Sandisk Corp. and SK hynix Inc. released the first High Bandwidth Flash (HBF) technical specification through the Open Compute Project on August 5, 2026, six months after the consortium began work in February. The spec is the first open technical standard of its kind in the memory and storage industry, and it is being positioned by both companies as the answer to a problem the AI inference era has been quietly building for two years: the gap between how much near-compute memory an inference server actually needs and how much High Bandwidth Memory (HBM) the supply chain can deliver.
StorageNewsletter reported the announcement in detail. The specification was developed through the HBF technology workstream under OCP, with Sandisk and SK hynix serving as primary contributors. Notably, Google and Tenstorrent joined as consortium members during the standardization process, contributing to technology validation and the establishment of the standard. The spec provides companies and developers designing AI inference systems and accelerators with a common technical framework for incorporating HBF technology where larger, near-compute memory capacity and higher bandwidth are needed to improve power and performance metrics and reduce total cost of ownership.
What HBF Is Actually For
Modern AI inference systems need high-bandwidth memory positioned close to compute cores, while the demand for greater near-compute memory capacity continues to grow with the requirements of large language models and emerging AI workloads. HBF technology is designed to address this need by combining high bandwidth with high capacity, helping data center system designers improve interactivity and throughput during model serving. As Sandisk Chief Technology Officer Alper Ilkbahar put it, AI inference is creating a new set of memory requirements, and HBF technology is designed to meet that moment. The spec defines system interface, electrical, and other technical guidelines for designing systems that interact with and use HBF technology, including basic performance expectations, the xPU-HBF host interface, reliability and packaging guidance for an HBF die stack, and a software user guide for read and write operations.
The functional pitch is that HBF lets an inference server hold more of a model’s working set close to the compute, which reduces how often the chip has to fetch from off-package DRAM or NVMe storage. For multi-modal models and long-context inference in particular, the working set has been outgrowing what HBM can hold at acceptable cost. HBF is being positioned as a tier that sits between HBM and conventional NAND, trading some of HBM’s bandwidth for substantially higher capacity per stack and substantially lower cost per gigabyte.
Why Google and Tenstorrent Joined
Google’s participation is the more strategically important of the two consortium additions. Google is the only hyperscaler that has both an internal TPU silicon program and a Cloud business that sells inference capacity to third parties. If Google’s TPU roadmap adopts HBF, that locks a second major AI accelerator vendor into the standard, alongside Tenstorrent’s RISC-V-based workload. A standard with one vendor is a custom interface. A standard with two competing accelerator architectures is an ecosystem. The spec’s release under the OCP framework, with openly available documentation, is also a deliberate move to keep the spec from being captured by a single commercial party’s roadmap.
Where HBF Sits in the Stack
As one of the first technical standards of its kind in the memory and storage industry, the specification helps give AI compute system designers added flexibility to build systems where HBF technology can coexist with High Bandwidth Memory, helping support ecosystem readiness. Sandisk and SK hynix proactively published the specification to position HBF technology as the de facto standard in the rapidly evolving AI storage market. Their strategy involves fostering an early-stage ecosystem, increasing the visibility of HBF technology’s adoption for customers, and accelerating market expansion and technological maturity through open collaboration and membership in the consortium. Sandisk will also deliver a keynote at FMS 2026 on NAND’s role in the AI memory hierarchy, alongside a panel discussion on HBF technology’s use and development needed for success, including architectural integration, technical challenges, standardization timelines, performance, and economics.
Industry quote
“AI inference is creating a new set of memory requirements, and HBF technology is designed to meet that moment. This specification helps give system designers a practical path to bring high-capacity, high-bandwidth memory closer to compute, while enabling more flexible architectures.”
Alper Ilkbahar, Chief Technology Officer of Sandisk, quoted by StorageNewsletter
What Comes Next
The spec is the foundational step, not the finished product. The next milestones are silicon that implements the xPU-HBF host interface, software that supports the read/write user guide, and at least one inference accelerator reference design that demonstrates HBF co-existing with HBM rather than replacing it. The open question is whether the industry treats HBF as a new tier in the memory hierarchy that gets built into every AI server from 2027 onward, or as a niche solution for long-context and multi-modal workloads. The OCP release gives the standard the institutional weight it needs to be the former, but the answer will come from the first commercial systems that ship with HBF on the bill of materials. That window is now the back half of 2027.
Why Open Standards Matter More Here Than in HBM
Analysis: HBM has been dominated by SK hynix, with Samsung and Micron in distant second and third positions, because the JEDEC standard for HBM is governed by a closed working group. HBF is different. The OCP release puts the interface, electrical, and packaging guidance in the open, which means any NAND vendor with engineering depth to design a high-capacity die stack can compete. The risk for Sandisk and SK hynix is that they have given away the standardization upside in exchange for ecosystem adoption. The reward is the same upside JEDEC gave DDR and HBM, which is that whoever’s dies ship in the most reference platforms owns the long-term volume curve.