April 30, 2026 – In February this year, SK Hynix and SanDisk jointly hosted the “Launch Event of the HBF Specification Standardization Alliance,” officially unveiling HBF (High Bandwidth Flash), a next-generation memory solution tailored for the AI inference era, and announcing their global standardization strategy.
Although the industry generally holds a positive view of HBF’s prospects and manufacturers are keen to move forward, its rollout pace seems slower than anticipated. It’s rumored that Nvidia has no immediate plans to adopt HBF. On one hand, it continues to rely on HBM, and on the other, it believes that enterprise-grade SSDs (eSSDs) can effectively address issues related to capacity and speed limitations.
Previous reports indicated that Nvidia is collaborating with Kioxia and SK Hynix to develop high-performance AI SSDs that support the PCIe 7.0 standard. These SSDs are aimed at partially replacing HBM as an extension for GPU memory.
To further enhance efficiency, Nvidia is also developing a software platform named “SCADA” (SCaled Accelerated Data Access). This platform bypasses the traditional architecture where CPUs read data from SSDs and then pass it to GPUs, significantly shortening the data path and improving the efficiency of both training and inference processes.
The SCADA approach places extremely high demands on SSD performance, driving the development of AI SSDs with a target of achieving 100 million IOPS.
In contrast to Nvidia’s cautious stance, Google is likely to emerge as a major customer for HBF. With its TPU ecosystem growing rapidly, Google is planning its next-generation TPU solution to further expand its computing capabilities. HBF is attractive to Google as it can partially replace HBM and also serve as an alternative to standard DDR memory.
From a technical perspective, HBF achieves significant capacity increases by vertically stacking NAND flash memory, similar to the stacking method used in HBM. As AI agent services expand, the role of “key-value” (KV) cache memory, which serves as long-term memory, becomes increasingly important. This role, previously filled by HBM, is gradually shifting to NAND flash memory.
