Sandisk plans to launch a pilot series of high‑efficiency HBF memory by the end of the year

New horizons of memory: Sandisk and competitors prepare for the release of HBF

Manufacturers of volatile and solid-state memory continue to refine technologies to meet growing market demands. One key player in this field is Sandisk, which plans to complete construction of a pilot line for producing *HBF* (High‑Bandwidth Flash) in Japan by the end of this year and move to mass production already in 2027.

What HBF is and why it matters
- HBF – a new type of NAND memory that uses vertical stacking similar to *HBM* (High‑Bandwidth Memory).
- With this approach, memory is placed closer to the GPU, significantly speeding up AI model inference.
- Although HBF’s speed may lag behind HBM, its capacity exceeds the latest by up to 16 times: a 16‑layer stack can hold up to 512 GB, and placing eight such stacks next to a GPU allows up to 4 TB.

Sandisk plan
Step | Timeline | Notes
---|---|---
Search for materials, components, and equipment for HBF | Already started (ETNews) |
Set up experimental line in Japan | Next half‑year |
Obtain first HBF samples | By the end of this year |
Mass production | 2027 (possibly six months earlier if everything goes as planned) |

Sandisk uses equipment and technologies already employed for HBM production, reducing supply delay risks.

Competition
- Chinese memory manufacturers are also preparing to launch HBF, intensifying competition.
- SK hynix – a competitor of Sandisk – supports HBF standardization and is already involved in development.
- Samsung Electronics has been working on similar memory since the start of the decade but has not yet announced specific market entry dates.

Market prospects
Peak demand for HBF is expected by 2030, when AI shifts from training large language models to inference. In this context, large volumes of memory close to the GPU will become critically important.

Thus, Sandisk and its competitors are actively developing HBF technology, which promises to radically change data storage approaches for high‑performance computing and AI inference.