Three manufacturers — South Korea's SK hynix, Samsung, and US-based Micron — control virtually all HBM supply. Each gigabyte of HBM requires three times the wafer capacity of standard DRAM.1 That physics-driven constraint is why no production ramp resolves the shortage quickly.
DRAM contract prices are forecast to rise 58–63% in Q2 2026 alone.1 Gartner projects a 125% full-year DRAM price increase for 2026.2 Both SK hynix and Samsung warn the shortage will persist through at least end of 2027.1
Demand is accelerating from every direction simultaneously. Counterpoint Research forecasts HBM demand from custom AI processors will grow 35x between 2024 and 2028.3 US hyperscalers, Chinese cloud giants, and European AI infrastructure builders are all competing for the same constrained wafer output.
HBM stacks multiple DRAM dies vertically using through-silicon vias (TSVs), delivering the wide memory bandwidth that large language models require. That stacking process demands more wafer area per unit of storage — a manufacturing reality that limits gigabyte output even as fab capacity expands.
Micron's stock has surged 162% year-to-date in 2026.1 Markets are pricing in a sustained premium, not a temporary spike. The company's Singapore facility is on a production ramp timeline targeting H2 2028.1 Until new capacity comes online across all three suppliers, supply growth remains structurally capped.
For AI infrastructure builders worldwide, the bottleneck is direct. HBM bandwidth determines model serving throughput. Constrained HBM supply means constrained AI inference capacity — regardless of GPU production volumes anywhere in the world.
The mismatch between 35x projected demand growth and supply expansion measured in single-digit percentages annually defines the structural case for elevated pricing through 2027.3 Wafer area is finite. HBM's manufacturing requirements are not shrinking. Until Micron's Singapore lines and expanded Korean fab capacity reach production scale, the constraint holds globally.
