SanDisk Prepares Breakthrough HBF Flash Memory for AI Infrastructure

SanDisk has revealed details regarding the new HBF (High Bandwidth Flash) memory standard, created together with SK hynix. This memory aims to reduce the costs of building AI infrastructure and lessen the industry's reliance on scarce GPUs and traditional HBM chips.

SanDisk expects to deliver the first physical samples next year, with mass production scheduled to begin in 2028.

HBF Efficiency

Comparing HBF to HBM, the former will feature higher read speeds and 6 to 8 times greater capacity. This is achieved through HBF’s architecture, which utilizes vertical stacks of 8 or 16 NAND dies with a total capacity of up to 512 GB. As a result, a computing system combining HBF memory with 4 GPUs can deliver the same token output when running large language models as a traditional setup requiring 8 GPUs with HBM memory.

With a throughput matching that of HBM and a lower overall cost, HBF memory could significantly reduce the deployment expenses of data centers.

Long-Term Contracts and Financial Outlook

The new memory standard will serve as a key growth driver for SanDisk. Company representatives revealed at an investor event that they project average annual revenue growth of 16–19% from 2028 through 2030, maintaining profit margins around 80%.

SanDisk has already secured agreements with 8 major clients, who are guaranteed to purchase up to two-thirds of SanDisk's total production volume by 2028.