HBF consortium releases first High Bandwidth Flash spec, targeting AI’s memory bottleneck

The High Bandwidth Flash consortium has released its first technical specification, roughly six months after launching in February 2026. The spec outlines a new class of NAND-based memory designed to tackle one of AI’s most stubborn infrastructure problems: there simply isn’t enough fast, affordable memory to feed the models.

Led by Sandisk and SK Hynix, the consortium counts Google and Tenstorrent among its members. The group’s goal is straightforward, even if the engineering isn’t: build an open standard for flash memory that can keep up with AI inference workloads without carrying the eye-watering price tag of high-bandwidth memory, or HBM.

What the spec actually promises

The initial HBF specification lays out some ambitious performance targets. Individual NAND stacks could support capacities up to 512 GB per package, a figure that dwarfs what’s currently possible with HBM modules.

Bandwidth ranges from hundreds of GB/s in baseline configurations to a peak of 3 TB/s when using UCIe connections. UCIe, or Universal Chiplet Interconnect Express, is an open standard for connecting chiplets inside a package.

The distinction between training and inference is worth pausing on. Training is the brute-force phase where a model learns from data. Inference is when that trained model actually does useful work, like answering questions, generating images, or running autonomous vehicle perception. Inference is where most of the compute spending is headed, and it’s far more memory-capacity-hungry than it is memory-bandwidth-hungry.

The memory wall problem

HBF proposes a tiered memory approach. Rather than relying exclusively on HBM for everything, systems could use HBM for the hottest data and HBF for the vast bulk of model parameters that need fast access but not the absolute fastest.

Sandisk and SK Hynix had been laying the groundwork well before the consortium’s formal launch. The two companies signed a memorandum of understanding in August 2025 to collaborate on high-bandwidth flash development. An advisory board was formed in July 2025, suggesting the technical conversations started even earlier.

Competitive landscape and what comes next

The HBF consortium isn’t operating in a vacuum. Samsung has been developing its own approaches to high-bandwidth flash memory, though outside this particular consortium framework. The decision by Sandisk and SK Hynix to pursue an open specification rather than proprietary solutions is a deliberate strategic choice.

Google’s presence in the consortium is notable because it validates demand from one of the world’s largest AI infrastructure operators. Tenstorrent, the AI chip company led by legendary chip architect Jim Keller, brings a different perspective as a hardware startup designing inference-focused processors that would directly benefit from cheaper, denser memory options.

For semiconductor supply chains, HBF could shift investment patterns. NAND flash fabrication is a mature, high-volume manufacturing process with well-understood economics. Repurposing that existing infrastructure for AI-optimized memory products is far less capital-intensive than building new HBM production lines, which require advanced packaging technologies that only a handful of facilities worldwide can handle.