SanDisk and SK Hynix Push High-Bandwidth Flash Standard for AI Workloads

A New Memory Blueprint for AI

SanDisk and SK Hynix introduced a technical specification for high-bandwidth flash (HBF) through the Open Compute Project, establishing a framework that could influence how AI infrastructure handles memory in the coming years. The specification addresses a specific bottleneck: inference systems that require faster access to data stored closer to the processor, particularly as workloads scale beyond what traditional storage hierarchies can support efficiently.

The Open Compute Project, an industry consortium focused on open hardware design, now hosts the HBF standard. By placing the specification in this forum, SanDisk and SK Hynix signal an intent to build industry consensus rather than pursue a proprietary path. The move matters because memory architecture remains fragmented across AI accelerator platforms, and a shared standard could reduce integration overhead for system builders.

Why Flash Bandwidth Matters in Inference

AI inference differs from training in its memory profile. Training runs often demand high-capacity, high-throughput storage for massive datasets, but inference workloads prioritize latency and sustained bandwidth when serving models to end users. As large language models and generative AI applications proliferate, inference clusters face mounting pressure to deliver responses in milliseconds while juggling concurrent requests.

High-bandwidth flash sits between DRAM and conventional SSDs in the memory hierarchy. It offers lower latency than networked storage and greater capacity than on-package DRAM, a combination that becomes critical when models exceed tens of billions of parameters. The HBF specification defines electrical interfaces, thermal envelopes, and protocol layers that allow flash modules to communicate directly with accelerators or CPUs at speeds previously reserved for volatile memory.

Regional Implications

The collaboration carries weight in Asia's semiconductor landscape. SK Hynix, headquartered in South Korea, ranks among the top three global memory manufacturers and has invested heavily in high-bandwidth memory for AI chips. SanDisk, owned by Western Digital, maintains significant operations in Japan and Malaysia, where it produces NAND flash at scale. Together, the companies command substantial manufacturing capacity and supply-chain reach across the region.

China's AI infrastructure developers, constrained by export controls on advanced GPUs, have turned to memory and storage optimization as a performance lever. An open standard for high-bandwidth flash could accelerate adoption of alternative inference architectures in Shenzhen, Shanghai, and Beijing, where companies are building custom accelerators to circumvent chip restrictions. Japan's AI research labs and India's hyperscale data centers also stand to benefit if the specification gains traction, lowering the cost of deploying inference at scale.

What Comes Next

The specification itself is a technical document, not a shipping product. System integrators and accelerator vendors must now decide whether to design around the HBF standard or continue with proprietary memory interfaces. Adoption hinges on whether major cloud providers and AI platform companies - Amazon Web Services, Microsoft Azure, Google Cloud, Alibaba Cloud, Tencent Cloud - incorporate HBF modules into their next-generation inference servers.

SanDisk and SK Hynix have not disclosed a timeline for commercial modules that comply with the specification, nor have they named design partners. Industry observers expect early implementations to appear in prototype systems within twelve to eighteen months, followed by volume shipments if the standard proves stable. The Open Compute Project will likely host reference designs and interoperability testing to smooth the path from specification to silicon.

For now, the HBF standard represents a bet that inference workloads will increasingly rely on a distinct memory tier, one that balances speed, capacity, and power efficiency in ways that neither DRAM nor traditional SSDs can achieve alone. Whether that bet pays off depends on how quickly the ecosystem rallies around a common approach to memory in AI systems.