SK Hynix and SanDisk have published the first standard specification for High Bandwidth Flash, a NAND-based memory designed to give AI accelerators terabytes of extra capacity without the cost of HBM. The spec, released August 4 through the Open Compute Project, follows a consortium the two memory makers launched in February.
HBF stacks NAND dies 8 or 16 high for up to 512GB per package, links to processors over UCIe, and offers three bandwidth grades from roughly 0.4 TB/s to about 3.0 TB/s. At the top end that approaches HBM4-class bandwidth, but the partners position HBF as an additional memory layer for AI inference and storage-heavy workloads, not a replacement for HBM.
The announcement lands at Flash Memory Summit, where a Google DeepMind engineer joins SK Hynix and SanDisk executives on a panel about breaking the memory wall with HBF. SK Hynix is also previewing the 375-layer 4D NAND wafer behind its next enterprise drives, which it plans to start mass producing early next year.
Standardization matters because AI systems are running short on memory capacity: GPUs need far more storage-adjacent memory for large models, and a common spec lets server makers mix suppliers. SanDisk and SK Hynix began their HBF partnership in August 2025, and the consortium is now courting the broader ecosystem to turn the spec into shipping products.
