AI Chip Bottleneck Shifts to Packaging: CoWoS Pricing, Lead Times
For two years, AI chip capacity was mostly constrained by wafer manufacturing; in the second half of 2026, the bottleneck has moved more clearly to advanced packaging. High-end packaging such as CoWoS is seeing planned price hikes and longer equipment lead times, while demand has spread from a few large GPU customers to multiple cloud vendors' custom chips. For teams building AI hardware and solutions, packaging capacity is becoming a scheduling constraint as important as wafer tape-out.
Price and lead time rise together
A BOC International report on November 11 noted that foundries and OSATs such as TSMC are preparing another round of price increases, with CoWoS packaging quotes planned to rise about 10%–20%;associated chip equipment lead times have also stretched from the previous 3–6 months to 6–8 months. Price and lead time rising together usually means one thing: capacity still trails demand. Even where expansion is underway, customers must pay more and wait longer to secure a firm slot.
Public estimates let us check the capacity ramp. Industry estimates put TSMC CoWoS monthly capacity in the low thirty-thousands of wafers in 2024, about 70,000 by end-2025, above 130,000 by end-2026, and over 200,000 expected for 2027. The ramp is not slow, but larger AI chip package areas and higher HBM content keep raising packaging intensity per device, so released capacity still struggles to match demand.
Demand spreads from one GPU customer to many
The advanced packaging demand base is widening. DIGITIMES reported in mid-September that TSMC CoWoS capacity remained short of demand through the end of Q3 2026; alongside high demand from a leading AI GPU maker, custom AI chips from AWS, Google and other cloud vendors keep adding new demand. The shift from a single large customer to multiple customers is a key reason the tight supply persists.
Yole Group forecasts the advanced packaging market could surpass traditional packaging for the first time in 2026, with related capex approaching USD 28 billion. Reuters reported on November 8 that Nvidia plans to invest billions of dollars to support ASE subsidiary SPIL in expanding CoWoS capacity, funding cleanroom and equipment build-out. ABF substrates in the packaging chain also face shortage and price pressure, and some competing packaging routes are likewise constrained by substrate supply.
| Dimension | Earlier stage | 2026 status |
|---|---|---|
| Demand source | Mainly leading AI GPU customers | Beyond GPUs: cloud custom chips and network chips join |
| Capacity constraint | Wafer tape-out pace was the main constraint | Packaging capacity, substrates and equipment lead times jointly limit volume |
| Supply action | One main packaging channel | Chip makers back multiple OSAT sites and qualify second sources |
Implications for the supply chain and solutions
Chip / hardware solution side
- Include packaging capacity and equipment lead times in project scheduling; don't treat tape-out as the only constraint
- Confirm packaging slots early for high-end chips; keep a backup packaging channel for critical projects
- Watch substrate and adjacent supply risks, so a single packaging choke point does not stall the whole device
Procurement / selection side
- Track advanced packaging quotes and lead times; leave room for packaging price rises in cost estimates
- Differentiate packaging routes by maturity and capacity availability, not just nominal chip specs
- Follow how 2027 new capacity lands to judge when the tightness eases
Overall, AI hardware capacity bottlenecks are spreading from wafer manufacturing to packaging, substrates and equipment supply. High-end packaging will likely stay tight in the near term, with price and lead time the most direct indicators of supply and demand. Going forward, track the release of new packaging capacity and how cloud custom-chip volume keeps pulling packaging demand, rather than reading only a single customer's order swings.

Comments
No comments yet. Be the first to comment.