Sandisk says customers are coming back for more NAND

AI inference workloads are fracturing the traditional NAND demand model. RAG pipelines, vector databases, and model checkpointing have created a structural storage floor absent from the 2019-2022 capex planning cycles. SanDisk's admission that customers underestimated demand is, in essence, a confession that the industry's capital allocation framework was calibrated for a pre-AI world. Technically, this is not a volume story. AI inference demands high-IOPS, low-latency NAND — QLC commodity tiers are being squeezed out while TLC/PLC with advanced controllers becomes the new baseline. Upstream, this accelerates 200+ layer 3D NAND and CBA architecture ramp; downstream, EDSFF enterprise SSDs see renewed pull, and the HBM→DRAM→NAND hierarchy is being re-architected. Strategically, with LTA pipelines already exceeding end-of-decade planned capacity, the industry is entering an allocation economy. The 2018 DRAM shortage playbook — Samsung and SK Hynix cherry-picking customers — is likely to repeat. Hyperscalers face supply-security pressure and may push dual-sourcing or in-house controller development. The competitive race now hinges on who ships 300+ layer products first in volume; Samsung's 286-layer and Micron's 232-layer mark the current frontier. Outlook: Over 12-24 months, NAND ASPs will hold elevated. The 2023-style deep trough is unlikely to recur. AI's structural floor is rewriting industry cyclicality — any future downcycle will be materially shallower, and 2025-2026 capex will be AI-directed rather than consumer-driven.