On August 4, 2026, at FMS 2026 (Future Memory and Storage), held in Santa Clara, California, USA, SK hynix and SanDisk officially released the first OCP technical specification for High Bandwidth Flash (HBF). The specification is made publicly available to the entire industry through OCP (Open Compute Project), the world’s largest open data center technology organization. It is positioned as an industry-open technical specification rather than a proprietary technical standard of any single vendor. This is a key milestone in the development of HBF technology and also marks HBF’s official entry into the stage of open standardization and ecosystem validation.
A Key Milestone in the Standardization Process
The pace of HBF standardization has exceeded industry expectations. As early as August 2025, SK hynix and SanDisk officially began joint standardization work on HBF. The two companies planned to jointly develop a new storage technology specification for AI applications based on their respective technological expertise. In February 2026, the HBF technical workstream was officially established under the OCP framework, bringing together companies across the industry value chain to participate in standards development.
From the launch of the workstream in February 2026 to the release of the first OCP technical specification, the entire process took only six months. Behind this rapid progress is the AI industry’s urgent demand for new storage solutions, as well as the open and collaborative approach to standards development. SanDisk and SK hynix were the main contributors, while Google and Tenstorrent also participated in the standardization process and provided support for technical validation and specification development.
HBF Positioned as an Intermediate Tier for AI Storage
Before HBF emerged, the storage architecture of AI computing systems was mainly composed of HBM (High Bandwidth Memory) and traditional enterprise SSDs. There was a clear gap in both performance and cost between the two.
HBM, with its DRAM architecture and 3D stacking technology, can deliver extremely high bandwidth and low access latency, making it an important storage medium for supporting real-time AI computing. However, HBM has relatively limited capacity and high cost. As the amount of data such as large-model weights and KV Cache continues to grow, relying only on HBM makes it difficult to economically accommodate all near-compute data. Traditional SSDs use NAND flash to provide large capacity at a relatively low cost, but their bandwidth and access latency are significantly behind those of HBM.
HBF is a new solution designed to address this storage-tier requirement. SK hynix positions it as a storage tier between HBM and SSD. It uses NAND flash to provide greater capacity, while using high-bandwidth interfaces and advanced packaging technologies to shorten the distance between data and computing units, thereby meeting the need for “larger capacity + higher bandwidth” in scenarios such as AI inference.
| Storage Type | Core Medium | Bandwidth Level | Capacity Characteristics | Cost Level | Key Application Scenarios |
|---|---|---|---|---|---|
| HBM | DRAM | Extremely high | Limited capacity | Very high | Real-time AI computing, frequently accessed data |
| HBF | NAND flash | High | Large capacity | Relatively low | AI inference, large-capacity near-compute data |
| SSD | NAND flash | Low | Very large capacity | Low | Cold data storage, persistent archiving |
Core Specifications of the First Technical Specification
The first OCP technical specification for HBF released this time covers system interfaces, electrical characteristics, basic performance requirements, packaging and reliability, and software read/write guidelines. It provides a unified technical framework for HBF product development and system integration across the industry.
In terms of capacity configuration, the specification provides two NAND chip stacking options: 8-layer stacking and 16-layer stacking. A single HBF module supports up to 512GB, further meeting the need for large-capacity near-compute storage in AI inference scenarios.
In terms of performance levels, the specification divides HBF into three grades, from Grade 1 to Grade 3, with data transfer bandwidth ranging from 0.4TB/s to 3.0TB/s. Products at different grades can be configured according to the bandwidth and capacity requirements of different AI systems.
Two Core Technical Features
HBF seeks to balance capacity and bandwidth. It relies on key technologies such as high-speed processor interconnects and advanced NAND stacking and packaging, while also emphasizing openness and engineering feasibility.
- The first is the adoption of the UCIe universal chiplet interconnect standard. HBF uses UCIe as the processor-side interconnect technology, providing an open interface foundation for high-speed connections between different types of computing processors, such as GPUs and CPUs, and HBF. Compared with traditional storage architectures, this design is expected to shorten the data access path and provide a more standardized interface for adapting HBF to computing chips from different vendors.
- The second is high-density NAND stacking and wafer bonding technology. HBF uses SanDisk’s BiCS flash technology and CBA (CMOS directly bonded to array) technology, combined with proprietary die stacking technology, to support high-density stacking of up to 16 layers. Previously released technical materials from SanDisk show that its HBF solution uses packaging interconnect structures such as TSVs and micro-bumps, and emphasizes reducing die warpage through proprietary stacking technology to support stacking with a high number of layers.
Industry Ecosystem and Mass Production Plans
As an open technical specification, HBF has emphasized industry-chain collaboration since its inception. An ecosystem involving companies from the storage, AI chip, cloud computing, and other fields has now taken shape.
The main participants in the HBF standardization process include storage vendors SK hynix and SanDisk, as well as AI chip companies Tenstorrent and Google. During the FMS 2026 summit, SK hynix, Google DeepMind, and SanDisk also jointly participated in a panel discussion on HBF, exploring the development direction of storage architectures for AI infrastructure.
In terms of product development, SanDisk has already advanced HBF chip development and disclosed in August 2026 that HBF technology was attracting increasing attention from the industry ecosystem. At its 2026 Investor Day held on August 13, the company further stated that an industry ecosystem supporting the application of HBF technology was taking shape.
Underlying flash memory technologies are also progressing at the same time. At this summit, SK hynix publicly unveiled its tenth-generation (V10) 375-layer 4D NAND flash for the first time. The company said that its performance per watt had improved by 2.5 times compared with the previous generation. SK hynix also showcased its next-generation storage technology roadmap for AI infrastructure.
Industry Impact and Future Outlook
The release of the first OCP technical specification for HBF is not only an important attempt in the development of storage technology, but is also expected to have an impact on the storage architecture of AI infrastructure.
First, it further enriches the tiered architecture of AI storage. In addition to computing chips and HBM, HBF seeks to provide a near-compute storage tier with larger capacity and higher bandwidth than traditional SSDs, offering a new storage option for data-intensive applications such as large-model inference and Agentic AI. SK hynix has also clearly positioned HBF between HBM and SSD to address the growing capacity and bandwidth requirements in the era of AI inference.
Second, it is expected to help AI infrastructure achieve a better balance between performance and cost. Instead of placing all high-capacity data in high-cost HBM, using HBF to carry some data that requires greater capacity while also requiring relatively high bandwidth could reduce the overall storage cost of the system and improve data access efficiency.
Third, it may promote further collaboration across the NAND, advanced packaging, chiplet interconnect, AI server, and other upstream and downstream industries. HBF involves high-density NAND stacking, wafer bonding, high-speed interconnects, and software read/write mechanisms. Therefore, its commercialization depends not only on the storage chips themselves, but also on the joint progress of computing chips, packaging, servers, and the software ecosystem.
Overall, the release of the first OCP technical specification for HBF means that this technology has moved beyond independent development by individual vendors and further entered the stages of open standardization, industry collaboration, and product validation. Through the open OCP framework, SK hynix, SanDisk, Google, Tenstorrent, and other participants are jointly driving the development of the technical specification, providing a unified technical foundation for different vendors to explore HBF applications.
As subsequent product samples, system validation, and ecosystem adaptation continue to advance, whether HBF can truly become an important storage tier in the AI era still requires further observation of its performance, cost, and reliability in actual AI inference systems. However, one thing is clear: as AI moves from training toward large-scale inference, the traditional “HBM + SSD” storage architecture is facing new challenges in capacity and bandwidth, while HBF is becoming one of the key directions the industry is exploring for this new storage tier.
