As AI scales, the industry’s cost conversation is shifting from compute to the data that powers it.
Much of the conversation about AI economics has centered on tokens: the cost per million tokens, tokens per dollar of GPU time, tokens generated per watt. These measures capture something real, but in my role leading strategy and corporate development at WD, the numbers we track most closely sit one layer beneath them, in the volume, movement, and value of the data that AI both consumes and produces. That is the focus of new research WD commissioned from IDC, a survey of 763 IT and business decision-makers across seven countries. Viewed through that lens, token economics looks less like the full picture and more like a useful simplification of a larger one.
The scale of that larger picture is significant. IDC’s Global DataSphere projects that the world will generate 274 zettabytes of data in 2026, growing to more than 718 zettabytes by 2030.* Installed storage capacity, by comparison, stands at roughly 13 zettabytes today and is expected to more than double, to over 26 zettabytes, over the same period. The gap between what is created and what is retained is not incidental. It reflects how the AI era is reshaping the value of data: much of what is generated will not be stored, and the portion that is retained is increasingly asked to do more, serving as training input, as retrieval context, as a compliance record, or as a resource for use cases not yet defined. Determining what data is worth keeping, and where it should reside, has become a strategic capital allocation question rather than an operational one.
Token economics describes one important part of that question well: the compute-intensive phase of AI, where speed is the primary constraint. The IDC survey provides a view into how storage priorities shift across the broader AI life cycle. Throughput and latency are the top storage priority in exactly one phase according to 45% of the respondents: model training. Across every other phase, including data ingestion, inference, storage of AI-generated data, and long-term retention, reliability leads or ties performance as the primary consideration by survey respondents. A senior product manager at a hyperscale cloud service provider described the practical stakes of this in one of IDC’s interviews: “You can think of having more storage as an optimization for keeping the GPUs busy. … hours of GPU idle time becomes very expensive, very fast.” Token-level costs can be understood as a downstream consequence of these infrastructure decisions, not the starting point for making them.
A related shift is underway in how organizations think about the life cycle of data itself. The traditional model assumed that data cools steadily over time, moving from hot to warm to cold and eventually becoming inactive. IDC’s research suggests that model is increasingly incomplete. In the survey, 74.3% of organizations reported that AI adoption has caused them to retain data for longer periods; 75.9% reported bringing archived, cold-tier data back online to support AI workloads; 96% expect they will need faster retrieval from that archived data to support inference or retrieval-augmented generation (RAG). In practice, data is not cooling in a single direction. It is warming and cooling repeatedly: information used for a training run or a RAG query may be set aside, then reactivated months later for a use case that did not exist when it was first captured. One IT decision-maker in storage distribution and systems integration described how this plays out operationally: organizations “deploy large SSD arrays for model training—then post-training they dismantle the SSDs and install HDDs at the back as backbone storage for inference.” That reflects a life cycle decision made asset by asset, not a static tiering policy.
These dynamics are also part of why AI infrastructure and cloud infrastructure are increasingly difficult to plan as separate initiatives. They draw on the same capital, often concurrently. One respondent in the same study estimated that for every unit of new compute deployed, organizations need two to three times more storage than in the past, not as a future requirement but as an immediate one. Separately, more than 90% of respondent organizations reported buying or selling more datasets than in previous years, with nearly half doing both. Data is increasingly treated as a tradable, reusable asset, which suggests that the infrastructure supporting it should be planned with the same strategic intent, rather than managed solely as a cost to minimize.
None of this diminishes the importance of compute. The findings are consistent with a view WD has long supported: the more useful framing is not HDD versus SSD, or storage versus compute, but a tiered system optimized end to end, from power and GPU utilization through every storage layer. The economics of AI at scale are shaped by the weakest link in that system, not the fastest one. What is changing is which layer receives strategic attention. For much of the past decade, that was compute. The findings in this study point to data, and the infrastructure that supports it, as the next major planning priority for technology and business leaders alike.
For WD, this reinforces why we view storage as a strategic component of AI infrastructure planning, not a downstream procurement decision. We commissioned this research with IDC to bring an independent, data-driven perspective to how these dynamics are unfolding across industries. The full study, including the underlying survey data and IDC’s complete analysis, is available here. As AI continues to scale, we believe the organizations that plan for the full data life cycle, and not just its most visible phase, will be best positioned to turn that scale into lasting value.
Sourcing note: statistics and quotations are drawn from IDC White Paper, Sponsored by WD, “Built for Scale: The Enduring Role of HDDs in the AI Era,” Doc. #US54789026, September 2026 (survey of 763 IT and business decision-makers across seven countries). Quote attributions use IDC’s published role identifiers; no individual respondents are named, consistent with IDC’s double-blind survey methodology and named-but-not-publicly-attributed interview protocol.
* Source: IDC Global DataSphere Forecast, 2026–2030
