Sandisk and SK hynix have published the first open High Bandwidth Flash (HBF) specification through the Open Compute Project, defining 512GB NAND stacks running up to 3.0TB/s. For UK buyers wrestling with Why AI Server Costs are Rising, it's the first credible, vendor-neutral roadmap for flash sitting between HBM and NVMe.
View the data behind this chart
| Entry Grade | Mid Grade | Top Grade | |
|---|---|---|---|
| HBF bandwidth per stack | TB/s0.4 | TB/s1.6 | TB/s3 |
What was actually announced
The spec, released on 4 August 2026, is the output of an HBF workstream that Sandisk and SK hynix formed inside the Open Compute Project just six months earlier, in February 2026, according to StorageReview.
Google and Tenstorrent joined the consortium during the standardisation process, and the finished document is now openly available to any company building AI accelerators or inference systems, not just the founding two vendors, per the same report.
Crucially, this is a technical specification covering the host interface, electrical guidelines, packaging and reliability rules, and a software guide for read and write operations — the groundwork needed before anyone can design silicon around HBF, rather than a shipping product announcement.
The numbers UK procurement teams should note
StorageReview reports HBF stacks arriving in 8-high and 16-high NAND configurations at up to 512GB per stack, connected over the UCIe interconnect standard used to link chiplets and accelerators.
Three performance grades span 0.4TB/s to 3.0TB/s, giving system designers a menu rather than a single fixed part. Additional reporting cited in the coverage suggests Sandisk's first-generation HBF is likely to land nearer 1.6TB/s, with 2TB/s-plus and eventual 3.2TB/s figures reserved for later generations — useful context for anyone modelling near-term deployments rather than the ceiling of the spec.
HBF is explicitly designed to sit alongside HBM in the same system, not replace it, which matters for anyone speccing GPU Accelerators where DRAM-class bandwidth still has to handle the hottest compute paths.
Why this matters for large-context AI workloads
The problem HBF is aimed at is well understood among UK buyers already reading up on High Bandwidth Memory (HBM) Explained: HBM capacity is limited and expensive, while model sizes and context windows keep growing faster than DRAM density.
HBF is pitched as the tier that holds model weights and long-context data that no longer fits in HBM but is too latency-sensitive for a PCIe-attached SSD. StorageReview frames this as the same capacity math already driving KV-cache offload to flash — more context held in a cheaper, denser tier means more tokens and more concurrent users served per accelerator.
That framing is directly relevant to anyone reading our Buying for AI Inference guidance, since HBF is squarely an inference-economics play, not a training-cluster upgrade.
Timeline reality check: don't buy the roadmap yet
Sandisk's chief technology officer, Alper Ilkbahar, called the spec a milestone "for the next generation of AI systems built to improve token economics at scale," but StorageReview is candid that all the figures are vendor-stated, and hardware still needs to reach the lab before anyone ships production systems built on it.
Corroborating coverage places first HBF samples later in 2026 with first devices arriving in early 2027, and separate industry commentary suggests broad deployment could stretch well beyond that. UK buyers should treat HBF as a 2027-plus procurement decision, not a line item for this year's refresh cycle.
There's also a technical caveat worth flagging to engineering teams: writes are understood to be limited relative to reads, so early software stacks will likely prioritise read-heavy inference rather than write-intensive workloads.
A crowding field, not a two-horse race
Sandisk and SK hynix aren't standing still on the flash side either — the same event brought BiCS10 QLC, a tenth-generation NAND die built with Kioxia that the companies say delivers a 60% bit density increase over their eighth-generation QLC, with no sampling date yet given.
Samsung used its own FMS keynote the same week to preview competing zHBM and zNAND-O concepts aimed at the identical memory-wall problem, while Kioxia has separately shown a prototype HBF-style module rated at 5TB capacity and 64GB/s bandwidth in an SSD-like form factor.
The practical takeaway is that publishing early through an open OCP spec, with Google and Tenstorrent already inside the tent, gives Sandisk and SK hynix a genuine head start in the standards race — but UK buyers evaluating AI and Analytics Storage Solutions should still expect multiple, possibly incompatible, flash-as-memory approaches before one becomes dominant.
View the data behind this chart
| Layer | Detail |
|---|---|
| HBM (DRAM) | Fastest, capacity-limited, near-compute |
| HBF (flash stack) | 512GB stacks, 0.4–3.0TB/s, persistent |
| NVMe SSD | Bulk capacity, PCIe-attached storage |
What UK buyers should do now
Nothing here changes a 2026 GPU order. HBF is a specification, not a component you can buy, and the honest read from StorageReview is that hardware validation comes first. Use the AI GPU Calculator against your actual context-window and concurrency targets before assuming a future flash tier solves today's HBM shortfall.
For teams speccing storage now, the sensible move is to keep SSD and NVMe Components plans flexible enough to accommodate a UCIe-attached tier later, and to track the OCP HBF workstream rather than any single vendor's marketing, since the spec is explicitly open to whoever adopts it.
Longer term, if HBF matures on schedule, it strengthens the case for inference-optimised architectures over general-purpose training clusters, which is worth factoring into multi-year capacity and refresh planning even while the near-term buying decision stays unchanged.
- 01StorageReview — High Bandwidth Flash Gets Its First Open Spec: 512GB Stacks and Up to 3.0TB/s · 4 August 2026
- 02HPCwire — What Is High Bandwidth Flash (HBF) and Can It Overcome the AI Memory Wall? · 4 August 2026
- 03Tom's Hardware — Sandisk and SK hynix unveil HBF spec: up to 16-Hi NAND stacks, 3TB/s bandwidth, UCIe · 4 August 2026
- 04The Register — GPUs could explode to multiple TB with new storage-inspired memory tech · 30 July 2026
- 05Blocks & Files — A window into HBF progress: engineering prof talks of decade ahead · 19 January 2026
- 06Blocks & Files — Kioxia eyes GPU memory stack with High Bandwidth Flash push · 20 March 2026
- 07Tom's Hardware — Kioxia's new 5TB, 64GB/s flash module puts NAND toward the memory bus for AI GPUs · 20 March 2026
- 08Blocks & Files — Sandisk and SK hynix working to standardize High Bandwidth Flash · 7 August 2025