AI buyers lock in capacity, driving Taiwan private investment to record NT$6.95T

This is not a capex story—it is a structural repricing of supply-chain power. Technical cascade: The AI compute bottleneck has migrated from wafer-node to HBM stacking and advanced packaging, then cascading into invisible chokepoints—photoresist, specialty gases, CMP slurries. The prepayment-plus-long-term-contract model permanently transfers pricing power from spot-market dynamics to upstream material and equipment suppliers. Echoing the 2018 memory cycle, when Samsung locked silicon-wafer capacity via three-year contracts, Shin-Etsu's bargaining leverage jumped roughly 40%. If equipment and materials exceed 50% of the NT$6.95T outlay, Taiwan, China's supply chain is shifting from manufacturing hub to pricing hub. Risk: The flip side of locked capacity is stranding. In 2023, when inference demand underwhelmed, CoWoS utilization dipped below 60%. If training-demand growth decelerates by 2027, prepaid orders become sunk costs. Tightening equipment export controls further concentrate equipment makers' overseas customer risk. Competitive dynamics: Samsung's HBM3E push and Intel's 18A node will chip at packaging share. Yet once buyers commit capital to lock capacity, the real cost of a second-source strategy spikes—paradoxically hardening the incumbent structure. 12–24 month outlook: Two digestion waves—HBM4 ramp in 2026Q2 and logic-chip expansion in 2027Q1—will pressure utilization. Materials pricing power will be the single largest profit-pool battleground.