AI training + inference dual demand drives HBM volume from 80kt in 2024 to ~200kt by 2026; SK Hynix / Samsung / Micron oligopoly.
Why HBM is not commodity DRAM
HBM3E stacks 12 DDR5 dies vertically through TSV, hitting 1.2 TB/s — 30x DDR5. NVIDIA H200 holds 141 GB of HBM3E per card; B200 holds 192 GB. The bandwidth wall is tighter than the compute wall in Hopper clusters — HBM is now 50%+ of GPU BOM cost.