SK hynix, SanDisk Publish First HBF Memory Standard

SK hynix and SanDisk Published the First High Bandwidth Flash Standard. Nothing Ships Yet.

The specification defines a NAND tier that sits between HBM and the SSD inside an AI accelerator, reaching 512GB and up to 3TB/s. It was released through the Open Compute Project with Google and Tenstorrent in the consortium — and it is a document, not a product.

SK hynix and SanDisk released the first specification for High Bandwidth Flash on 4 August, timed to the Future of Memory and Storage conference in Santa Clara. The Korea Herald carried it at 05:27 UTC. HBF is a new memory class positioned between DRAM-based HBM and the SSD, trading latency for capacity on inference workloads.

The specification

Capacity reaches 512GB through 8-high and 16-high NAND die stacks. Bandwidth is defined in three grades running from roughly 0.4TB/s to 3.0TB/s, over a UCIe interconnect. The document covers connection interfaces, electrical characteristics, reliability guidance, die-stack packaging and software I/O. It went out through the Open Compute Project; the consortium, launched in February 2026, includes Google and Tenstorrent.

A document, not a part

Nothing here can be bought or sampled. There is no HBF product, no announced customer and no commercialisation date — SK hynix commits only to SanDisk prototypes and a pilot line in the second half of 2026. Secondary coverage that pins mass production to 2027 has picked up a figure attached to a different item in the same announcement: a 375-layer 4D NAND with a claimed 2.5× gain in performance per watt. Estimates for HBF's actual commercial arrival range from 2027 to around 2030 depending on which outlet you read, which is a way of saying nobody knows.

Who is in it

Google and Tenstorrent are consortium participants, not announced buyers. SK hynix's own text names Google; some coverage upgrades this to Google DeepMind, which the release does not say.

Why flash

Serving large models is now constrained by memory capacity rather than raw compute, and HBM is the expensive way to add it. Kioxia pitched a similar thesis at the same conference from the SSD side.

Why It Matters

Routing a memory tier through an open standards body, with Google signed on, is an attempt to make the answer to the HBM cost wall an industry default rather than a proprietary product — and to set it outside Nvidia's control.