AI's data compounds — but the money lands in memory, not storage

Western Digital's chief product officer has been running a memorable one-liner over the past year: Storage persists it. It is the public face of a bigger argument the storage and memory supply chain is now selling: that AI is less a compute problem than a data problem, that tokens and their derivatives "persist, accumulate, and compound," and that Jensen Huang's five-layer model of AI infrastructure is missing a sixth layer — storage. On the strength of that narrative, the suppliers have re-rated hard: Western Digital is up roughly 180% this year, and Seagate trades near $910 at about 65 times trailing earnings.

The careful investor's first move is to decide whether the claim is engineering and economics-achievable, or positioning. The pitch is only half of what is actually happening — and it is the less valuable half.

A sixth layer called storage

The persistence idea is real at the mechanism level. Agentic systems hold persistent context, and each query leaves artifacts — session state, retrieved context, model checkpoints, synthetic data — that later workloads reference and reuse. That is genuine compounding, and it is why AI data keeps growing. The supply chain is right that data does not vanish when a GPU finishes a request.

But notice what the argument does with that fact. Western Digital's version turns "data persists" into a case that the economics of AI must shift toward the cheapest, slowest kinds of storage — hard drives that retain infrequently used data at scale. That is a marketing structure: a company that sells the media for the long tail arguing that the long tail is where the value now lives, and framing it as if NVIDIA's model were incomplete for omitting it.

What an AI server actually costs now

Check where the cost of an AI server actually sits, in dollars, not in the sales pitch. Memory now accounts for roughly 40–50% of a data-center server's bill of materials, up from about a third a year earlier. Compute's share has fallen from majority to minority; the industry now describes the server as a memory appliance. That is the per-unit inversion underneath the "data compounds" headline: the thing being sold by the truckload is no longer priced by its GPUs but by its fast memory.

It makes sense why. Inference is bound by bandwidth — by moving model weights and context quickly enough — not by how much cold data a machine can store. The binding constraint is high-bandwidth memory and DRAM, not disk space. So the expensive, scarce, compounding asset right now is memory, and persistence is just where the byproducts go when they are cheap enough to keep.

The pricing tells the same story. DRAM costs for AI servers roughly doubled in the first quarter of 2026 and are expected to rise about fourfold across the year — a scarcity story, not a volume story. Micron's latest reported quarter is the realized version: revenue near $24 billion, roughly double the prior quarter, at about a 73% gross margin. The global memory industry is on track to more than quadruple by 2027, past a trillion dollars in revenue. Contrast that with the hard-drive sellers: Western Digital's gross margin sits around 45% — a healthy figure, but it is a volume-and-persistence business, not a pricing-power one.

The efficiency that prunes the pile

There is a second reason to treat the "compounding data" story with suspicion, and it is the counterforce to everything the storage sellers are selling. The bytes a model must move per token are falling on the order of 30% a year, largely through KV-cache compression, sparsity, and aggressive context pruning. The very efficiencies that make inference cheap enough to deploy at scale are also shrinking the data each output leaves behind. The mechanism that generates "data that compounds" is the same mechanism that compresses and prunes it.

That is the test the pitch never answers: whether compounding data grows faster than the efficiency gains erasing it. One trend supports the storage bull case; the other cuts directly against it, and the sellers only ever quote the first.

So the two sides of this trade answer to different standards of proof. Micron's is a pricing supercycle that is already in the reported numbers — realized margins, and a price-to-earnings ratio near 23 only because trailing earnings are already close to their cyclical peak. Its risk is the wall of new memory supply scheduled to land in 2027 and after. Seagate and Western Digital's is a forecast: that compounding data converts into durable storage demand, priced at roughly 65 times trailing earnings and 36 times EBITDA respectively, against a trend — falling bytes per token — moving the other way. Pricing power is a fact. Compounding is a projection. The "data persists" headline pulls you into the trade either way, but it matters a great deal which of the two you are actually paying for.