When companies budget for artificial intelligence, they focus on graphics processing units and compute power—but a new report from Western Digital published October 22 reveals they're missing a massive hidden cost. An eight-second AI-generated video clip delivered to users as a 3.97MB file leaves more than 500MB of back-end data behind—logs, cached context, renders, backups, and safety records—representing a 130-fold increase over what the user actually sees. The finding highlights how data storage expenses, often overlooked in initial infrastructure planning, are quietly becoming one of the hardest costs to predict as inference workloads scale.
The data accumulation problem is accelerating rapidly across enterprises. A WD-sponsored IDC survey of 763 IT and business decision makers found that nearly 95% reported increased data volume from generative AI over the past 12 months, with roughly 61% seeing growth of 25% or more. McKinsey projects AI inference demand will expand at a 35% compound annual growth rate through 2030, compared to 22% for training, making up more than 40% of total data center demand by decade's end. For longer, higher-resolution enterprise video generation, the back-end footprint swells into hundreds of gigabytes per clip. The same IDC survey revealed that 96% of respondents anticipate needing faster archive retrieval to support inference or retrieval-augmented generation.
The report emphasizes that inference runs continuously every time a model answers a prompt, unlike training which is finite and can be budgeted as a discrete project. A senior product manager at a hyperscale cloud service provider quoted in the report said keeping GPUs busy requires thinking of storage "as an optimization"—noting that "hours of GPU idle time becomes very expensive, very fast" when processors wait for data to arrive. The report finds that if storage spend grows 25% year over year while compute budgets follow more predictable cycles, storage ends up outpacing the budget line receiving the most attention, yet the storage bottleneck stays hidden until it surfaces in operating costs and retrieval delays rather than capital expenditure lines.
The report explains that while compute demand rises and falls with training cycles and hardware refreshes, retained data volume doesn't decrease when compute demand slows—model logs, synthetic outputs, and inference results keep accumulating. Information that once moved into archives and rarely surfaced can now become input for future AI workloads, changing the role of retained data. A team that budgets AI as "compute plus a rounding error for storage" may discover in 12 to 18 months that the rounding error has become a significant fraction of total program cost, right when leadership decides whether to expand. The report illustrates this with concrete scenarios: a retailer generating AI product videos for its catalog or an insurer running claims documents through a retrieval-augmented generation pipeline both face compute costs that are known and budgeted, but storage costs compound with every video generated and every claim processed because none of that back-end data disappears once the task finishes.
The report recommends organizations ask three questions before the next budget cycle: What's our data retention policy for AI-generated content and who sets it? Is storage procurement tied to training-time performance needs or to the actual life cycle of the data we're producing? If AI usage doubled next year, would storage cost double with it, or worse? Most organizations cannot answer all three with confidence, according to the report, which argues that enterprises need economically scalable capacity—infrastructure designed for the full AI data life cycle that treats archives as active assets rather than cold storage. The report concludes that addressing this challenge requires recognizing that storage decisions play out across fleets where power consumption, reliability, density, and operational efficiency determine whether systems remain economically sustainable as capacity scales. Organizations that scope AI infrastructure spend too narrowly on compute will find a model idling while data loads costs about as much as one idling for lack of a GPU, but the problem won't become visible until it's already eating into margins. The strategic choice isn't whether to invest in storage architecture, but whether to design for the compounding data reality before or after it becomes a budget crisis that derails expansion plans.