Teams compare AI cloud platforms on GPU availability and hourly rates, then discover the real constraint is somewhere else: their data. Datasets are heavy, they accumulate, and they are often subject to rules about where they may be stored and processed. Two questions, where must the data live and what does it cost to move it, can rule out an otherwise attractive option before you write a line of code.

Data Gravity Is Not a Metaphor

Compute is portable in a way data isn’t. You can point a job at a different cluster in an afternoon, but a multi-terabyte dataset has to physically travel, takes time to copy, and then exists in two places that must be kept consistent. Training pipelines that stream data every epoch make this worse, because latency and throughput between storage and GPUs directly shape how fast the job runs.

The practical rule: put compute where the data is, unless there is a strong reason to do otherwise. If the platform you like is in a different place from your data, the move itself becomes part of the project cost.

Residency: Know the Rule Before You Pick the Region

Data residency means requirements about the geographic location where data is stored or processed. These can come from law, from contracts with customers, or from internal policy. They differ by country, sector and data type, and they change, so this article can only give you the shape of the question. Have your legal or compliance contact confirm what applies to you.

Questions to put to them early:

• Which datasets are restricted to a country or region, and which are not?
• Does the restriction cover storage only, or also processing, backups and logs?
• Does it apply to derived artifacts, such as embeddings or trained weights that might encode information from restricted data?
• Are there rules about who may administer the infrastructure, not just where it sits?

That third question is the one teams forget. Whether a model trained on regulated data inherits the restriction is a judgment for qualified advisors, but you need to ask it before the training run, not after.

Egress: The Cost That Hides in the Architecture

Egress is data leaving a provider’s network or crossing certain boundaries. Many providers charge for some kinds of transfer and not others, and the rules differ between providers and between types of traffic. Check each vendor’s current pricing documentation instead of assuming; do not rely on a rule of thumb from a blog post, including this one.

What matters is identifying the transfer paths your design creates:

• Ingest: moving your existing dataset in.
• Cross-region reads: compute in one region reading storage in another.
• Checkpoints and artifacts: model files copied out to other systems.
• Serving: responses returned to users or to services hosted elsewhere.
• Exit: moving everything out if you leave.

The exit path is the easy one to ignore and the one that most affects your negotiating position.

A Hypothetical Example

Say a team stores several terabytes of labeled images in object storage with their current provider and wants to train on a different platform that offers better GPU availability. Their first plan is to leave the data where it is and have the new platform read it over the network each epoch.

On paper that avoids a migration. In practice, every epoch re-reads the full dataset across a provider boundary. Training slows because the GPUs wait on the network, and each pass generates transfer charges. After a few runs the team realizes that a one-time copy of the data to the new platform would have been cheaper and faster, even counting the effort of keeping the two copies in sync. They also find that some images contain personal data restricted to a specific region, so those can’t move at all and need a separate path.

The lesson is not that copying is always right. It’s that this had to be worked out on a whiteboard before the first run rather than from the bill.

A Pre-Migration Checklist

Question	Why it matters	Who to ask
Which datasets are region-restricted?	Determines which regions are even eligible	Legal / compliance
How large is each dataset and how often does it change?	Sets one-time copy cost and sync effort	Data engineering
Does training stream data or read it once?	Streaming across regions hurts speed and cost	ML engineers
Where do users and downstream services sit?	Drives serving egress and latency	Platform / product
What are the provider’s transfer terms?	Egress rules differ by vendor and path	Vendor documentation
How would we export everything on exit?	Reveals lock-in before you commit	Infrastructure

Design Patterns That Reduce the Pain

A few approaches come up repeatedly:

1. Stage once, train many. Copy the dataset to the training region a single time and treat that copy as the working set.
2. Split by sensitivity. Keep restricted data in its required region and run those jobs there, while using wherever the best capacity is for non-restricted work.
3. Ship models, not data. Trained weights are usually far smaller than the data behind them. Where policy allows, moving the artifact is cheaper than moving the corpus.
4. Cache close to compute. If you must read from a remote store, put a local cache in front so repeat epochs don’t cross the boundary again.

Frequently Asked Questions

Is moving data between regions always expensive? Not always, since terms vary by provider and by traffic type. Check current documentation, and estimate using your actual dataset size and read pattern.

Can we just keep data on-premises and burst to the cloud? Sometimes, but streaming over a wide-area link usually bottlenecks training. It works better when you copy a working set up front.

Does residency apply to model weights? It can, depending on the data used and the rules involved. Get a qualified legal opinion before training on restricted data.

Should we encrypt data we move? Yes, in transit and at rest, as a baseline. Encryption does not by itself satisfy a residency requirement, though.

How do we find out what exit would cost? Add up your stored volume and look up the provider’s current transfer terms for moving data out. Do this before signing, not at renewal.

Next Step

List your three largest datasets, mark which are region-restricted, and write down where each one is read from during training. That one-page map will settle more of the platform decision than any benchmark.