Memory is going to stay expensive through 2028, and the same shortage that keeps it expensive is what guarantees the crash that follows. That's my verdict, and it isn't really about AI. Micron, Samsung and SK Hynix are signing 2027 DRAM contracts above 2026 levels because HBM3E and HBM4 eat roughly three times the wafer area per bit that DDR5 does. Every Blackwell order is capacity that doesn't ship in a laptop. DDR5 modules have doubled in some retail channels. TrendForce has logged double-digit quarterly contract increases all through 2025. Read that as a signal about construction schedules, not just prices. Three years of tight supply is an invitation.
The invitation is already being accepted. Micron is building in Idaho and New York, Samsung is expanding Pyeongtaek, and SK Hynix has Yongin. Those projects take three to four years from groundbreaking to wafers, which puts meaningful new output in 2028 and 2029. Nobody can time this market, so everybody builds at once, because no one wants to be the supplier who blinked and lost HBM share to a rival. That is the mechanism. Supply shows up in a cluster while the HBM mix keeps pulling wafer area away from commodity DRAM, and then one quarter the mix shifts back. Capacity added in 2026 lands regardless.
This is what makes the current run so dangerous for the people paying for it. HP, Dell and Lenovo run 3 to 5 percent operating margins, so they pass memory costs straight to buyers and trim base configs; a 16GB laptop becomes a 12GB laptop. Hyperscalers sign multi-year deals, bury memory in AI capex, and don't flinch. Consumers do flinch, and they blink first. They hold a three-year-old machine another year, and that demand destruction arrives fast, landing on suppliers who committed capex two years earlier at prices they now can't cover. If you need a machine in 2026 or 2027, buy it and stop watching contract pricing. If you can wait, the glut is the thing worth waiting for.