✦ Packaging: CoWoS, CoPoS & the Second MoatNarrow moat

TSMC (TSM) — the future bets

The AI bottleneck moved from transistor to package, so TSMC is quadrupling the chokepoint it already owns.

Every AI accelerator that matters is not just fabbed but assembled by TSMC — GPUs and memory stacked on interposers through CoWoS — and that assembly step has been the binding constraint of the AI boom. The response is a second moat under construction: capacity driving toward roughly 130,000 wafers a month by late 2026, nearly four times late-2024 levels, with the supply-demand gap narrowing from about 20% to 10% by year-end1. Beyond CoWoS waits CoPoS — panel-level packaging with pilot production targeted for 2027 — and an overflow system routing hundreds of thousands of wafers to partners like Amkor and SPIL.

CoWoS monthly capacity, wafers~33KLate 2024~130KLate 2026 (target)Supply-demand gap guided from ~20% to ~10% by end-2026; CoPoS panel pilot 2027; Amkor/SPIL overflow.
Quadrupling the chokepoint: the assembly step that rationed the AI boom becomes the second moat — sold out before it is built.

The wager inside the expansion: packaging is where TSMC's process lead converts into system-level lock-in, because a customer who fabs at TSMC and packages at TSMC has nowhere else to take the finished design. The risk is the classic capacity trap — quadrupling a chokepoint is glorious until demand pauses, at which point it is empty cleanrooms at $60-64 billion of annual capex2. Watch the gap closing on schedule and utilization staying tight as the new lines land.

Moat trajectory: Widening

Widening at the chokepoint: capacity toward ~130K wafers a month, the supply gap halving toward 10%, panel-level CoPoS piloting next. Packaging is where the process lead becomes system lock-in — and the expansion is the rare kind whose demand is already sold out.

The number that tests this moat
Reported
CoWoS capacity, late 2026
~130K wafers/month (~4×)

The AI bottleneck migrated from transistor to package, and TSMC is quadrupling the chokepoint it owns — with the supply-demand gap guided to halve toward 10% by year-end and CoPoS panels piloting in 2027. The trap is overshoot: watch utilization as the lines land, because an empty advanced-packaging cleanroom is the expensive kind.

Source: TrendForce CoWoS coverage (Jun 2026) ↗
References
  1. ReportedCoWoS toward ~130K wafers/month late 2026; gap narrowing 20% → 10%; CoPoS pilot 2027; Amkor/SPIL overflow.
    TrendForce — TSMC CoWoS capacity expansion (Jun 15, 2026): toward ~130K wafers/month by late 2026 (~4x late-2024); supply-demand gap narrowing from ~20% to ~10% by end-2026; OSAT overflow to Amkor/SPIL; CoPoS panel pilot targeted 2027 — 2026 · publ. Jun 15, 2026 · source ↗
  2. Reported2026 capital spending guided to $60–64B.
    TSMC 2Q26 earnings call transcript (2026 capital budget raised to US$60-64B); the ~US$20B cost of one leading-edge fab is an industry estimate — 2026 guidance · publ. July 16, 2026 · source ↗
Sources
Generated September 23, 2026