Many still don't adequately price the growing demand for NAND. When AI is discussed, everyone generally focuses on GPUs. But behind the scenes, a perhaps quieter but far more permanent transformation is happening on the storage side, especially NAND. Why? Because the way AI is used has changed. Previously, a model would receive a single question, generate the answer, and close the process. Today, with agentic AI, things have moved to a completely different point. Systems are no longer just structures that respond; they are transforming into structures that remember, track, operate, and work over multiple cycles. At the heart of this transformation is the KV cache. The model stores the intermediate information it creates for each token and reuses it in subsequent steps. So, as the process gets longer, the memory load grows, and as it grows, it becomes more difficult to manage.

This is where the real issue begins. This data is now too large to fit into HBM or DRAM. Since GPUs are expensive, constantly recalculating the same data isn't practical. Therefore, this state needs to be stored at a lower level. This is where NAND comes in. NVIDIA's CMX approach actually points to this very problem. KV is making the cache a natural part of the inference process by moving it to a more accessible, more scalable storage layer. This means NAND is no longer just "storage," but an active component of the AI ​​infrastructure.

I think the most important change here is this: Performance is no longer determined solely by computing. If storage is slow, the GPU waits. If the GPU waits, the cost increases. If the cost increases, the system becomes inefficient. Therefore, NAND's role is not just to provide capacity, but also to maintain throughput. Moreover, the workload is very different from classic enterprise storage. We're talking about a large-block, read-heavy, long-lived, and almost immutable data structure. This means that ordinary SSD architectures will not be sufficient.

In the big picture, I see this: As AI grows, not only the need for computing but also the need for memory and storage increases exponentially. In fact, in some use cases, the storage side may grow faster than most people expect. Therefore, it's no longer correct to see NAND as a mere background component. In my opinion, this side is one of the layers of the AI ​​infrastructure cycle that is at least as important as the GPU, but much less discussed. This is not investment advice.