SanDisk has projected that key-value cache, or KV cache, workloads will account for 35% of NAND flash demand inside AI data centers by 2030. The forecast was reported by CryptoBriefing on August 13, 2026. It signals a notable shift in how storage manufacturers expect artificial intelligence infrastructure to evolve over the next several years.
KV cache is a technical component used in large language model inference. It stores intermediate data, known as key and value pairs, generated during the attention process inside transformer models. This data allows AI systems to avoid recomputing information for every new token generated in a response. As models handle longer conversations and larger context windows, the size of this cache grows substantially.
That growth creates a memory bottleneck. GPU memory, while fast, is limited and expensive. Data center operators have increasingly looked to tiered storage systems, including high-performance NAND flash, to offload KV cache data without sacrificing too much speed. SanDisk's projection suggests this offloading pattern will become a dominant use case for flash storage in AI facilities.
NAND flash is the memory technology used in solid-state drives. It has traditionally served roles such as storing training datasets, model checkpoints, and general data center storage needs. A 35% share attributed to KV cache workloads alone, if realized, would represent a meaningful reallocation of storage demand toward inference-specific tasks rather than training or general-purpose storage.
The forecast arrives as AI companies continue scaling inference capacity to serve growing user bases. Training large models once dominated headlines about compute and storage needs. Inference, the process of running trained models to generate responses, is increasingly described by hardware makers as the larger long-term driver of infrastructure spending. SanDisk's projection fits that broader narrative, framing inference-related memory demands as a defining feature of AI data center design through the end of the decade.
Storage and semiconductor companies have been repositioning product roadmaps around AI-specific workloads for several years. Flash memory makers, DRAM producers, and specialized memory architecture firms have all pointed to inference scaling as a source of future demand. SanDisk's specific KV cache figure adds a quantified data point to that trend, even as the broader industry continues refining how it measures AI-driven storage consumption.
The 2030 timeframe gives the projection a multi-year runway, during which model architectures, context window sizes, and inference optimization techniques could all change. Techniques such as cache compression or alternative memory hierarchies could alter how much NAND capacity KV cache workloads ultimately require. SanDisk's estimate should be read as a directional forecast rather than a fixed outcome.
Market Impact
If SanDisk's projection holds, NAND flash manufacturers could see a new, durable source of demand tied specifically to AI inference rather than training cycles. That could influence capital spending decisions across the memory industry, as suppliers weigh capacity investments against expected inference-driven growth.
For data center operators, the forecast highlights storage architecture as a growing cost and design consideration alongside GPU procurement. Companies building AI infrastructure may need to account for flash storage tiers optimized for KV cache offloading, potentially reshaping procurement strategies for both memory and storage components over the coming years.
SanDisk's projection adds a concrete figure to the ongoing conversation about how AI inference is reshaping data center hardware needs. Whether the 35% share materializes will depend on how quickly inference workloads scale and how storage architectures adapt.
Frequently Asked Questions
What is KV cache in AI systems?
KV cache stores key and value data generated during a language model's attention process, allowing it to avoid recomputing information for each new token.
Why does KV cache affect NAND flash demand?
As KV cache grows with longer AI conversations and context windows, data centers increasingly offload it to storage tiers like NAND flash instead of relying solely on limited GPU memory.
What did SanDisk project?
SanDisk projected that KV cache workloads will account for 35% of AI data center NAND demand by 2030, according to a report from CryptoBriefing.
Does this forecast affect crypto markets?
The report concerns AI infrastructure and storage demand rather than cryptocurrency markets directly, though it reflects broader trends in data center and hardware spending.