Silicon Analysts
Supply Chain

Why CoWoS Lead Times — Not Wafer Capacity — Are the Real AI Bottleneck

By Silicon Analysts
4 min read
Foundry Economics

Executive Summary

The standard read on AI hardware supply — 'TSMC is adding CoWoS capacity, so the shortage will ease' — uses the wrong metric. Monthly capacity is rising fast (an estimated ~75–80k toward a 120–130k WPM target by end-2026), but it is being pre-booked faster than it comes online: an estimated 85%+ of 2026–2027 capacity is already locked, and NVIDIA alone reportedly booked more than half of the expansion. The result is that lead times stay at 52–78 weeks even as capacity nearly doubles. The metric that actually gates AI accelerator supply is the booking window — lead time over time — not instantaneous WPM, and almost no one publishes it as tracked data.

1Capacity is rising, lead times aren't falling. TSMC's CoWoS monthly capacity is ramping toward a ~120–130k WPM target by end-2026, yet both CoWoS-S and CoWoS-L remain fully booked at 52–78 week lead times.
2Pre-booking is why. An estimated 85%+ of 2026–2027 CoWoS capacity is already locked; NVIDIA reportedly booked >half of the expansion. New WPM is spoken for before it exists.
3Demand is outrunning supply. CoWoS demand roughly tripled — ~370k wafers (2024) → ~670k (2025) → ~1.0M (2026E) — so even a near-doubling of capacity doesn't close the gap.
4Watch the booking window, not the capacity headline. 'Capacity coming online' headlines mislead; the lead-time index is the leading indicator of when supply actually loosens.

Every few weeks brings a headline that TSMC is expanding CoWoS capacity, usually framed as relief for the AI hardware shortage. The framing is intuitive and mostly wrong — or at least, it watches the wrong number. The constraint on AI accelerator supply in 2026 is not how many packaging wafers TSMC can produce per month. It's how far out the booking window stretches. And the booking window is not shrinking.

Capacity Is Rising Fast — and It Doesn't Matter (Yet)

TSMC's CoWoS capacity is, by any historical standard, ramping aggressively. Monthly capacity is moving from roughly 75–80k wafers toward a target of 120–130k WPM by the end of 2026 — close to a doubling inside a single year.

Source: TrendForce / DigiTimes / Morgan Stanley, 2025–2026 (target estimates)

Yet across that ramp, both CoWoS-S and CoWoS-L have stayed fully booked, with lead times holding around 52–78 weeks. If capacity alone drove availability, a near-doubling would have visibly loosened the queue. It hasn't. That disconnect is the whole story.

The Mechanism: Capacity Gets Booked Before It Exists

The reason capacity additions don't translate into shorter lead times is pre-booking. An estimated 85%+ of TSMC's 2026–2027 CoWoS capacity is already locked, and NVIDIA alone reportedly committed to more than half of the expansion. When a customer books capacity that won't come online for four quarters, that future WPM is already spoken for the moment TSMC announces it.

So the marginal wafer of new capacity doesn't go to the back of a shrinking line — it goes to fill an order placed months ago. Demand makes this worse: CoWoS wafer demand has roughly tripled, from ~370k in 2024 to ~670k in 2025 to an estimated ~1.0M in 2026. A supply curve doubling while a demand curve triples does not produce slack.

This is why "capacity coming online" is a misleading signal in a pre-booked market. The number that captures the real state is the booking window — the lead time — because it encodes how much of future capacity is already committed, not just how much exists.

Lead Time Is the Leading Indicator

For anyone planning AI hardware — procurement teams, server OEMs, fabless chip companies, or investors modeling supply — the practical implication is to track lead time over time, not capacity announcements:

  • Lead time rising or flat while capacity grows = demand is still outrunning supply; the shortage persists regardless of expansion headlines.
  • Lead time trending down across consecutive periods = capacity additions are finally exceeding new bookings, or demand is cooling. That is the signal that supply is loosening — and it shows up in the booking window before it shows up anywhere else.

The catch: almost no one publishes lead time as tracked data. It lives in prose — "fully booked," "sold out through 2027" — as point-in-time snapshots that are stale the moment they're written. The same is true of the %-of-capacity-locked figure. These are exactly the two series that matter most for timing, and exactly the two that are least available as data.

Tracking the Booking Window

That gap is why we built the Allocation Dashboard to track allocation status, lead times, and %-locked by node and packaging technology as a time series — with the same data exposed through a machine-readable API (/api/v1/allocation) and the get_foundry_allocation MCP tool, every datapoint tagged with source, date, and confidence tier. For the current snapshot of who holds what across CoWoS, leading-edge logic, and HBM, see the companion Foundry Allocation Status tracker.

A caveat worth stating plainly: most of these figures are forward projections that trace to a single sell-side estimate reproduced across outlets. The point of a tracked, provenance-tagged series isn't to assert a precise number — it's to watch how the booking window actually evolves, which is the only honest way to call the turn.

The Bottom Line

CoWoS capacity is nearly doubling in 2026, and the AI hardware shortage is not over. Those two facts are only contradictory if you watch capacity. Watch the booking window instead: while lead times sit at 52–78 weeks and 85%+ of capacity is pre-locked, supply stays tight no matter how many expansion headlines arrive. The day the lead-time index starts falling is the day the constraint finally breaks — and that will be visible in the data long before it's visible in shipments.

References & Sources

<!-- Auto-populated from frontmatter sources by the rendering pipeline -->

Sources & Methodology

Data Verified PublicAll data sourced from public filings, press releases, and published reports

Methodology

This analysis is based exclusively on publicly available information including quarterly earnings calls, investor presentations, SEC/regulatory filings, published analyst reports, industry conference proceedings, trade publications, and government disclosures. All cost models use cross-validated benchmarks derived from these public sources. No proprietary, classified, or confidential information is used.

Public Sources

  1. [1]
  2. [2]
  3. [3]
    36Kr (citing Morgan Stanley). "Per-Customer CoWoS Allocation Split for 2026". 2026.
  4. [4]
    Silicon Analysts. "Foundry Allocation Dashboard + API". Jun 2026.

The views expressed on this site are my own and do not represent those of my employer. This is a personal research project for educational purposes. All data is sourced exclusively from public filings, press releases, and published industry reports. No proprietary or confidential information is used.

Related Analysis

Free Weekly Briefing

Weekly semiconductor analysis in your inbox

Get our weekly briefing with AI chip analysis, foundry updates, and supply chain intelligence.

View past issues & subscribe

Explore Our Tools