The Memory Supercycle: When Demand Outruns What Physics Can Build

The mechanism is a lead-time mismatch, not a willingness-to-invest problem. SK hynix's chairman put a new ('greenfield') memory fab's build time at more than five years — meaning capacity ordered today lands near the tail end of the shortage window he's forecasting through 2030, not before it. TSMC's CEO said the same about advanced logic nodes: 'Building the capacity and then ramping it up now takes five to seven years. There are no shortcuts.' ASML, the sole supplier of the EUV lithography machines both companies depend on, is 'investigating' a further 30% capacity expansion for 2028 specifically because customer order signals are strong enough to justify it — but its own CFO wouldn't commit to a number, since matching supply to a demand curve that's still rising is inherently a moving target. Within memory specifically, the gap compounds further: high-bandwidth memory (HBM) consumes far more wafers per bit than ordinary DRAM and carries higher margins, so new capacity keeps tilting toward HBM — which starves standard DRAM supply even as chipmakers add wafers overall, explaining why DRAM contract prices reportedly rose roughly 63% quarter-over-quarter even as capacity investment accelerated.