What Happened
Semiconductors · Standards · AI Capability · Specification Briefunveiled 4 Aug 2026 · record to 27 Aug 2026
A new memory tier between HBM and SSD gets its first open specification
High Bandwidth Flash Gets Its First Open Standard
SK hynix and Sandisk unveiled the first standard specification for High Bandwidth Flash (HBF) on 4 Aug 2026, the opening day of the FMS 2026 conference in Santa Clara. HBF stacks NAND flash the way HBM stacks DRAM, aiming at a new memory tier between HBM and ordinary SSDs for AI inference workloads — model weights, RAG data, and the KV cache that grows as context windows and AI agents run longer. The specification, HBF 0.7, covers up to 512GB of capacity, three bandwidth grades from about 0.4TB/s to 3.0TB/s, and the UCIe chiplet interconnect. It was published through the Open Compute Project as an open industry standard rather than a proprietary technology, and Google and Tenstorrent have joined the HBF Consortium alongside SK hynix and Sandisk. Nothing here has shipped: SK hynix's own timeline puts the full specification at early 2027 and first samples at early 2028.
By the Numbers the first specification's own figures
- 512GBmaximum HBF capacity in the first specification, across 8-high and 16-high NAND die stacks
- 3.0TB/stop of the specification's three bandwidth grades; the lowest is about 0.4TB/s
- 2×GPU efficiency Sandisk says it demonstrated — 4 GPUs with 4TB of HBF matching 8 GPUs with 192GB of HBM on tokens per second, in one inference case it chose
- early 2028when SK hynix says first HBF samples are expected; the full specification is due early 2027
What's Confirmed, Graded the two companies' own material, against a vendor demonstration and a forecast
| Standing | What | Detail |
|---|---|---|
| Confirmed | That the specification was published through the Open Compute Project, as an open industry standard | SK hynix's own release states the specification “was disclosed through the Open Compute Project (OCP) ... positioning it as an open standard for the entire industry rather than a proprietary technology.” Sandisk's own release states the same independently: it was “released within the Open Compute Project framework, meaning the information is openly available to the industry.” |
| Confirmed | That Google and Tenstorrent are participating in the HBF Consortium | SK hynix's own release: “With Google and Tenstorrent currently participating in the HBF consortium, the consortium aims to expand its reach.” Sandisk's own release, independently: “Google and Tenstorrent joined as consortium members during this standardization process, contributing significantly to technology validation.” |
| Confirmed | The specification's own figures: up to 512GB, three bandwidth grades, and the UCIe interconnect | SK hynix's own release: capacity “up to 512GB based on two stack configurations (8-high and 16-high NAND dies)”; bandwidth “categorized into three grades (Grade 1–3), delivering scalable performance from approximately 0.4TB/s to 3.0TB/s”; and UCIe as the interconnect “linking HBF technology and processors.” |
| Confirmed, vendor's own demonstration | That Sandisk demonstrated roughly 2× GPU efficiency using HBF | Per Forbes's own first-hand account of Sandisk's FMS presentation: “For 4TB HBF only memory with 4-GPUs compared to 192GB HBM with 8 GPUs the tokens generated per second was almost the same, resulting in 2X GPU efficiency and lower total cost.” This is Sandisk's own demonstration of its own technology, in a scenario it chose; Sandisk has not stated the specific inference workload, and no independent benchmark of the comparison is publicly available. |
| Company timeline, not yet realised | That the full HBF specification will publish in early 2027 and first samples will follow in early 2028 | Per Forbes's account of SK hynix's own timeline slide: the “0.7 HBF standard” unveiled at FMS 2026 is followed by “the full HBF spec ... expected to be released in early 2027 with HBF samples available in early 2028 and full HBF flash production after that.” Neither company's own written press release states these dates directly; they are SK hynix's own stated plans, not results. |
Timeline
From Partnership to Published Standard August 2025 to August 2026, from the two companies' own material and Forbes's own coverage
- Aug 2025SK hynix and Sandisk begin an initial standardisation partnership for HBF, per SK hynix's own release.
- Feb 2026The HBF Consortium formally launches.
- 3 Aug 2026Sandisk publishes its own newsroom release on the first OCP HBF technical specification.
- 4 Aug 2026FMS 2026 opens at the Santa Clara Convention Center; SK hynix and Sandisk unveil the first HBF standard specification (HBF 0.7).
- 5 Aug 2026Sandisk delivers its own FMS keynote, “NAND: The Versatile & Scalable Foundation of the AI Era.”
- 6 Aug 2026Sandisk, SK hynix and Google hold a joint panel, “Breaking the Memory Wall with High Bandwidth Flash”; FMS 2026 closes.
- 13 Aug 2026At its own “In Focus 2026” investor day, Sandisk details its BiCS8-to-BiCS10 NAND roadmap.
