NAND TAM growth is slower and more cyclical than DRAM or HBM, driven primarily by bit demand growth (~30–40% annually) partially offset by $/GB price declines as layer counts increase. The base case reflects moderate enterprise SSD expansion driven by AI storage workloads, steady smartphone content growth, and disciplined supply from major NAND makers post-2024 correction. Bear case assumes oversupply returns in 2027–2028 as all major suppliers ramp 300+ layer nodes simultaneously, driving ASP declines that outpace bit demand growth. Bull case assumes enterprise NAND becomes a critical AI infrastructure component — warm storage tiers for RAG (retrieval-augmented generation) systems create a step-change in enterprise SSD attach rates, and automotive NAND demand accelerates faster than expected with ADAS adoption.
“Enterprise SSD demand accelerating as hyperscalers expand AI training infrastructure. Each LLM training run requires petabytes of high-throughput storage — enterprise NAND is the primary medium. QLC SSDs gaining share at cloud scale due to lower $/GB vs. TLC.”
“NAND contract pricing recovering from the 2023–2024 correction but pace is uneven. Enterprise SSD pricing firming; consumer NAND (PC, mobile) still facing mild oversupply. Kioxia and WD/Sandisk reducing wafer starts to support price stabilization — discipline holding so far.”
“SK Hynix 321-layer 4D NAND achieving volume production with 30% higher bit density vs. prior gen. Samsung and Micron on similar roadmaps. Higher layer counts compress $/GB, expanding NAND into workloads previously served by HDDs — particularly warm storage tiers at hyperscalers.”
“Consumer NAND demand soft. Smartphone NAND content per device growing modestly but unit shipments flat. PC NAND recovery slower than DRAM — average SSD capacity per PC growing but ASP pressure from oversupply partially offsets volume gains.”
“Automotive NAND demand growing as ADAS systems require high-endurance, wide-temperature-range flash storage for map data, sensor logs, and OTA update buffers. Content per vehicle growing ~25% annually; still a small fraction of total NAND TAM but a durable long-tail growth driver.”
Every foundation model training run requires petabytes of high-throughput storage for datasets, checkpoints, and activations. Hyperscalers are building dedicated AI storage clusters using enterprise SSDs — a workload that barely existed three years ago. As model sizes grow and training runs lengthen, NAND storage requirements scale super-linearly. Enterprise SSD is the primary beneficiary; HDD warm storage competes but NAND wins on latency and throughput.
Agentic AI systems using retrieval-augmented generation require persistent, low-latency vector stores and document archives. At scale, RAG infrastructure is a new enterprise NAND workload — always-on, randomly accessed, high endurance. If RAG becomes standard architecture for enterprise AI deployments, it creates a sustained pull for high-endurance enterprise SSDs beyond the training data use case.
QLC (4 bits per cell) NAND delivers the lowest $/GB of any flash tier, making it viable for hyperscaler warm storage at a cost close to HDD. Micron, Samsung, and SK Hynix are all scaling QLC production. If QLC achieves the endurance and reliability thresholds required for cloud workloads, it displaces a portion of the HDD installed base and significantly expands the total addressable NAND market.