AI's data deluge is keeping Seagate's hard drives spinning

Roughly 90% of the storage Seagate ships now goes to datacenters. That figure, which chairman and CEO Dave Mosley repeated on the company's Q4 FY2026 earnings call, is the structural answer to a question Wall Street stopped asking years ago: who actually buys the spinning disk in the age of the GPU?

The quarter ended July 3 produced the receipts. Revenue hit $3.6 billion, up 50% year-on-year. Full-year FY2026 revenue reached $12.2 billion, up 34% from $9.1 billion. Non-GAAP gross margin hit a record 52.7%, up from 37.9% the year before. Free cash flow for the year was $3.105 billion, also a record. Seagate retired $1.4 billion of debt and returned $810 million to shareholders (TechTimes). The next quarter is guided to $4.1 billion, plus or minus $100 million, which would imply 56% year-on-year growth.

The "why now" is a tiered-storage argument that has very little to do with training. Inference, especially agentic inference, does not just call a model; it keeps the conversation. Each user turn produces a key-value cache, a working memory of tokens the model has already attended to, and that cache has to live somewhere between turns. Cheaper memory (DRAM) and faster flash (NVMe SSDs) hold the hot, latency-sensitive portion. The cold bulk, the layer that holds multi-hour transcripts, tool-call histories, retrieval corpora, and the long tail of agent state, sits on mass-capacity nearline drives. Seagate has published on exactly this tiered-storage split. The consequence is that bulk-capacity demand is now a function of how many agentic sessions are running, not how many foundation models are being trained.

Capacity per drive is doing the work, not unit volume. CFO Gianluca Romano told analysts that drives now contain 15% to 20% more disks and heads than a year ago, while unit volumes are "absolutely flat." Each disk pack holds more platters (HAMR, heat-assisted magnetic recording, lets the read-write head fly closer to the surface and pack bits tighter) and each drive holds more terabytes. The Mozaic 4 platform, which uses HAMR, is ramping with major cloud providers and supports drives up to 44 TB. Mozaic 5, which Seagate says will deliver more than 5 TB per platter, remains on track for qualification shipments in late 2027.

Mosley said the "vast majority" of nearline exabytes are now allocated into calendar 2028 under long-term supply agreements, and that "we are not seeing customers pull back on planning horizons." Hyperscalers are reserving capacity years in advance, a behavior more associated with commodity chemicals or LNG than with consumer storage. The lead-time signal in the market is consistent: TrendForce reported in September 2025 that nearline HDD lead times had stretched from weeks to more than 52 weeks, and that QLC SSD contract prices were projected to rise 5% to 10% quarter-on-quarter in the fourth quarter of 2025 as North American cloud providers considered SSDs for cold-data tiering.

Three structural caveats qualify the picture. Seagate's SSD business is, in The Register's phrasing, "utterly trivial"; the company is structurally concentrated on nearline HAMR drives sold to a small set of hyperscalers and large enterprises. The mechanical limits of HDDs, millisecond seek times, rotational latency, and failure modes that flash largely avoids, do not disappear because the workload is cold. And the AI storage buildout, like the AI compute buildout it shadows, depends on a capex regime that has its own cycle: if model-training spend cools, the inference-tier argument survives, but the surrounding data-center footprint may not. As of the Q4 call, Mosley said demand is still growing, and 2028 allocations imply the company is not yet seeing a pullback.

The next tell is Mozaic 5. Qualification shipments are slated for late 2027; if they land on schedule and at the per-platter densities Seagate describes, the capacity-per-drive curve keeps compounding, and the mass-capacity tier keeps widening the gap between the GPU and HBM headline and the storage revenue that funds it.