Here's a concrete benchmark: one customer running four AI models around the clock writes 9,168 files—roughly 0.68 TB—to object storage every six hours, with each file averaging 75 MiB. Akamai ran the numbers and found the bottleneck isn't the cloud storage bucket at all; it's the client-side instances and network bandwidth. What deserves our attention is that the industry signal here runs far deeper than this single case.
What this is
This is a customer solution Akamai published on its engineering blog: four AI models (likely image or video, since 75 MiB per file is on the heavy side) produce an output file every few seconds. The customer's local arrays overheated, backup windows collapsed, and throwing more servers at the problem didn't help—so they migrated to Akamai's own S3-compatible object storage.
The author built a browser-runnable tool using Node.js + AWS SDK and benchmarked across the Los Angeles and Chicago regions. The conclusion is blunt: storage-layer throughput is nowhere near its ceiling. The question has shifted from "can the platform keep up?" to "how fast do you want to run?"
Industry view
The real signal here is that the AI infrastructure bottleneck is quietly moving from compute to the data pipeline. A text model spitting out a few thousand tokens per call is an entirely different order of magnitude from a multimodal model producing megabytes per frame.
From the bullish angle, cloud storage and cross-region bandwidth budgets will quickly catch up to GPU budgets inside AI projects and become the next cost center under scrutiny.
But we have to flag this: the post is essentially an Akamai solutions-marketing piece, and its methodology and conclusions tilt toward the vendor's own products. A 0.68 TB / six-hour workload isn't exotic for top-tier internet companies—Akamai is really reaching for mid-sized AI teams. There's also a risk the post deliberately soft-pedals: piling all your AI outputs onto a single cloud vendor creates lock-in that's harder to escape than GPU lock-in. Switching storage is cheap; data migration is where the real bill lives.
Impact on regular people
For enterprise IT: over the next year, the share of cloud storage and cross-region traffic budgets inside AI projects will climb noticeably. Stop using the old "GPU rules everything" spreadsheet to calculate total cost of ownership for AI.
For careers: traditional ops and data engineers are becoming hot again inside AI projects. People who understand object storage and data lifecycle management will be scarcer than pure prompt engineers.
For the consumer market: every AI-generated image and video you scroll past is paying for cloud storage on the back end. This is the real reason many AI video products charge per generation instead of daring to offer flat-rate subscriptions.