and more efficient, and TSMC's lead here is widening. CoWoS is advanced packaging — the step that bonds logic dies to high-bandwidth memory on one substrate. Every modern AI accelerator needs it, and it is the industry's actual bottleneck.
A structural quirk makes this page unusually useful: Taiwanese listing rules require TSMC to report revenue monthly. Everyone else reports quarterly, up to 45 days late. That makes TSMC's monthly release the highest-frequency AI demand indicator available anywhere.
Check TSMC's monthly revenue before reading any other category in this file. If it is decelerating, every average computed afterwards is measuring a receding tide.
Bull case
- Q2 2026 revenue US$40.20B, +33.7% YoY (+36.0% in NT$). Gross margin 67.7%, operating margin 60.3%, net margin 55.6%. EPS NT$27.25, +77.4% YoY
- Q3 2026 guided US$44.6–45.8B at 65–67% gross margin, with a steep 2nm ramp
- Near-monopoly economics: ~72% of foundry, 90%+ of leading edge, ~95% of AI accelerators. This is a moat, not a lead
- N2 in volume production in 2026 — the process lead is widening, not narrowing. Q2 wafer mix: 2nm 3%, 3nm 30%, 5nm 33%
- CoWoS and N3 sold out through end-2026, with lead times into 2027. Wei: "Our packaging capacity is so tight it is now limiting our customers' growth."
- COUPE silicon photonics in production, with co-packaged optics mass production from H2 2026, PIC capacity heading to 25K wafers/month by 2028, and a roadmap to 50x system compute by 2029
- Software-like margins from a manufacturer, a fortress balance sheet, a growing dividend, and FY26 capex of $52–56B funded internally
Bear case
- Taiwan geopolitics is a binary tail. A single headline can reprice the stock below $350, and no amount of operational excellence hedges it
- Customer concentration in HPC. HPC is the majority of revenue and the AI customers within it are few; an AI capex pause hits disproportionately
- Margin dilution is guided, not hypothetical: roughly 2–3% from the N2 ramp and another 2–3% from overseas fabs in FY26
- Overseas expansion carries real cost: Arizona labour, water and construction economics are worse than Taiwan's
- Export controls can restrict what TSMC may manufacture and for whom, independently of demand
- Still a cyclical industry, and margins near 68% gross are close to any reasonable ceiling
Major customers
- Apple — historically the largest single customer at roughly a quarter of revenue (verify the current figure). Leading-edge capacity is typically allocated to Apple first
- NVIDIA — the AI driver. Reportedly booked ~60% of CoWoS capacity through 2026, plus more than half of the 2026–27 expansion
- AMD — EPYC and Instinct
- Broadcom — the custom AI accelerators for Google, Meta, OpenAI and Anthropic are all fabricated here
- Qualcomm, MediaTek — mobile SoCs
- Marvell, Intel (partially), and essentially every fabless designer of consequence
Read-through — TSMC is the floor of the entire stack. Nearly every chip company in this file is a customer: NVDA, AVGO, MRVL and AMD all depend on its capacity, and ANET cited wafer fab shortages and 52-week-plus lead times as its own binding constraint. Most usefully, TSMC's VP of Advanced Packaging named the four remaining bottlenecks — lasers, optical fibres, fibre connectors and product testing — which maps directly onto LITE and COHR, GLW, APH and VIAV. That is the single most valuable sentence on this page: TSMC has told you where the money goes next.
What would change the view
Strengthens: AI capex continuing, N2 yields and pricing holding, CoWoS staying sold out, gross margin holding near 66%, US expansion progressing on plan.
Weakens: cross-strait escalation; a hyperscaler capex cut; Intel or Samsung winning a marquee node customer; an Apple or NVIDIA order air-pocket; the monthly revenue releases decelerating.
Specifically worth tracking: China–Taiwan headlines, hyperscaler capex guidance, margin against the guided N2 and overseas dilution, CoWoS capacity expansion pace, and the monthly revenue release.
Update log
19 Sep 2026 — Converted to the standard template. Brought the financials current from Q1 to Q2 2026 — the body had still been quoting Q1 revenue of $35.9B and a 66.2% gross margin. Q2 was US$40.20B at 67.7% gross and 55.6% net, with EPS +77.4%. Added the Q3 guide and the Q2 node mix. Removed the standalone thesis-tracker table and snapshot block, which duplicated the columns.
7 Sep 2026 — Verdict unchanged: Conviction. No new company print in that refresh; the value was the customer-side confirmation. NVDA (26 Aug) guided FY2028 growth of ~70% and called itself supply-constrained; AVGO (2 Sep) AI revenue +221%; AMD (4 Aug) datacentre +107%; ANET (4 Aug) cited wafer fab shortages and 52-week lead times; GOOG (22 Jul) "still in a supply-constrained environment."
31 Aug 2026 — TrendForce/TSMC presentation: SoIC + CoWoS to deliver 50x system compute by 2029; COUPE in production with CPO mass production from H2 2026; PIC capacity to 25K wafers/month by 2028; silicon photonics to exceed 50% of the transceiver market by 2027. NVIDIA and Broadcom named as early COUPE customers. Four remaining bottlenecks named: lasers, optical fibres, fibre connectors, product testing.
16 Jul 2026 — Q2 2026 reported. Revenue US$40.20B (+33.7%), gross margin 67.7%, operating margin 60.3%, net margin 55.6%, EPS NT$27.25 (+77.4%). Q3 guided US$44.6–45.8B at 65–67% gross margin.
Probabilities: strengthen ~62% / weaken ~32%, the latter carrying a fat tail — low odds, high severity, geopolitical.
Research and education only — not investment advice.
Made with Bullet