TSMC Is Quietly Building the Backbone of AI

TSMC Is Quietly Building the Backbone of AI

TSMC CoWoS Capacity Expansion — Overview (as of late 2026)

CoWoS (Chip-on-Wafer-on-Substrate) remains the dominant advanced packaging technology for high-end AI accelerators. TSMC has scaled it aggressively, but demand from NVIDIA and other AI/HPC customers has kept the market tight even as capacity multiplies.

Capacity Trajectory

Period Estimated Monthly CoWoS Capacity (TSMC) Notes
End 2023 ~13,000 wafers Severe bottleneck phase
End 2024 ~40,000 wafers Rapid early ramp
End 2025 ~70,000–80,000 wafers Continued expansion
End 2026 ~120,000–140,000 wafers (commonly cited ~130,000) Record levels; still short of demand
2027 Further growth (reports of >60% increase or ~170,000–200,000) Supply-demand gap expected to narrow
End 2028 ~260,000 wafers (some estimates ~270,000) Roughly double end-2026 levels

TSMC has stated a compound annual growth rate above 80% for CoWoS capacity from 2022 to 2027. Industry estimates (TrendForce, Economic Daily News, UBS, and others) consistently point to the end-2026 figure near 130,000 wafers per month and a doubling by end-2028. These are largely supply-chain and analyst figures rather than formal TSMC guidance.

A persistent supply-demand gap has been reported around 20% in 2026, with expectations it could narrow toward 10% by year-end 2026 and improve further in 2027 as new capacity comes online.

Expansion Methods

  1. Internal expansion: New and expanded lines concentrated in Taiwan (notably AP7 in Chiayi, plus sites in Tainan, Taichung, and Zhunan). TSMC is also building advanced packaging capacity in Arizona as part of its U.S. investment program.
  2. Outsourcing to OSATs: TSMC has increased outsourcing of CoWoS steps, particularly the wafer-on-substrate (WoS/oS) stage, and has reportedly expanded outsourcing of the chip-on-wafer (CoW) bonding step to partners such as ASE (including SPIL). This accelerates capacity without waiting solely on TSMC’s own fab buildout.
  3. Long-term partner agreements:
    • Seven-year deal with Xintec (TSMC-affiliated) for a dedicated CoWoS on-substrate line inside a TSMC facility in Taiwan’s Central Taiwan Science Park; mass production targeted for 2029.
    • Agreements involving GlobalFoundries for U.S. silicon interposers (ramp targeted ~2028).
    • Partnerships with Amkor (including Arizona capacity).
  4. Related technology roadmap: TSMC is producing ~5.5-reticle CoWoS packages today and has outlined a ~14-reticle version for 2028 capable of integrating roughly 10 large compute dies and 20 HBM stacks. Longer-term moves include CoPoS (Chip-on-Panel-on-Substrate) and System-on-Wafer approaches.

Customer Allocation and Demand Drivers

NVIDIA has been estimated to take roughly 50–60% of TSMC’s CoWoS capacity in 2026 (reports of ~800,000–850,000 wafers reserved for the year). Other significant users include Broadcom (custom ASICs for hyperscalers), AMD (Instinct series), AWS/Trainium-related designs, Marvell, and MediaTek. The top AI chip designers collectively absorb the large majority of available advanced packaging and HBM.

Larger package sizes and higher HBM stack counts per accelerator increase the effective demand per unit, offsetting some of the pure wafer-capacity growth.

Geographic and Strategic Context

Most CoWoS capacity remains in Taiwan. U.S. expansion (TSMC Arizona advanced packaging + Amkor Peoria) is progressing but will represent a minority of leading-edge output in the near-to-medium term. Intel’s EMIB-T is the main alternative architecture; estimates place its equivalent capacity well below TSMC’s even by 2028.

Summary

TSMC is executing a multi-year CoWoS expansion that has taken capacity from low-teens of thousands of wafers per month in 2023 toward ~130,000 by end-2026 and a targeted ~260,000 by end-2028. The ramp combines internal fab expansion, increased OSAT outsourcing (especially ASE/SPIL and Amkor), and new long-term partner lines. Despite this growth and a narrowing supply gap, demand from NVIDIA and other AI accelerator makers has kept CoWoS constrained. Larger future packages and continued AI infrastructure build-out are expected to sustain high utilization into the late 2020s.