HBF and running big models locally
aihardwarememory
SanDisk’s High-Bandwidth Flash stacks NAND the way HBM stacks DRAM — 512 GB per package, ~1.6 TB/s read bandwidth, eight stacks putting you at 4 TB on a single card. First silicon is expected 2H 2026, with SK hynix now co-developing and the spec going through OCP.
The interesting part: that’s enough to run DeepSeek V4 Pro, Kimi K2.7, or GLM 5.2 unquantized, on one card, locally. No sharding, no cloud dependency, no data leaving the machine. The memory bottleneck that made large open-weight models impractical to self-host just got a credible solution on the roadmap.
— Felix