Over the past two years, whenever AI compute came up, investors' first reaction was almost always GPUs: Nvidia's shipment pace, TSMC's advanced process nodes, HBM capacity schedules. But entering 2026, market attention is quietly broadening. As large models shift from “training” to “inference,” storage is moving from a supporting role in compute to the second battlefield that decides the efficiency of the entire AI infrastructure. For US investors, the storage sector's story deserves to be retold.
1. In the Inference Era, GPU “Starvation” Exposes the Storage Bottleneck
When training large models, data is mostly loaded into the cluster in one pass, so the bottleneck lies mainly in chip compute and interconnect bandwidth. Inference, however, is an around-the-clock online service with massive concurrency: every moment, data must be moved from disks to memory and then into HBM to feed the GPU. If data movement can't keep up, even the strongest GPU just idles—this is the “memory wall” problem the industry often mentions.
More importantly, inference has spawned several brand-new storage demands:
- KV Cache: Large-model inference needs to cache intermediate results; the longer the conversation and the larger the context, the more memory and cache space is consumed, directly boosting demand for high-capacity memory;
- Vector databases and RAG retrieval: Enterprises deploying AI applications widely adopt retrieval-augmented generation, which requires vectorizing massive documents, storing them, and retrieving them with low latency—directly driving enterprise SSD purchases;
- Multimodal data lakes: Unstructured data such as video, images, and audio are the raw material for AI training and fine-tuning; data lake capacity is growing in exabytes, making cold data storage an unexpected hit.
In other words, inference turns both “storing enough” and “fetching fast” into hard requirements, giving every layer of the storage chain—HBM, DRAM, NAND flash, enterprise SSDs, nearline HDDs—a place of its own.
2. Beyond HBM, NAND Becomes the Second Main Battlefield
If HBM is the “crown” of AI storage, then NAND flash is the “foundation” now being repriced by the market.
Enterprise SSDs Enter the 100TB Era
AI data lakes care most about cost and power per unit of capacity, and QLC flash hits that pain point squarely. Leading vendors' enterprise QLC SSDs have already surpassed 100TB per drive, with power consumption far below hard-drive arrays of the same capacity, and they are rapidly replacing traditional HDDs in data centers. Large-model corpora, training checkpoints, and vector stores are all moving onto high-capacity SSDs.
The supply side matters just as much. Around 2023, storage makers went through a round of deep losses and broadly tightened NAND capital spending, enforcing produce-to-order capacity discipline; after 2025, AI demand suddenly surged, the supply-demand gap widened rapidly, and contract prices kept climbing. Micron has even announced it will gradually shrink its consumer retail business and tilt capacity toward data centers—the big storage makers' “bet on AI” stance is now unmistakable.
Nearline HDDs Make an Unexpected Comeback
Interestingly, hard drives—widely seen as a “sunset” technology—have also found a second spring in the AI wave. Hyperscale cloud providers' cold and archived data still rely heavily on high-capacity nearline HDDs, per-drive capacity is iterating beyond 30TB, and order visibility for Seagate, Western Digital, and their peers has clearly improved. The storage-compute division of labor is simple: hot data races for speed on SSDs; cold data wins on cost with HDDs.
3. An Investment Map of the US Storage Supply Chain
For everyday investors, the storage rally can roughly be followed along three threads:
- HBM and DRAM leader: Micron is the only US-listed storage maker spanning all three markets—HBM, DRAM, and NAND. Its HBM products have entered the supply systems of top AI chip customers, making it one of the names with the greatest earnings elasticity in the storage rally;
- The NAND camp: SanDisk, after spinning off from Western Digital, focuses purely on flash for clean exposure; Kioxia, though not US-listed, keeps influencing sector sentiment as its capacity and pricing moves pass through industry prices;
- The HDD duo: Seagate and Western Digital command an oligopoly in nearline drives; AI data lakes lift volume and price together, and both companies maintain steady dividends and buybacks, offering both cyclical and income attributes.
In addition, the order pace of server makers such as Dell and Super Micro, along with capex guidance from the major cloud giants, are high-frequency weathervanes for storage sentiment—after all, storage demand is the “downstream echo” of AI capital spending.
4. After the Hot Money, Don't Forget Storage's Cyclical Nature
A reminder: storage is a classic strongly cyclical industry, and the script of “shortage—price hikes—capacity expansion—glut” has played out repeatedly in history. The sustainability of the current rally depends on three things:
- Whether AI inference demand keeps beating expectations: If the ramp-up of agent applications and multimodal inference slows, the elasticity of storage prices will fade quickly;
- How restrained makers stay about expansion: Once capital spending spins out of control again, the supply-demand balance could reverse rapidly, and the price-hike cycle could end at any moment;
- Technology-route variables: New technologies such as CXL memory expansion and processing-in-memory may reshape storage's competitive landscape and valuation logic over the medium to long term.
To sum up in plain words: in the edifice of AI compute, everyone used to stare only at the “engine” (GPUs); now we find that the “warehouse and conveyor belts” (storage) can be just as much of a bottleneck. HBM is the open card; NAND and HDD are the hidden laggards catching up. For novice investors, rather than chasing a single leader at elevated prices, it is better to treat storage as the “second engine” of an AI compute portfolio, while closely watching two leading indicators—makers' capital spending and cloud vendors' orders—to sense cyclical turning points in time. The storage story has only reached its midpoint, but the iron law of cyclical industries never stays away: keep tracking and respect the cycle—that is the way to endure.
