A cost outside the frame
When people talk about the cost of artificial intelligence, the recurring image is that of GPUs, accelerator racks, and waitlists for silicon. According to an in-depth feature published by MIT Technology Review and produced by WD, this representation overlooks storage: much of the back-end data generated every day by prompts, images, and video must be retained, and the expense of doing so is among the hardest to plan for.
The starting point is the difference between training and inference. Training is finite: it begins, consumes compute, and ends, so it can be treated as a project with a defined budget. Inference, by contrast, does not let up: every request leaves behind material to be archived, and the volume grows as use of the models spreads.
The gap between what the user sees and what remains
The most concrete example concerns AI-generated video. Based on internal measurements, WD reports that an 8-second clip, delivered to the user as a 3.97 MB 720p file, leaves more than 500 MB of associated data on the back end: logs, cached context, intermediate renders, backups, and security logs. That is an increase of about 130 times compared with the visible file. For a company producing longer videos at higher resolution, the back-end footprint can reach hundreds of gigabytes per clip.
The phenomenon is part of a broader growth. McKinsey forecasts that inference demand will increase at a compound annual rate of 35% through 2030, compared with 22% for training, exceeding 40% of total data center demand by the end of the decade. A global IDC white paper sponsored by WD, based on a survey of 763 IT and business decision-makers, reports that nearly 95% of respondents have seen data volumes increase because of generative AI over the past 12 months, and about 61% have seen growth of 25% or more.
The resource that does not shrink when compute declines
Demand for compute fluctuates with training cycles, product launches, and hardware upgrades, but the volume of stored data does not decrease in the same way: logs, synthetic outputs, and inference results continue to accumulate and are retained longer. This changes the role of storage, because information once destined to remain inactive can become input for new workloads, including retrieval-augmented generation.
The economic consequence is that storage can exceed the most closely watched budget line. A product manager at a cloud provider quoted in the report sums up the issue: more storage is needed to keep GPUs busy, because idle hours waiting for data quickly become expensive. The limitation of the analysis is that the figures cited come partly from WD's internal measurements and from a survey sponsored by the company, so they reflect the perspective of a storage vendor. It remains to be seen how these ratios translate into actual costs for individual organizations.
