Moving a GPT-3 scale training set once costs roughly $3,875 in egress fees. Moving the LAION-5B dataset once costs roughly $15,800. Neither number includes the engineering time, the re-architecture, or the second move you'll eventually need to make.
Moving a GPT-3 scale training set once costs roughly $3,875 in egress fees, at AWS's standard tiered rates. Moving the LAION-5B dataset, the 240 terabyte corpus behind Stable Diffusion, costs roughly $15,800 once. Those are one-time numbers, for one move, at relatively modest AI-era scale. Enterprises operating at petabyte scale are looking at bills in the tens or hundreds of thousands of dollars every time a dataset needs to relocate, and in AI infrastructure, datasets relocate constantly: to a new GPU provider, a new region, a new training run.
This is data gravity, and it is quietly becoming the single most expensive architectural constraint in AI infrastructure. Compute gets the budget attention. Storage decides whether that compute can actually reach the data it needs, and increasingly, whether it can afford to.
The term was coined by Dave McCrory in 2010: large concentrations of data attract applications, services, and infrastructure toward them, the way mass attracts mass under gravity. The larger the dataset, the stronger the pull, and the more expensive and complex it becomes to move. Applications end up migrating toward the data rather than the other way around, because moving the data is the costlier option every time.
In AI infrastructure specifically, this plays out as a direct architectural tension. You want the cheapest or fastest available GPU capacity, which this week might be at a different provider than last week. Your training data, however, is sitting wherever you first landed it, often a hyperscaler you chose years before AI workloads were part of the plan. Every time compute and data are not in the same place, one of them has to move, and moving data is almost always the more expensive option.
Cloud was supposed to make data portable. In practice, it concentrated data inside a small number of providers who charge specifically for the act of moving it back out. AI has intensified every dimension of that problem at once.
Training and inference pipelines now routinely work with petabyte-scale corpora, logs, embeddings, and checkpoints. A petabyte is roughly 13 years of continuous high-definition video. Moving that volume is a materially different problem than moving a few terabytes of application data ever was.
Enterprises may need AWS today, a specialized neocloud tomorrow, and a different provider next quarter depending on what capacity is actually available. Every provider switch drags the training data along with it, and the egress bill on that move can exceed whatever savings motivated the switch in the first place.
Agents that query data, trigger processes, and coordinate across systems need fast, at-scale access to enterprise context spread across multiple environments. Each additional system an agent touches increases both governance risk and the pull toward centralizing that data somewhere, which reintroduces gravity even when the original goal was distributed access.
Roughly 80 percent of enterprise data still sits on-premises, according to industry analyst commentary on enterprise storage patterns, while the AI tools that need it run in the cloud. Bridging that gap is not a one-time migration, it is an ongoing tax on every workload that needs both sides.
Egress fees are the most visible cost, but far from the only one. Gartner estimates that organizations spend 10 to 15 percent of their total cloud bill on egress charges, against a global public cloud spending base Gartner projected to exceed $900 billion in 2025. Cloudflare has separately documented that AWS's egress fees represent nearly an 8,000 percent markup over wholesale transit costs in North American markets, a figure that is hard to explain as pure cost recovery rather than a deliberate toll on data leaving the platform.
The cost is not only the transfer itself. Data gravity forces architectural compromises that show up elsewhere in the budget: duplicated backup and disaster recovery infrastructure maintained in place rather than consolidated, compute positioned suboptimally because the data can't move to where cheaper or faster compute lives, and abandoned or delayed migrations because the egress bill alone kills the business case. Flexera's 2026 State of the Cloud Report found 53 percent of organizations now cite security and compliance risk with cloud-based AI as their top challenge, a governance burden that compounds every time data sprawls across another environment to chase available compute. This connects directly to why storage functions as the anchor of the AI infrastructure stack rather than a detail to sort out after compute is chosen.
The math is straightforward once you put real numbers against it.
| Dataset | Size | One-Time Egress Cost |
|---|---|---|
| GPT-3 scale training set | ~45 TB | ~$3,875 |
| LAION-5B (Stable Diffusion data) | ~240 TB | ~$15,800 |
| Mid-size enterprise data lake | 500 TB | $40,000+ |
These are one-time costs, calculated at standard hyperscaler tiered rates. In practice, AI infrastructure rarely involves a single move. A dataset that shifts between a hyperscaler and a neocloud for training, then back for inference, then out again for a provider renegotiation, pays that egress bill multiple times over the life of the project. At petabyte scale, an organization moving even 5 percent of a 50 petabyte estate annually is looking at a bill that can run into eight figures.
The most direct fix is architectural: run training and inference workloads against data where it already lives, rather than centralizing data to chase available compute. Query-in-place engines, external table formats, and data lakehouse architectures increasingly make this practical without duplicating the dataset itself.
Zero-egress providers exist specifically to remove this cost from the equation. Backblaze B2 and Wasabi both offer S3-compatible object storage without the egress penalty that turns a routine provider switch into a five-figure surprise. For any dataset likely to move between GPU providers over its lifetime, this is not a nice-to-have, it is the difference between a switch that pencils out and one that doesn't.
Not every dataset needs to be positioned for frequent movement. Hot training data and active inference caches justify premium, high-bandwidth positioning. Cold archives and infrequently accessed logs do not, and tiering them appropriately reduces both storage cost and the blast radius of any future egress bill.
Design new AI data pipelines with platform independence as a first-class requirement rather than an afterthought. Avoiding proprietary transformation tools and APIs where open alternatives exist keeps the option to move open, even if you don't exercise it immediately.
Model the recurring cost of data movement between platforms over the life of a workload, not just the cost of the initial migration. Architectural decisions made today generate egress bills for years, and treating that as a sunk one-time cost rather than an ongoing operating expense is how budgets get blindsided.
Compute is mobile. Data is not, until you architect it to be, and every dollar you save on cheap GPUs can disappear the moment you have to move the data to reach them.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
