Every AI project we discuss with IT leaders has a GPU budget, a model budget and, increasingly, an energy budget. Almost none has a line for the data itself. The market numbers show the same imbalance: Gartner forecasts worldwide AI spending of $2.7 trillion in 2026, with nearly $1.5 trillion going to AI infrastructure and about $3.1 billion to the category it calls AI data.
That is the gap: the unstructured data AI projects depend on (files, images, instrument output, documents, logs) sits scattered across storage systems that were never designed to feed a model.
Moving, finding and governing that data is real work, and when nobody budgets for it, it quietly becomes the slowest part of the whole pipeline.
We at NODEUM spend our days designing and engineering enterprise software for moving and archiving unstructured data for research centres, HPC sites and AI Factories and teams.
Here is what we see, and what we think CIOs should plan for before the next AI budget is signed.
The money goes to compute, the delays come from data
AI investment conversations start with accelerators, and for good reason: they are expensive, scarce and easy to count. But a GPU cluster only produces value when the right data is on fast storage at the moment a job starts.
In practice, the data needed for training or fine-tuning is spread across a parallel file system, a few NAS shares, object storage, tape and one or more clouds. Someone has to find it, copy it, check it and clean up afterwards.
That work rarely appears in a project plan. It is absorbed by storage administrators, data engineers and researchers who write one more script, open one more ticket and wait one more day. In Iron Mountain's "Information Supercycle" research (October 2026, conducted by Vanson Bourne), four in ten organisations said they spend at least 41% of their AI project timelines on manual data preparation. The cost is real; it is simply spread across people and weeks instead of sitting on an invoice.
Why unstructured data AI pipelines break at scale
Structured data has had decades of tooling: databases, warehouses, ETL, catalogues. Unstructured data has mostly had file systems and good intentions. Gartner estimates that unstructured data makes up 80–90% of all new enterprise data. Three things make it hard to feed into AI pipelines.
It is scattered. The same project's data can live on four kinds of storage, owned by different teams, with different access rules. Nobody has a single view of what exists and where.
It is big in the wrong way. The challenge is not only petabytes but millions or billions of files. Tools that work well for a few thousand files slow to a crawl when they have to walk a tree of that size before copying anything.
It has no memory. When a dataset is copied to fast storage for a training run, there is usually no record of which version was used, where it came from or where it went afterwards. That makes results hard to reproduce and audits hard to answer.
None of this is solved by buying more GPUs. Faster accelerators actually make it worse: the more compute you add, the more often it sits idle waiting for data.
The hidden line items
When we map a typical AI data workflow with a customer, the same unbudgeted costs come up again and again:
Idle accelerators. Jobs queued behind data staging, or launched on incomplete datasets and restarted.- