SK hynix and Sandisk Target AI's Real Bottleneck: HBM Scarcity in Inference

HBM scarcity is pushing AI's bottleneck away from compute

The next AI repricing may hit the memory layer before the GPU layer. HBM is already sold out through 2026, and industry analysis says the constraint is shifting from compute to memory. When supply tightens at one point in the stack, pricing power tends to concentrate there first.

That shift is being driven by AI workloads. The industry is moving from batched training to inference runs continuously across many deployments. Those workloads need both bandwidth and capacity, so bigger models do not help if the system cannot keep data flowing to the processor fast enough.

That is why High Bandwidth Flash is more than a product announcement. Sandisk and SK hynix have signed an MOU to work together to establish the specification for High Bandwidth Flash, and they are creating an OCP workstream to advance the standard. The goal is not just a new part; it is an attempt to define a broader inference-memory category.

With entire 2026 HBM production is sold out, investors are increasingly focused on which architecture can ease the memory constraint first.

What HBF is and where it fits in the AI stack

HBF is a NAND flash-based solution contained in an HBM package. SK hynix and Sandisk describe it as a new memory layer between HBM and SSD. The basic idea is to keep HBM for the hottest data the GPU needs immediately, then add a wider, flash-based tier for larger working sets instead of pushing everything down to conventional storage.

HBF is meant to supplement HBM, not replace it

The pitch is practical: supplement traditional DRAM-based HBM with a higher-capacity layer that can hold more data close to the accelerator at a lower cost per unit of capacity. In that sense, HBF is about expanding the memory hierarchy, not eliminating HBM.

That matters because inference is not only a speed test. In production LLM serving, the memory bus is the actual constraint. As context windows and concurrent users grow, systems need more room to keep active state without stalling.

The real opportunity is the gap between fast memory and dense storage

HBM is fast but expensive and scarce; SSDs are denser but too slow to solve the memory bottleneck on their own. HBF only matters if it narrows that gap enough to keep more model state and session data in a faster tier.

Standardization is central to that case. If HBF becomes an open interface, adoption depends less on a single vendor's roadmap and more on whether other designers build around it.

What would move HBF from concept to market test

Sandisk says it will begin sampling HBF in the 2nd half of 2026. That keeps the timeline close enough that investors no longer need to treat HBF as only a slide-deck concept.

What to watch: - Whether sampling leads to multi-vendor designs rather than a single joint announcement. - Whether early systems show real relief when working sets grow, not just strong bench numbers. - Whether the OCP standardization effort turns into an ecosystem that customers can actually deploy.

Open standards matter only if the ecosystem adopts them

The key question is no longer whether HBF addresses a real architectural need. It is whether the market treats it as an open interface worth building around or dismisses it as another vendor-backed spec. Right now, a dedicated workstream under the Open Compute Project is already underway, and the next proof point is sampling HBF in the 2nd half of 2026.

Why an open standard could matter

If HBF becomes an open interface, the value shifts from a single product launch to ecosystem influence. Sandisk and SK hynix are launching a dedicated OCP workstream and framing interoperability as central to adoption. If that works, the companies helping define the interface could capture more of the value chain than those that simply react to it.

There is also a broader structural reason not to dismiss that path. As AI systems have evolved, limitations in memory bandwidth and data transfer efficiency have become more pronounced, and in production inference the memory bus is the actual constraint. That keeps the focus on memory capacity and movement, not just raw compute.

Why standardization alone does not prove demand

A standard does not force adoption. Even constructive industry framing points to a longer buildout before HBF becomes a major revenue category, which means the opportunity may arrive before the earnings visibility does. Skeptics can also argue that OCP standardization does not by itself prove hyperscaler mandates, software support, or efficient use of the flash tier in real workloads.

That is also where behavioral bias can distort the picture. Investors already understand sold-out capacity, long-term supply agreements, and rising margins in HBM, so a newer memory tier can be easy to underweight simply because it does not fit the familiar narrative.

Until those signals show up, this remains constructive for SK hynix's AI-memory breadth and for Sandisk's role higher in the stack, but it still looks more like optionality than conviction. The main risk is that HBF becomes technically plausible and commercially late.