Sandisk has unveiled NAND-based technologies for AI inference workloads, including High Bandwidth Flash, updated NAND products and enterprise solid-state drives.
The announcement focuses on how storage and memory designs are adapting as AI systems demand more capacity, bandwidth and lower power use for inference, not just training.
Sandisk's latest work spans High Bandwidth Flash, or HBF, a NAND-based memory approach for AI inference, alongside next-generation 3D NAND and enterprise SSD products. It also updated its AI Data Cycle framework, which maps storage requirements across stages including raw data archives, data lakes, training, inference and generated-content repositories.
The company is targeting workloads tied to AI deployment and operation, including fast data lakes, model storage, retrieval-augmented generation and key-value cache workloads. Those areas have become more prominent as AI models handle longer context windows and businesses shift from building models to running them repeatedly in production.
Storage focus
Among the products highlighted is BiCS10 QLC NAND, a new 3D NAND generation designed to improve density, performance and power efficiency for AI, cloud and data centre storage. QLC, or quad-level cell, technology stores more bits per cell than earlier forms of NAND, increasing capacity but also creating engineering challenges around endurance and speed.
Sandisk also pointed to enterprise SSD configurations for key-value cache workloads, including high-endurance options, future PCIe Gen 6 technology and ultra-high-capacity E3.S designs with reduced DRAM placement.
The emphasis reflects a broader shift in AI infrastructure spending. While graphics processors remain central to model execution, memory and storage suppliers are seeking a larger role by arguing that data movement and access patterns can become bottlenecks once models are deployed at scale.
High Bandwidth Flash is the clearest example of that strategy. Sandisk is positioning HBF as a NAND-based memory technology for AI inference, where large volumes of data must be stored and accessed efficiently, but the cost and power demands of other memory types can limit deployment.
Inference demands
The updated AI Data Cycle framework is intended to show where different storage layers fit across the AI workflow. By separating archives, fast data lakes, training environments, inference systems and repositories for generated output, Sandisk argues that no single memory technology suits every task.
That systems view has grown more important as AI operators look beyond raw compute to the economics of serving models in everyday use. Inference workloads can run continuously, and their costs are shaped not only by processors but also by how quickly data can be retrieved, how much memory is needed to hold model state and how much energy the process consumes.
Khurram Ismail, Chief Product Officer at Sandisk, outlined the company's position on those constraints.
"AI infrastructure has entered a new phase where the ability to store, move, and access data efficiently has become as important as compute itself," said Khurram Ismail, Chief Product Officer at Sandisk.
He added: "As inference workloads grow, customers need memory and storage technologies that can deliver ideal performance, capacity, and power efficiency without adding complexity. Building on decades of flash innovation, Sandisk is advancing next-generation NAND, and emerging technologies such as HBF, to help advance the industry."
Market context
Sandisk's messaging comes as chipmakers and storage vendors compete to define the next layer of AI infrastructure beyond accelerators. Demand for larger models and longer context windows has increased interest in hardware that can feed data to processors more efficiently and store larger working sets close to where they are used.
NAND flash has traditionally played a larger role in storage than in main memory, but suppliers are now exploring how flash-based designs can address some AI bottlenecks, particularly where lower cost per bit and lower power consumption are priorities. The trade-off is usually lower speed than established high-bandwidth memory technologies, which means vendors must position flash for specific tasks rather than as a universal substitute.
For Sandisk, that means linking HBF, QLC NAND and enterprise SSDs to a broader architectural argument. The company is effectively saying that AI deployment now depends on a hierarchy of storage and memory products, with flash taking on a greater share of inference-related workloads such as model storage, retrieval layers and cache-heavy serving systems.
The latest product framing suggests Sandisk wants to shift discussion of NAND away from conventional storage markets and closer to the centre of AI data centre design. The result is a sharper focus on how flash can be used not only to hold more data, but also to manage the cost and energy demands of operating AI services at scale.