AI inference generates a storage bill that compounds beyond compute budgets

When executives picture the cost of AI, they tend to picture GPUs: racks of accelerators, enormous power draw, and a waitlist for silicon. An analysis published by MIT Technology Review — produced by storage manufacturer WD rather than the outlet's own editors — argues that this picture leaves out a cost that quietly compounds: storing everything inference generates.

Training retires; inference never does

AI infrastructure planning has traditionally centred on training, which is a bounded exercise. A training run starts, consumes its compute, and ends, so it can be scoped and retired like any other project. Inference behaves differently: it runs continuously, every time a model answers a prompt, and every request leaves residue behind. According to the piece, that back-end data — from billions of daily prompts, images, and videos — is among the harder infrastructure costs to plan for and is routinely missed in initial AI budgeting.

A 130x data multiplier per generation

The headline number comes from WD's own measurements. An eight-second AI-generated video, handed to the user as a 3.97MB 720p file, left more than 500MB of associated data on the back end: logs, cached context, intermediate renders, backups, and safety records. That is roughly 130 times the size of what the user actually received, and at enterprise scale — longer clips, higher resolutions — the piece puts the back-end footprint in the hundreds of gigabytes per video.

Demand-side projections point the same direction. McKinsey, cited in the article, expects AI inference demand to grow at a 35% compound annual rate through 2030, versus 22% for training, accounting for more than 40% of total data center demand by the end of the decade. A WD-sponsored IDC white paper surveying 763 IT and business decision makers found that nearly 95% reported data volume increases from generative AI over the past year, with about 61% seeing growth of 25% or more.

Storage is a GPU utilisation problem

The piece frames slow storage as a compute cost, not just a capacity cost. "The biggest challenge is to ensure that the GPUs are not idle, waiting for data to show up — hours of GPU idle time becomes very expensive, very fast," said a senior product manager at a hyperscale cloud provider, quoted in the report. A model stalled while data loads costs roughly as much as one idling for lack of a GPU, but the bottleneck stays invisible because it surfaces in operating expenses and retrieval delays rather than as a capital line item next to the accelerators.

Archives turn into active inputs

Compute demand rises and falls with training cycles and product launches, but retained data does not shrink back down when it does. Model logs, synthetic outputs, and inference results keep accumulating — and their role changes, because material that once sat untouched in archives can now serve as an input to future AI workloads. In the IDC survey, 96% of respondents said they expect to need faster archive retrieval to support inference or retrieval-augmented generation. Predictably for a hard-drive maker, WD argues that HDDs, with their cost-per-terabyte advantage, are the right foundation for these warm, cool, and cold tiers, with faster tiers reserved for instantly accessed data.

Why it matters

Two caveats are worth stating plainly: the article is sponsored content, and its central measurement comes from WD itself, a company that sells the storage it says enterprises will need. The framing serves the vendor.

Even so, the underlying mechanics are a legitimate budgeting concern for adopters. Compute costs are largely forecastable — a per-generation or per-query rate finance can model. Storage costs are not: they compound with every video generated and every document processed, because none of that back-end data disappears when the task ends. A team that treats storage as a rounding error next to compute may find, within 12 to 18 months, that it has become a significant fraction of total programme cost — right around the time leadership is deciding whether to expand.

The piece closes with three questions most organisations apparently cannot answer confidently: what the retention policy for AI-generated content is and who owns it; whether storage procurement tracks the actual life cycle of the data being produced rather than training-time performance; and whether storage costs would merely double — or worse — if AI usage doubled next year.