NAND & AI storage

NAND and AI storage is the least glamorous layer of the memory hierarchy: the flash chips and hard drives that hold the mountains of data AI models train on and generate. It matters because every token, trained or served, touches a storage tier, and the capacity for those tiers was planned years before the AI demand shock arrived. That mismatch has turned the cheapest link into a binding one, with nearline hard-drive lead times stretched out badly and enterprise SSD pricing inflecting as hyperscalers rush to provision the vast data lakes that feed AI pipelines.

If storage stays tight the effect is quiet but real: data-centre build-outs slip while waiting for drives, and the cost floor of running an AI pipeline rises even where compute is plentiful. Under-investment in the last cycle is precisely what seeds the next shortage. The exposure sits with the flash and controller specialists that package and sell this capacity, from SanDisk on the memory side to controller houses like Phison, Silicon Motion, Longsys and Biwin that sit between raw NAND and the finished drive.