TL;DR
AI inference creates logs, cached context, renders, backups and safety records in addition to the result delivered to a user. A Western Digital measurement cited by MIT Technology Review found more than 500MB of back-end data associated with an eight-second video returned as a 3.97MB file. The report argues that growing, retained data can add capacity and retrieval costs that AI budgets focused on compute may miss.
AI systems can leave a substantial volume of data behind each response, adding a storage burden that may be missed when organizations plan AI infrastructure mainly around processors and computing costs, according to a MIT Technology Review report published October 22. In one measurement from Western Digital cited in the report, an eight-second AI video delivered as a 3.97MB file was associated with more than 500MB of back-end data, including logs, cached context, renders, backups and safety records.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The measurement illustrates the difference between the file a user receives and the data created or retained to produce and manage it. The report says the 500MB-plus back-end footprint was about 130 times the delivered file size. That figure comes from WD’s internal measurement of a particular video workload; it is not a universal ratio for all AI systems or outputs. The report says longer, higher-resolution enterprise video could produce footprints in the hundreds of gigabytes per clip, but does not provide a measured example for those larger workloads.
The report also cites a McKinsey projection that AI inference demand will grow at a 35% compound annual growth rate through 2030, compared with 22% for training. McKinsey projects inference will account for more than 40% of total data-center demand by the end of that decade. These are forecasts, not observed future outcomes. A WD-sponsored IDC survey of 763 IT and business decision-makers found that nearly 95% reported higher data volumes from generative AI over the prior 12 months; roughly 61% said growth was at least 25%.
Inference takes place whenever a model responds, unlike a training run that has a defined start and end. The report says the associated records and outputs can accumulate after a request is complete, with organizations retaining them for operational, safety or later-use purposes. It cites a senior product manager at a hyperscale cloud service provider saying storage helps keep GPUs supplied with data, because idle accelerator time is expensive. The report does not identify the manager by name.
Storage Can Outlast Compute Spikes
For organizations running AI at scale, the concern is not simply how large the immediate output is. Retained logs, generated media and other supporting data can add up across repeated requests, while storage needs may persist even when a training project or product launch ends. That can make storage a recurring operating cost rather than a one-time supplement to compute spending.
The report argues that storage constraints can also affect performance. If data cannot be retrieved quickly enough, processors may wait instead of working. A senior product manager quoted in the report described keeping GPUs from sitting idle while waiting for data as a central challenge. That connection makes storage planning relevant to both capacity budgets and accelerator utilization.
The scale and cost will vary by application, retention choices and system design. Still, the examples point to a budget risk: teams forecasting AI expenses only through per-query or per-generation compute costs may not account for the data accumulated behind those activities. The report suggests those costs could become more visible as organizations expand deployments, although it provides no general cost estimate or documented forecast for when storage will overtake compute budgets.
Products Worth Considering
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Training Runs to Continuous Inference
AI infrastructure planning has often focused on training, which consumes substantial compute during a bounded run. Inference is different: it continues as users submit requests, and each request may create outputs and operational records. The report’s central point is that the data footprint follows ongoing use, not just model development.
Some of that data may later serve as an input to other work. The report describes archives becoming more active as organizations reuse retained information for inference and retrieval-augmented generation, or RAG. In the WD-sponsored IDC survey, 96% of respondents anticipated needing faster archive retrieval to support inference or RAG. That result records respondents’ expectations, not a measured industry-wide increase in retrieval speed or demand.
The report presents a tiered storage approach as one possible response: fast storage for frequently accessed data and lower-cost, higher-capacity options for large volumes retained over time. It highlights hard disk drives as a cost-per-terabyte option for high-capacity retention. This is a storage-industry argument in a report that cites WD data and a WD-sponsored survey; the source does not provide an independent comparison of specific technologies or a universal recommendation for every workload.
The Cost Depends on What Stays
The report does not establish how much storage costs will rise across the industry, or how much of that cost will be passed on to customers. Its video measurement is specific to a WD internal workload, and the survey records decision-makers’ reported experiences and expectations. Neither establishes a standard amount of back-end data per AI request.
Organizations’ retention policies, model designs, output formats and access patterns can all affect the volume kept and the expense of retrieving it. The report does not quantify how much generated or operational data companies delete, compress or reuse, nor does it give a representative comparison of storage costs against compute costs. The precise budget impact—and whether storage becomes a bottleneck in a particular deployment—remains dependent on those details.
Planning Around the Data Lifecycle
The report calls for organizations to plan storage across the full life cycle of AI data, rather than sizing infrastructure only for immediate, frequently accessed workloads. That includes deciding what to retain, how long to retain it, how quickly it may need to be retrieved and whether archived material will support future inference or RAG.
For finance and technology teams, the practical next step is to track storage alongside compute as deployments grow: measure the data generated per workload, the volume retained over time, retrieval requirements and the cost of each storage tier. The report offers no specific timetable or industry-wide standard for these decisions. As inference volumes rise, future measurements of real-world data retention and total system costs will show whether the storage burden matches the concerns raised in the report.
Key Questions
What does AI inference leave behind?
Beyond the result shown to a user, a system may generate or retain logs, cached context, intermediate renders, backups and safety records. The mix and amount depend on the workload and its data policies.
How large was the video example in the report?
WD reported that an eight-second clip delivered as a 3.97MB, 720p file had more than 500MB of associated back-end data. The report describes this as about 130 times the output file size; it is a measurement of that workload, not a general rule for AI video.
Does the report show that storage costs now exceed compute costs?
No. It argues that storage may be undercounted in AI budgets and could grow as data accumulates, but it does not provide evidence that storage spending already exceeds compute spending across the industry.
Why might archived AI data need faster access?
Organizations may reuse retained information in later inference or retrieval-augmented generation workflows. The WD-sponsored IDC survey found that 96% of respondents anticipated a need for faster archive retrieval for inference or RAG.
Source: rss
