Sandisk Expects AI Workloads to Drive 60GB to 90GB of NAND Flash Demand Per CPU
Sandisk is seeing a major shift in how companies design memory and storage for artificial intelligence systems. According to Sandisk Vice President Rajeev Nagabhirava, some AI inference and AI agent-orchestration workloads could require around 60GB to 90GB of NAND flash for every CPU.
The comment was shared during the SEMICON Taiwan 2026 Memory Summit, where Nagabhirava discussed HBF-based KV cache management and the growing role of NAND flash in modern AI infrastructure.
As AI models become more advanced, the way they use memory is changing quickly. Inference workloads, especially those involving large language models and AI agents, often need to handle huge amounts of temporary data. One important part of this process is the KV cache, which stores key-value data used by AI models to generate faster responses during inference.
Traditionally, AI systems have leaned heavily on DRAM and high-bandwidth memory for performance-critical tasks. However, as deployments scale, companies are looking for more efficient ways to balance cost, capacity, and speed. This is where NAND flash could become increasingly important.
Sandisk’s estimate of 60GB to 90GB of NAND per CPU highlights how AI infrastructure may evolve beyond conventional memory setups. Instead of relying only on expensive memory technologies, system designers may use a layered approach that combines different types of memory and storage. NAND flash can provide large capacity at a lower cost, making it attractive for AI inference servers and agent-based workloads that need to manage significant data movement.
AI agents add another layer of demand. Unlike simple chatbot interactions, agentic AI systems can plan tasks, call tools, maintain context, and coordinate multiple operations over time. These activities can increase memory pressure, especially when many agents are running at once. Efficient KV cache management becomes essential to keep performance high while avoiding excessive infrastructure costs.
The growing interest in HBF-based KV cache management suggests that storage companies are preparing for a future where NAND flash plays a deeper role in AI performance optimization. While NAND is not a direct replacement for faster memory, it can help support larger AI workloads by acting as a high-capacity tier within the system.
For data centers, this could influence future hardware purchasing decisions. If AI inference continues to expand across cloud services, enterprise platforms, and edge deployments, demand for NAND flash may rise alongside CPUs, GPUs, and specialized AI accelerators.
Sandisk’s outlook reflects a broader trend in the AI hardware market: memory and storage are becoming just as important as compute power. As companies scale AI services, they must rethink how data is stored, accessed, and moved through the system.
The key takeaway is clear: AI infrastructure is entering a new phase where NAND flash could become a critical part of the performance and cost equation. With certain workloads potentially requiring up to 90GB of NAND per CPU, storage capacity may become a defining factor in how efficiently AI inference and AI agent systems operate.
