Sandisk and SK hynix Advance Global Standardization of High Bandwidth Flash with First OCP Technical Specification

OCP standardization, not product proof, is the first real development

Sandisk and SK hynix have started a dedicated workstream under OCP to standardize High Bandwidth Flash, with a kick-off at Sandisk headquarters last week. That matters more than the marketing language around the technology. At this stage, HBF is a standards initiative aimed at a potential gap in the AI memory hierarchy, not a deployed product with field results.

Why inference, not training, is the focus

Both companies frame the opportunity around inference rather than training. Inference runs continuously, serves many users, and is described as more sensitive to memory bandwidth, capacity, and power constraints at scale. That makes the memory hierarchy more important economically: the goal is not only speed, but also enough fast capacity to keep model weights close without putting HBM-level memory everywhere.

The technical idea is a new tier between HBM and SSD

HBF is pitched as a bridge layer between HBM and SSD. Sandisk has proposed a 16-die-plus-base-die flash stack that fits the same footprint as HBM, with optimized read characteristics for model weights. SK hynix and Sandisk say the architecture targets HBM for high-bandwidth performance, with HBF handling higher-capacity storage in the same system block.

The first samples are targeted for the second half of 2026, with accelerators featuring HBF expected to sample in early 2027. If the ecosystem embraces the concept, standardization could turn HBF from a company-specific demo into a reusable system component.

HBF's value proposition depends on system architecture, not flash speed alone

The architectural case: fewer hops for model weights

Modern models require billions of parameters. The HBF thesis is that keeping more of those parameters closer to the processor can reduce the time and energy needed to fetch them during inference. That is a different claim from saying flash is now generally fast enough to replace existing memory tiers. The opportunity is more specific: a new memory tier that helps systems hold larger model sets near the compute engine.

Why OCP matters more than headline specs

A joint dedicated workstream under the Open Compute Project is the clearest early signal because standards bodies are where hardware ecosystems turn new concepts into something engineers can design around. If HBF remains a vendor-specific story, it is likely to stay niche. If it becomes an adopted new memory layer between HBM and SSD, the architecture could matter well beyond any single product launch.

What could limit adoption

There are at least two reasons to stay measured. First, a new memory tier needs broad ecosystem support, including board design, packaging, and accelerator integration. Second, demand is not expected immediately across the industry: one industry view cited in coverage says demand for complex memory solutions, including HBF, is expected to expand around 2030.

There is also a competitive-pressure point. If HBM keeps improving in speed and capacity while costs fall, the economic case for inserting another layer becomes narrower. HBF does not need to beat HBM on raw bandwidth; it needs to offer a better balance of capacity, power, and system cost for certain inference workloads.

Investor read: pre-revenue optionality, not a near-term earnings driver

Treat HBF as samples due out later this year-driven optionality rather than a near-term revenue contributor for Sandisk or SK hynix. The deployment timeline is still too early to model meaningful sales, so the right lens is strategic: can HBF become a standardizable system block before the market settles on another solution?

What to watch next

• Whether OCP moves from launch to published specification
• Whether sample dates hold or slip
• Whether accelerator and board designers publicly build around HBF
• Whether later demand timelines remain consistent with the current 2026-2027 sample path

Invalidation signals

If ecosystem participation stays narrow, if timelines keep slipping, or if HBM advances fast enough to keep capacity and cost under control, HBF may remain an interesting concept rather than a material new memory category.