Each gigabyte of HBM requires three times the wafer capacity of standard DRAM.1 That single manufacturing constraint explains why AI chip makers face a memory bottleneck that no quarterly production ramp can quickly resolve.
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 SK hynix and Samsung both warn the shortage will last at least through end of 2027.1
The demand side is accelerating simultaneously. Counterpoint Research forecasts HBM demand from custom AI processors will increase 35x between 2024 and 2028.3 GPU makers and hyperscalers designing custom silicon are all competing for the same constrained wafer output.
HBM differs from standard DRAM in both architecture and cost structure. Stacking multiple DRAM dies vertically with through-silicon vias (TSVs) delivers the wide memory bandwidth that large language models require. But the stacking process demands more wafer area per unit of storage, limiting how fast suppliers can scale gigabyte output even when fab capacity expands.
Micron Technology's stock has surged 162% year-to-date in 2026.1 The market is pricing in a sustained pricing environment, not a temporary spike. Micron's Singapore facility is on a production ramp timeline targeting H2 2028.1 Until that capacity comes online, supply growth remains structurally capped.
For AI infrastructure builders, the bottleneck is direct. GPU clusters assembled today pay a premium not just for compute but for the memory bandwidth that makes that compute usable. HBM bandwidth determines model serving throughput. A constrained HBM supply means constrained AI inference capacity, regardless of how many GPU dies a manufacturer can produce.
The 35x projected demand growth in custom AI processor memory underscores the scale of the mismatch.3 Hyperscalers designing their own AI chips — a trend accelerating across the industry — are each adding HBM demand that compounds against a supply base growing at a fraction of that rate.
The structural case for elevated HBM and DRAM pricing through 2027 rests on physics as much as market dynamics. Wafer area is finite. HBM's manufacturing requirements are not shrinking. Until new fab capacity reaches production scale, the constraint holds.