HBF could cut AI memory costs but squeeze NAND supply

HBF is not "cheap NAND" — it is a structural reordering of the memory hierarchy. When inference capacity is delivered at one-tenth HBM4's per-GB cost, HBM4's pricing anchor is permanently reset. The real killer is wafer consumption: 3D stacking plus advanced packaging means each GB of effective capacity devours several times the wafer area of planar NAND. Once Samsung, SK Hynix, and Micron tilt capacity toward HBF, consumer and data-center SSDs enter structural shortage within 12 months. History rhymes: in 2023, HBM cannibalized conventional DRAM capacity and server memory prices nearly tripled. NAND's "HBM-ification" will be more violent — the player base is more fragmented, capacity elasticity worse. Kioxia and SanDisk, if they cannot match advanced-packaging capability, will be demoted from NAND suppliers to HBF "raw-material vendors," with the profit pool captured upstream by OSAT houses. Within 24 months: AI chipmakers embed HBF into inference BOMs; NAND spot prices spike in a 2021-style pulse; OSAT capacity in Taiwan, China (ASE, Powertech) becomes the new geopolitical chokepoint. Whoever locks in advanced-packaging capacity holds the "memory oil" of the AI inference era.