With the increasing parameters and context lengths of AI models, the capacity of High Bandwidth Memory (HBM) is struggling to meet the memory demands of large AI models. Recently, Sandisk announced a collaboration with SK Hynix to develop a new storage product, HBF (High Bandwidth Flash), designed for AI model inference scenarios. The product aims to achieve high-speed GPU interconnection using BiCS and CBA wafer bonding technology. According to Sandisk’s official website, the HBF under development could offer 8–16 times the capacity of existing HBM, potentially expanding GPU storage to 4TB. Given their database expertise, companies like Alibaba may have the potential to develop data infrastructure software for HBF storage.
GF Securities highlights the following key points: 1. **Growing Memory Demands of AI Models** As model parameters and context lengths expand, memory requirements during inference continue to rise. For instance, NVIDIA’s Blackwell Ultra chip features eight HBM3E memory units with 288GB capacity and 8TB/s bandwidth. Despite advancements in HBM standards toward higher bandwidth and DRAM density, trillion-parameter models with context lengths exceeding 128K are outpacing HBM’s capabilities.
2. **HBF as a Potential Solution** In August 2025, Sandisk and SK Hynix unveiled plans for HBF, targeting AI model inference with superior capacity and speed. Sandisk’s HBF could deliver 8–16 times the storage of current HBM, enabling GPU storage expansion up to 4TB. Its high capacity and transfer rates make it well-suited for large-scale AI inference tasks.
3. **Opportunities for Database-Savvy Firms** Companies with database expertise—such as Huawei, Alibaba, StarRing Tech, and PingCAP—could develop HBF-optimized data infrastructure software. While tech giants like Huawei and Alibaba may build proprietary solutions for their AI models, third-party providers like StarRing Tech and PingCAP could cater to firms like Zhipu AI and MiniMax.
4. **HBF’s Potential to Drive Software Adoption** StarRing Tech’s distributed database, ArgoDB, is optimized for flash storage, leveraging fine-grained encoding to maximize random read/write performance. With further development, the company could adapt its solutions for HBF and other emerging storage media. However, product timelines depend on R&D cycles and strategic priorities.
5. **Risks and Challenges** HBF hardware remains about a year from maturity, with potential delays in overcoming technical hurdles. Its viability in AI inference is untested, posing uncertainties in market acceptance and commercialization. Performance gaps between HBF and traditional storage media may also impact software effectiveness and adoption.
Comments