Technology
A training cluster is an extremely expensive machine that must never wait. Storage earns its place in an AI system by delivering training data fast enough to keep thousands of accelerators fed and by absorbing checkpoints quickly enough that saving state is not a visible pause.
In one sentence
AI storage is the high-throughput file and object storage layer that holds training datasets, model checkpoints and served model weights, sized for aggregate bandwidth to many nodes at once rather than for single-stream speed.
Two access patterns dominate, and they are opposites. Training data is read as an endless stream of small-to-medium objects in random order, by every node at once. Checkpoints are enormous writes that happen periodically, all at the same instant, from every node at once — a burst the storage system must absorb without stalling the run.
Both are aggregate-bandwidth problems rather than latency problems, which is why AI storage looks different from the storage behind a database: flash, parallel access from many clients, and a network fast enough that the storage rather than the wire is the limit.
A long training run will lose nodes. The standard defence is to save the full model and optimiser state periodically so a failure costs one interval rather than the whole run. That state is large, every node writes at once, and everything stops while it happens — so checkpoint write bandwidth translates directly into wasted accelerator time.
Each step consumes a batch of examples that must be fetched, decoded and transformed before the accelerator needs them. If the pipeline falls behind, the most expensive hardware in the building idles. Practical systems overlap loading with computation and cache aggressively on local flash.
In production the dominant storage question is how fast a large set of model weights can be loaded onto an accelerator — when starting a service, recovering from a failure, or switching which model a node hosts. Minutes of load time per node is a real constraint on how dynamically a fleet can be scheduled.
Technology dependencies are solved by engineering; supply dependencies are solved by building something, which takes years.
The bandwidth these systems need is only economic on flash, which ties AI storage to the memory industry's other capacity cycle.
NAND and SSDsStorage bandwidth is delivered over the same fabric as everything else; a storage tier faster than its network is a storage tier that cannot be used.
AI networkingFrameworks, checkpointers and data loaders all speak a small number of common storage interfaces. Without them every storage vendor would need its own connector in every framework, and the fleet could not be re-platformed.
Other pages in this map that name AI storage as something they cannot do without.
What each company supplies at this step, and — where a public figure exists — its share of this specific market — with what that share measures, the period it covers and who published it. Some rows also show the company’s own reported revenue for the segment covering this step, which is a different thing: it says how much this business matters to that company, not how much of the market it holds. Not a ranking and not a recommendation.
Supplies the high-throughput flash storage platform used behind large training clusters.
Supplies enterprise file and object storage used as the data layer for AI clusters.
Supplies the drives and media underneath large AI data lakes.
Cloud segment — 88.9% of Western Digital’s reported revenue ($11.5B)FY2026 · company filings · not a market share
Supplies high-capacity drives used for the cold tier of AI datasets.
Supplies the parallel file system used to keep accelerators fed during training.
Supplies the storage arrays and the servers around them as one system.
Supplies the high-throughput storage used in most national-scale AI clusters.
Supplies Storage Scale, the parallel file system under a large share of national and research clusters.
Holds most of the world's training data as objects, and rents the parallel file layer that stages it to the cluster.
Supplies the object store teams run themselves when the data cannot leave their own building.
Supplies the file layer that presents one namespace across sites, so a run does not wait on a copy.
Supplies the flash under the fast tier; every platform above is assembled from someone's NAND.
Supplies the enterprise drives most of these platforms are populated with.
Whether checkpoint frequency rises with cluster size to the point where write bandwidth becomes a design constraint on run length.
Whether serving fleets that swap models frequently make weight-load time a purchasing criterion.
Whether the cold tier stays on spinning disk as datasets grow, or flash economics take it too.
Each page explains one technology in plain language, states what it depends on, and names companies by what they supply at that step. Company roles are described qualitatively and deliberately carry no market shares, revenue figures or rankings — those change faster than an explainer can, and a stale number is worse than none. Ticker links point at company pages on this site and are provided for reference only.
Nothing here is investment advice, a recommendation, or a forecast. A company named on a page about a technology is not thereby a good investment, and the chokepoints described are structural facts about supply chains rather than predictions about prices. Technology moves; where a page describes something as unresolved or in development, that was true when it was written.