SK Hynix and SanDisk Publish the First Open Standard for High Bandwidth Flash, Targeting AI Inference Memory
The two memory makers released the first OCP technical specification for High Bandwidth Flash, a new memory tier between HBM and SSDs built to fix AI inference bottlenecks. Google and Tenstorrent joined the standardization effort.
SK hynix and SanDisk released the first Open Compute Project technical specification for High Bandwidth Flash at Flash Memory Summit 2026 in Santa Clara on August 3. HBF is a new memory tier designed to sit between High Bandwidth Memory (HBM) and SSDs — cheaper and denser than HBM, far faster than flash storage, and purpose-built for the capacity-hungry, bandwidth-starved side of AI inference that HBM alone can’t economically cover.
The published spec supports capacities up to 512GB per stack with bandwidth grades ranging from roughly 0.4TB/s to 3.0TB/s, UCIe die-to-die interconnect support, and a 375-layer 4D NAND design SK hynix says is 2.5 times more power-efficient than prior generations. The document covers the full stack an implementer needs: system interface guidelines, the xPU-to-HBF host interface, reliability and packaging guidance for the die stack, and a software user guide for read/write operations.
The bottleneck HBF is aimed at is real and getting worse. HBM capacity has scaled slower than model sizes and KV-cache demands, and GPU/accelerator vendors have been forced to either over-provision expensive HBM or accept throughput cliffs when inference workloads exceed on-package memory. HBF’s pitch is a middle tier: NAND-based, so it’s cheaper per gigabyte than HBM by a wide margin, but architected with enough bandwidth to keep large models resident near the compute die instead of round-tripping to a discrete SSD.
This is the second standardization milestone in a year for the format — SanDisk and SK hynix first partnered on HBF standardization in August 2025, and the Open Compute Project formally stood up an HBF technology workstream in February 2026. Since then, Google and Tenstorrent joined the consortium to help validate the technology and shape the spec, a signal that hyperscalers and accelerator designers outside the two founding memory makers see enough value to commit engineering time before silicon ships.
Publishing through OCP rather than keeping HBF proprietary is the strategically interesting part. An open standard means any accelerator vendor — not just SK hynix and SanDisk’s direct partners — can design against a common interface, which is exactly the dynamic that made HBM itself a multi-vendor ecosystem instead of a single-supplier lock-in. If HBF gets broad adoption, it changes the AI accelerator memory hierarchy from “HBM plus off-package SSD” to a three-tier stack, giving chip designers a genuine option between the two that didn’t exist as a standard before this week.
Don’t expect HBF-equipped accelerators on shelves immediately — this is a specification release, not a product launch. But for anyone tracking where the AI memory bottleneck goes next, this is the document vendors will now design silicon against.