SanDisk sees KV cache driving 35% of AI data center NAND workloads by 2030

If you’ve ever wondered where all the data generated during an AI chatbot conversation actually goes, the answer is: it piles up fast, and the storage industry is scrambling to keep pace. SanDisk, freshly spun out from Western Digital as of February 2025, is placing a major bet that one specific piece of AI infrastructure, the KV cache, will account for 35% of all AI data center NAND flash workloads by 2030.

What KV cache actually does, and why it matters

KV cache stands for key-value cache. When an AI model processes your prompt, it generates key-value pairs at each layer of its neural network. Storing those pairs lets the model avoid redundant computation when generating subsequent tokens in a response. Without KV cache, every new word the model produces would require recalculating attention across the entire input sequence from scratch.

This cached data grows linearly with sequence length and session counts. As AI models handle longer context windows and serve more simultaneous users, KV cache storage needs balloon. Traditionally, that data lives in high-bandwidth memory (HBM) or DRAM. SanDisk’s pitch is to offload KV cache to NAND flash storage, enabling the scaling of context windows without a proportional increase in compute or power costs.

The numbers behind SanDisk’s bet

SanDisk’s forecasts, presented at the Future of Memory and Storage 2026 (FMS 26) event, estimate that KV cache alone could drive 75 to 100 exabytes of additional NAND demand by 2027, with that figure potentially doubling by 2028. The broader enterprise SSD market is projected to grow at approximately 35% compound annual growth rate through 2030. Data centers are expected to become the single largest segment of NAND consumption by 2026, overtaking consumer devices like smartphones and laptops.

SanDisk showcased PCIe Gen5 enterprise drives capable of accommodating up to 256 TB, specifically optimized for KV cache workloads. These drives leverage SanDisk’s BiCS10 QLC (quad-level cell) NAND technology, which prioritizes density and endurance. High-endurance ratings matter here because KV cache involves constant write and overwrite cycles that would quickly degrade consumer-grade SSDs.

The AI inference shift reshaping storage markets

SanDisk’s KV cache thesis is a derivative of a larger trend: the AI industry’s pivot from training to inference. As inference workloads scale, token intensity grows with longer conversations, more complex reasoning chains, and multi-modal inputs, all contributing to larger KV caches per session. HBM costs roughly 10 to 20 times more per gigabyte than NAND flash, making NAND-based tiering a compelling cost reduction opportunity for cloud providers and AI companies.

What this means for the competitive landscape

SanDisk isn’t operating in a vacuum. Samsung, SK Hynix, Micron, and Kioxia are all pursuing enterprise SSD opportunities in AI data centers. Samsung and SK Hynix have dominated the HBM market that currently handles most KV cache duties. As a pure-play flash storage company post-spinoff, SanDisk no longer carries Western Digital’s legacy hard drive business alongside its NAND ambitions.

The 35% figure for KV cache’s share of AI data center NAND workloads by 2030 depends on several variables: how quickly AI model architectures evolve, whether alternative approaches to KV cache compression or eviction reduce storage needs, and how aggressively hyperscalers invest in NAND-based tiering infrastructure.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.