Part of the day: NAND flash

Assembled by the nightly build from this catalog’s own entry for NAND flash (2.07, Layer 2 — Memory & storage). The tables below are rebuilt every night; first published 2026-08-31, refreshed 2026-09-24.

Non-volatile flash memory die — the raw storage medium later packaged into SSDs — that holds data and model artifacts when the system is powered off.

NAND is the upstream input to the enterprise SSDs and controllers described next (2.08), and it sits below DRAM (2.02) and above HDD (2.09) in the usual data center storage hierarchy: faster and more expensive than spinning disk, but non-volatile and far cheaper per unit than DRAM. AI workloads write checkpoints frequently during training and stage large datasets locally ahead of GPU reads, and inference deployments cache model weights on fast local flash to shorten cold starts — all of which push endurance and capacity requirements above what a typical enterprise server workload needs. For an investor, NAND remains a cyclical, capital-intensive commodity business already consolidated among a small number of large integrated device makers that produce both the die and, through captive or partner controllers, finished drives — so pricing tends to move with the industry’s broader supply cycle rather than with AI demand in isolation.

AI delta: Frequent checkpointing during training and local weight caching for inference push NAND endurance and capacity needs above typical enterprise storage workloads.

Who makes it

The full entry — sub-categories, news, videos and sources — lives on the NAND flash page.