mstack

TSM

5 min read

TSM — Taiwan Semiconductor Manufacturing

TL;DR

TSMC is the bottleneck of the AI buildout. N3 capacity is structurally short — AI takes ~60% of 2026 N3 output and an estimated 86% in 2027 — and TSMC is “kingmaker” across Nvidia, AMD, Google TPU, AWS Trainium, Broadcom ASICs, Apple, Qualcomm, and MediaTek.1 Two anchor tenants (Apple + Nvidia) now fund the roadmap; A14 (1.4nm) is architected to serve both mobile and HPC, rebalancing node share back toward Apple.2 Near-term, effective N3 utilization >100% in 2H26 and CapEx smashing 2025 records.1

Business

Pure-play foundry. Revenue $122B in 2025, gross margin 59%+; N3 is the leading node for 2026 AI accelerators and premium smartphones, N5/N4 the volume workhorse, N2 ramps with Apple anchor, A16 (backside power) optimized for HPC, A14 targets both. Advanced packaging (CoWoS, InFO, SoIC) contributes 10%+ of revenue vs. 6-7% in 2018.2 HPC revenue mix grew from 36% (2020) → 58% (2025); smartphone shrank 46% → 29%.2

Thesis

  • Dual anchor tenants fund the roadmap. For a decade Apple alone funded TSMC’s leading edge (8K → 130K monthly wafers; 24B spend 2014-2025; >50% share at node launches). Nvidia’s AI cash now makes a second anchor — Nvidia consumes more N3 than Apple by 4Q27 per SemiAnalysis model.2
  • N3 is structurally oversold for the next two years. All major accelerator families (NVDA Rubin/4NP→3NP, AMD MI350X/MI400, Google TPU v7/v8, AWS Trainium3, Meta MTIA) converge on N3 in 2026. Effective utilization >100% in 2H26; cleanroom space, not capex, is the binding constraint.1
  • Kingmaker pricing power. AI customers get priority over smartphones; multi-year compute commitments from hyperscalers de-risk capex. Smartphone BOM cost inflation (memory-price led) likely to soften handset demand; freed N3 wafers go straight to accelerators.1
  • A14 rebalances back to Apple + HPC. TSMC is designing A14 for both mobile and HPC from the start, with a separate backside-power variant for HPC. Model implies Apple reclaims 67% node share on A14 — a reversal of the “Apple loses leverage at N2” narrative.2
  • Advanced packaging is a growing lever. CoWoS grew 14× 2018→2025 (8.4B). Apple InFO $3.5B+ in 2025. CoWoS outsourcing to Amkor/ASE/SPIL gives TSMC capacity flex without committing more capex.21

Risks

  • Capacity diversion to Intel / Samsung. AI demand so exceeds N3 that customers are exploring Intel Foundry (18A) and Samsung SF4/SF4X — Samsung has entered Nvidia’s datacenter supply chain (LP30/LP35) and Tesla AI5/AI6 dual-sourced TSMC+Samsung. Share erosion at the highest end is a tail risk.1
  • Geopolitics. Taiwan concentration risk remains the backdrop; Arizona ramp is real but not a capacity substitute.3
  • Memory choke point above foundry. HBM4 yield at ~11 Gb/s pin speed (SK Hynix leading, Micron lagging) can rate-limit accelerator output even with N3 wafers in hand.1
  • Smartphone recovery shifts mix back. If rising memory prices flow through and smartphones de-stock, recovery in 2027 could pull capacity back toward consumer ASICs — though in current setup this is a release valve for AI, not a threat.1

Recent catalysts

  • 2026-04-15 — ISSCC 2026: Nvidia + Broadcom CPO detail — validates TSMC CoWoS + optics packaging roadmap.4
  • 2026-04-15 — Chipstrat, Substrate — advanced substrate discussion, TSMC mentioned in supply chain context.5
  • 2026-04-10 — FundaAI, Preview: TSMC 1Q26 — continued front-end demand; Chinese demand resilience (CXMT DUV procurement); Intel/Samsung Foundry recovery underway.6
  • 2026-03-12 — SemiAnalysis, Great AI Silicon Shortage — N3 is “the biggest constraint”; effective util >100% in 2H26.1
  • 2026-01-08 — SemiAnalysis, Apple-TSMC: the partnership that built modern semiconductors — historical deep dive + 2030 outlook.2
  • 2026-01-01 — SemiAnalysis, CPO book scaling — TSMC packaging roadmap context.7

Second-order reads

  • 2026-04-13 — FundaAI, Preview: ASML 1Q26 — “TSMC is revising its EUV procurement volumes upward due to a shortage in advanced node capacity.” Positive for both TSMC (signals urgency on ramp) and ASML (EUV shipments).8
  • 2026-03-24 — SemiAnalysis, Nvidia Inference Kingdom Expands — Groq LP30 on Samsung SF4 (non-TSMC) is framed as incremental capacity Nvidia gets without touching TSMC allocation. Read: TSMC N3 is the binding resource Nvidia jealously guards.9
  • 2026-02-09 — SemiAnalysis, CPUs are Back — hyperscaler ARM-based CPU rollouts (Graviton, etc.) add another class of N3 customer.10
  • 2026-01-13 — SemiAnalysis, Interconnects Beyond Copper — TSMC CFET made “much further progress than anyone expected” at IEDM; 2D FET MoS₂/WSe₂ GAA research disclosed. Pre-commercial but relevant to sub-A14 node timeline confidence.11
  • 2026-01-04 — Irrational Analysis, Fabulous Failures — contrasts TSMC’s execution against Intel/Samsung stumbles, reinforces the “TSMC as kingmaker” thesis.12

Valuation & positioning

Pending. The corpus argues TSMC is volume-constrained not demand-constrained through 2027 — multiple expansion case if utilization + capex translate to margin leverage. Need a specific earnings print / guide to anchor a number.

Sources

INTC — competing foundry; potential overflow beneficiary TSEM — specialty foundry (not direct competitor, complementary) ASML AMAT LRCX KLAC ONTO CAMT — equipment pull-through from TSMC capex NVDA AMD AVGO MRVL — N3 customers, the fuel behind the shortage ARM QCOM — mobile leading-edge customers MU SNDK — memory is the adjacent choke point above TSMC TTMI — advanced substrate/PCB pull-through on CoWoS

Footnotes

  1. SemiAnalysis — The Great AI Silicon Shortage — 2026-03-12 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9

  2. SemiAnalysis — Apple-TSMC: The Partnership That Built Modern Semiconductors — 2026-01-08 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  3. SemiAnalysis — TSMC Overseas Fabs: A Success — 2025-12-01 ↩

  4. SemiAnalysis — ISSCC 2026: Nvidia and Broadcom CPO — 2026-04-15 ↩

  5. Chipstrat — Substrate — 2026-04-15 ↩

  6. FundaAI — Preview: TSMC 1Q26 — 2026-04-10 ↩

  7. SemiAnalysis — Co-Packaged Optics (CPO) Book Scaling — 2026-01-01 ↩

  8. FundaAI — Preview: ASML 1Q26 — 2026-04-13 ↩

  9. SemiAnalysis — Nvidia: The Inference Kingdom Expands — 2026-03-24 ↩

  10. SemiAnalysis — CPUs Are Back: The Datacenter CPU — 2026-02-09 ↩

  11. SemiAnalysis — Interconnects Beyond Copper: 1000 Transistors to 10 Trillion — 2026-01-13 ↩

  12. Irrational Analysis — Fabulous Failures — 2026-01-04 ↩

Graph View

Backlinks