Product signal
SK hynix and Sandisk Publish HBF Specification
SK hynix and Sandisk have published an open High Bandwidth Flash specification for AI inference servers, defining modules up to 512GB and roughly 0.4TB/s to 3.0TB/s.
SK hynix and Sandisk have published the first open High Bandwidth Flash, or HBF, standard specification, defining flash-memory modules of up to 512GB with approximately 0.4TB/s to 3.0TB/s bandwidth for AI inference servers. This is a concrete interface and ecosystem move: it does not make HBF a replacement for existing high-bandwidth memory, but it adds an open capacity-and-bandwidth tier for inference-system design.
The concept becomes an open standard
SK hynix says it and Sandisk jointly published the first HBF standard specification through the Open Compute Project. The announcement specifies capacities up to 512GB and three bandwidth levels from roughly 0.4TB/s to 3.0TB/s. Unlike a single vendor’s internal roadmap, publication through OCP gives controller, module and server participants a shared definition around which to build implementations and test interoperability.
The mechanism is an added inference-memory tier
HBF is flash-based but targets server-class bandwidth. Tom’s Hardware identifies AI inference servers as the intended market, suggesting this is not a general consumer-storage refresh but an attempt to change capacity, bandwidth and system-cost tradeoffs in inference workloads. Its potential role is to give system designers another resource-allocation option alongside traditional storage and high-bandwidth memory, rather than to directly replace one component.
Standardization comes before adoption
For infrastructure buyers, an open specification immediately reduces interface uncertainty: server vendors can assess module layouts, controller support and software scheduling, while cloud and inference operators can test workload economics. The strongest reservation is that a published specification is not deployment at scale; compatible hardware, system announcements and workload benchmarks still need to establish its practical effect.
What to watch next
Watch for OCP ecosystem controller, module and server reference designs; compatibility announcements from major server or cloud vendors; and public tests reporting HBF capacity, throughput, latency and power behavior on defined inference workloads. Those signals will determine whether HBF becomes a new memory tier or remains a specification.