The market equates artificial intelligence with GPUs, HBM memory and fast flash, and treats hard drives as a sunset technology. That is a mistake. Training and deploying AI models generates enormous volumes of data that have to be stored durably and cheaply: training sets, successive versions of model weights, checkpoints, inference logs, vector databases for RAG and multimodal data (image, video, audio). The vast majority of that data is cold or warm, rarely read, and economically it has to land on the cheapest medium per terabyte, meaning high-capacity nearline drives.
The core of the thesis is cost-per-terabyte economics. Flash remains several times (roughly 5-7x) more expensive per terabyte in the capacity class, and that gap is not closing quickly, because NAND makers are steering limited capacity toward high-margin enterprise SSDs instead. As a result, the marginal petabyte of AI data ends up on spinning platters. Counterintuitively, AI is not killing hard drives, it is expanding their market.