HDD vs NAND Flash: How AI Data Centres Use Both for Cheap Storage

The article argues that AI data centres should use a tiered design—HDD vs NAND flash—because each storage type solves different problems in training and inference pipelines. NAND flash (SSDs) is positioned closest to GPUs for low latency and very high random IOPS, powering model weights, KV caches, and vector indexes. HDDs are kept for the bulk, long-retention capacity tier: large training corpora, synthetic/augmented data, stabilized checkpoints, logs, and growing inference outputs. Key performance and cost points cited include Seagate’s estimate that HDDs are ~6× cheaper per TB than NAND SSDs, alongside much lower energy and embodied carbon per TB—important at hyperscale. On endurance, flash requires controller-level management (ECC, wear leveling, over-provisioning) because cells wear out with program/erase cycles. The piece references SNIA guidance that write amplification and workload patterns strongly affect real flash lifetime. A practical workflow is described: store authoritative datasets and archives on HDD-backed object/file storage; stage hot shards on NVMe/SSD before runs; use flash for bursty checkpoint writes; then migrate stabilized checkpoints and artifacts back to HDD to recover expensive flash capacity. Cloud and vendor best practices (e.g., staging/caching guidance) are presented as consistent with this tiered approach. For crypto traders, the immediate market signal is indirect: this is an infrastructure/IT efficiency narrative rather than a crypto protocol or regulation event.