Future TechnologyHigh Bandwidth Flash Just Got Its First Open Spec, and Samsung Answered With zHBM
Key Takeaways
- SK hynix and SanDisk published the first open High Bandwidth Flash spec through the Open Compute Project: stacks up to 512GB, roughly 0.4TB/s to 3.0TB/s, connected over UCIe.
- Samsung showed zHBM, successor to z-NAND, claiming around eight times HBM5 performance. That figure is the vendor's and is not independently verified.
- HBF is not a rival to HBM. It exists to stop expensive HBM being spent on inference workloads that need capacity more than bandwidth.
A memory stack holding 512GB, wired to a processor over UCIe, running at up to 3.0TB/s. That is the headline for High Bandwidth Flash, and as of Future of Memory and Storage 2026 in Santa Clara it is an open standard rather than one vendor's slide deck.
What High Bandwidth Flash actually is
SK hynix and SanDisk published the first open HBF specification through the Open Compute Project. It defines stacks of up to 512GB across three bandwidth grades, running from roughly 0.4TB/s at the low end to 3.0TB/s at the top, connected to the processor over UCIe.
The word doing the work is flash. HBF stacks NAND where HBM stacks DRAM. That makes it slower and much cheaper per gigabyte, and it makes writes expensive. Which is fine, because the workload it is aimed at barely writes at all.
Samsung's answer is zHBM
Samsung used the same show to put zHBM in front of people, the successor to its z-NAND line, with a claim of around eight times HBM5 performance. Treat that number as a vendor claim until somebody outside Samsung measures it. The strategic read is clearer than the spec sheet: Samsung is betting on pushing DRAM-like memory harder, while SK hynix and SanDisk are betting on bringing flash closer.
Why the industry needed two answers at once
Because the squeeze is real. HBM3E contract prices are up close to 20 percent, and 2026 and 2027 DRAM and HBM capacity is effectively fully allocated. You cannot buy your way out of this one.
The split that matters is training versus inference. Training is bandwidth-hungry: you are moving enormous amounts of data through the processor constantly. Inference is capacity-hungry: you need the weights resident, you read them relentlessly, and you almost never write. That second shape is flash-shaped. Spending HBM on it is like renting a warehouse and using it as a corridor.
If HBF works, the cost curve for running models bends in a way the cost curve for training them has stubbornly refused to.
What to watch
Publishing a spec is the cheap part. The things to watch are whether a shipping part appears with a price attached, and whether anyone beyond SK hynix and SanDisk signs on. An open standard with two backers is a proposal. An open standard with Micron, Samsung or a hyperscaler behind it is a roadmap.