The Argument
An Open Standard, Built Quickly why two companies chose to publish through OCP rather than keep HBF proprietary
SK hynix's and Sandisk's own framing is that AI inference is opening a gap HBM alone cannot close at reasonable cost. Sandisk CTO Alper Ilkbahar put it in the company's own release: “AI inference is creating a new set of memory requirements, and HBF technology is designed to meet that moment. This specification helps give system designers a practical path to bring high-capacity, high-bandwidth memory closer to compute… It is an important milestone for the HBF ecosystem and for the next generation of AI systems built to improve token economics at scale.” SK hynix EVP Kim Chun-sung, in his company's own release, frames it as a redesign of “overall data processing structures” rather than a single product launch. Both statements are the companies' own characterisation of why the specification matters, not independently verifiable facts. What is independently confirmed is the choice of vehicle: rather than keep HBF as a two-company proprietary technology, SK hynix and Sandisk published it through the Open Compute Project, and Google and Tenstorrent joined the consortium during the standardisation process. Per Forbes's own account of the FMS material, Tenstorrent CEO Jim Keller joined the HBF Consortium's board alongside David Patterson of UC Berkeley and Raja Koduri of Oxmiq Labs — three named individuals with standing across chip design and computer architecture, not employees of either founding company. Building an open standard, and recruiting that board, only makes sense as a deliberate attempt to grow an ecosystem quickly around a memory tier that does not exist as a shipping product yet. What complicates the argument is exactly how early this still is. What was published on 4 Aug 2026 is version 0.7 of the standard, not the full specification, which SK hynix's own timeline places at early 2027; first samples follow at early 2028. And the single figure most likely to travel — Sandisk's own 2× GPU-efficiency demonstration — is a vendor showing its own technology in one scenario it chose, not an independent or third-party benchmark, and Sandisk has not named the workload behind it.
- 6months from the HBF Consortium's launch (Feb 2026) to the first published specification
- 4organisations named in the consortium so far: SK hynix, Sandisk, Google and Tenstorrent
- 3named HBF Consortium board members: Jim Keller (Tenstorrent), David Patterson (UC Berkeley) and Raja Koduri (Oxmiq Labs)
What Others Add
The NAND Doing the Stacking Sandisk's own roadmap, shown at FMS and again at its own investor day
| Generation | Layers | TLC | QLC |
|---|---|---|---|
| BiCS8 | 218-layer | Compute eSSD, up to 32TB | Storage eSSD, up to 256TB |
| BiCS10 | 332-layer | 1Tb per die, sampling | 2Tb per die, sampling |
The Rest of FMS 2026's HBF Material the tiered-memory framework, the consortium's board, and a separate SK hynix NAND announcement
Tiered memory
Five tiers, not one
- SK hynix's own FMS slides, per Forbes, add G0.5 (3D-stacked DRAM, above HBM) and G2.5 (CXL-pooled memory, multiple TBs) around the existing G1 (HBM)
- HBF itself sits at G1.5: about 1TB of capacity, Tb/s-class bandwidth
- SK hynix states the goal as maximising xPU performance and quality of service, and improving TPS/MW and Token/$
Consortium
A wider board than two founding companies
- Google and Tenstorrent joined the HBF Consortium as members during the standardisation process, per both founding companies' own releases
- Tenstorrent CEO Jim Keller joined the consortium's board, per Forbes's own account, alongside David Patterson (UC Berkeley) and Raja Koduri (Oxmiq Labs)
- Publishing through OCP, rather than as a two-company technology, is what makes that board membership possible at all
Same event, separate product
SK hynix's own 375-layer NAND
- SK hynix's own release: a tenth-generation (V10) 375-layer 4D NAND wafer and products, shown publicly for the first time at the same FMS booth
- SK hynix's own claimed improvement: 2.5× the power efficiency of the previous generation
- This is a separate SK hynix NAND product, not a component of the HBF specification; mass production of eSSDs based on it is planned for early 2027
Conclusion
What the Specification Establishes and what is still a stated plan, not a shipped result
What the specification establishes
That SK hynix and Sandisk published the first HBF standard specification on 4 Aug 2026, through the Open Compute Project, confirmed independently by both companies' own releases as an open industry standard rather than a proprietary one. That Google and Tenstorrent are participating in the HBF Consortium, also confirmed independently by both companies. Forbes's own account of the conference material adds that Tenstorrent CEO Jim Keller has joined the consortium's board alongside David Patterson and Raja Koduri — a single-source account, and neither company's release states it. That the specification itself covers up to 512GB across two NAND stack heights, three bandwidth grades from about 0.4TB/s to 3.0TB/s, and the UCIe interconnect. And that Sandisk demonstrated, in a scenario and workload of its own choosing, 4 GPUs with 4TB of HBF matching 8 GPUs with 192GB of HBM on tokens per second — its own vendor demonstration, not an independent benchmark.
What's still a plan, not a result
That the full HBF specification will follow in early 2027, and that first samples will exist in early 2028, are SK hynix's own stated timeline, not events that have happened — nothing has shipped yet, and full production is later still. Whether HBF proves complementary to HBM, as both companies frame it, rather than a competitor to it, and whether Sandisk's 2× efficiency figure holds outside the one scenario it demonstrated, are both open. So is how far the consortium's membership grows beyond SK hynix, Sandisk, Google and Tenstorrent.
What to Watch signals this specification's own timeline points to
- Whether the full HBF specification actually publishes in early 2027, as SK hynix's own timeline states.
- Whether first HBF samples materialise in early 2028, and which other companies join the consortium beyond SK hynix, Sandisk, Google and Tenstorrent.
- Whether an independent benchmark of HBF against HBM appears, beyond Sandisk's own demonstrated scenario.
- Whether SK hynix's five-tier memory hierarchy (G0.5 to G2.5) is adopted as a framework by other memory vendors, or stays a single company's own way of describing the market.