AI Inference Leaves 130x More Data Than Output, Driving Hidden Storage Costs
Updated
Updated · MIT Technology Review · Oct 22
AI Inference Leaves 130x More Data Than Output, Driving Hidden Storage Costs
3 articles · Updated · MIT Technology Review · Oct 22
Summary
An 8-second AI video clip returned as a 3.97MB file can leave more than 500MB of back-end data—about 130 times the user-visible output—adding a large storage burden after each inference request.
That footprint keeps compounding because inference runs continuously, unlike finite training jobs, with logs, cached context, renders, backups and safety records accumulating long after compute demand ebbs.
McKinsey projects inference demand will grow at a 35% CAGR through 2030 versus 22% for training, and exceed 40% of total data-center demand by decade-end.
A WD-sponsored IDC survey of 763 decision makers found nearly 95% saw generative AI lift data volumes over the past year, while 96% expect faster archive retrieval will be needed for inference or RAG.
The report argues CIOs and CFOs should budget AI around the full data lifecycle—not just GPUs—using tiered, economically scalable storage as retained data becomes a reusable asset.