Explosive, sustained demand — Hyperscaler and enterprise AI spending (training + inference + agentic workloads) keeps rising faster than capacity. Demand now spans GPUs, custom ASICs, server CPUs, and related chips.
HBM memory is the primary bottleneck — Samsung, SK hynix, and Micron have largely sold out 2027 HBM output (and much of 2028). HBM production crowds out regular DRAM (roughly 3+ bits of standard DRAM sacrificed per HBM bit). Shortages are expected to persist well beyond 2027, potentially into 2028–2030. New memory fabs take years to build.
Advanced packaging (especially CoWoS) remains constrained — TSMC and others are expanding capacity aggressively, but demand growth still outpaces it. Larger multi-reticle packages further reduce effective output per wafer. Meaningful relief is not expected until later.
Advanced-node wafer capacity is fully booked — 3nm/2nm (and still 5nm) lines at TSMC and peers stay tight. New leading-edge fabs require 2–4+ years to come online.
Substrates, materials, and equipment lag — IC substrates, ABF, specialty laminates, and glass fiber face shortages (projections of ~20% material gaps in 2027). Key equipment lead times stretch to 24–30 months.
Long structural lead times across the chain — Power infrastructure, cleanrooms, specialized tools (e.g., EUV), and supporting components cannot scale quickly enough to match AI growth rates.
SMIC Co-CEO Zhao Haijun stated the AI industry is still in its early stages with robust demand; many customers still face shortages of computing resources. Demand is driven by data centers and private clouds
Overall, capacity additions are large but still trail demand growth, so the imbalance is expected to continue through 2027 (and in several segments beyond).
