SanDisk Introduces High Bandwidth Flash Memory to Address AI's Memory Bottleneck

SanDisk, in collaboration with SK hynix, has launched a new memory standard known as High Bandwidth Flash (HBF), aimed at overcoming one of the most pressing challenges in AI infrastructure: the need for fast and affordable memory. This innovative technology combines NAND flash with advanced engineering techniques to deliver performance that rivals traditional High Bandwidth Memory (HBM) while offering significantly greater storage capacity.

Performance and Capacity

The HBF technology is designed to bridge the gap between conventional SSDs and HBM, providing a solution that balances high capacity with substantial bandwidth. In simulations, HBF achieved read bandwidths of 12.8 TB/s with a capacity of 4 TB per GPU, which could dramatically reduce the number of GPUs required for complex AI models. This capacity is reported to be 8 to 16 times that of current HBM offerings, making it a cost-effective alternative for AI applications.

Technical Specifications

• Each HBF stack can support up to 512 GB, with bandwidth capabilities reaching up to 3 TB/s when using Universal Chiplet Interconnect Express (UCIe) connections.
• HBF's performance is particularly beneficial for inference workloads, where the majority of AI processing occurs, as it emphasizes read-heavy operations.
• The technology utilizes proprietary CMOS Directly Bonded to Array techniques to enhance performance by minimizing signal routing bottlenecks.

Industry Collaboration and Future Plans

SanDisk has formed a consortium with SK hynix, Google, and Tenstorrent to standardize HBF technology under the Open Compute Project. The first technical specifications were released in August 2026, with product samples expected in 2027. This collaborative effort aims to create an open standard that can be widely adopted across the semiconductor industry, potentially reshaping investment patterns in AI infrastructure.

Implications for AI Development

The introduction of HBF is poised to significantly impact the deployment of large language models and other AI applications by providing a more efficient memory solution that reduces costs and enhances performance. As AI continues to evolve, the ability to leverage existing NAND manufacturing processes for high-performance memory products could lead to more scalable and economically viable AI systems.

New Developments in SanDisk's AI Memory Strategy

• SanDisk shares surged approximately 16% following a positive long-term growth outlook focused on AI demand, advancements in NAND flash technology, and an expanding enterprise storage business.
• The company's fiscal Q3 2026 results highlighted significant profit growth driven by AI data center storage needs, particularly through NAND-based products and enterprise SSDs for hyperscale environments.
• Since its spin-off from Western Digital, SanDisk's stock has increased by over 5,900% in certain measurement periods.
• Analysts predict that memory supply will remain constrained until at least 2030, providing SanDisk with pricing power amid tight supply conditions.
• Despite reporting revenue of $8.97 billion in early August 2026, SanDisk's stock experienced a drop due to softer forward guidance, illustrating the volatility in the market.