SK hynix and SanDisk Unveil First HBF Standard-New Memory Layer Between HBM and SSDs Could Reshape AI Inference

SK hynix and SanDisk introduced the first High Bandwidth Flash standard

SK hynix and SanDisk have unveiled the first standard specifications for High Bandwidth Flash (HBF), a memory solution positioned between HBM and SSDs. The announcement came at Future of Memory and Storage (FMS) 2026, and the specification was published through the Open Compute Project. The standard defines three bandwidth grades ranging from approximately 0.4 TB/s to 3.0 TB/s, while supporting capacities of up to 512 GB using 8-high and 16-high NAND die stacks. shift from training to inference

Why AI inference is the catalyst

The timing matters because the AI workload focus is moving from model training toward inference. In that phase, systems need not only speed but also larger memory capacity and better power efficiency. According to SK hynix and SanDisk, HBF was designed to fill the gap between HBM's high performance and SSDs' high capacity, while supporting both capacity expansion and power efficiency in AI inference infrastructure. ensure both capacity expansion and power efficiency required for AI inferencing

HBF also uses the UCIe interconnect standard, which should make it easier to integrate with different processors, including CPUs and GPUs. That makes it more of a platform question than a single-product story.

Bull case: a new memory tier for inference workloads

The bullish view is straightforward: HBF could serve as a dedicated memory layer for live AI traffic. In that role, it would offer more capacity than HBM and higher bandwidth than a typical SSD. The expectation is that better scalability and power efficiency can lower the overall cost of serving inference workloads. secure scalability and power efficiency in AI inference infrastructure

Bear case: faster flash is still flash

The skeptical view is that this is not a new problem with a new name. Stacked NAND can read faster, but it remains flash memory. IEEE Spectrum quoted one analyst saying that while NAND is generally slow, for reads, it can be coaxed to go pretty fast. That supports the idea that HBF may help narrow the gap, but it does not prove that flash can fully replace the role of HBM.

Competition is also moving quickly. Samsung is already engaged in early HBF development, while Kioxia has shown prototype effort in the space. Samsung Electronics has begun conceptual design and early development That raises an important watchpoint: being first to publish a standard is not the same as owning the market long term.

What matters next

For now, the main takeaway is simple: SK hynix and SanDisk have established the first published HBF standard, and the early debate is less about whether the tier can help inference than about who controls the specification and how quickly vendors can turn it into shippable products.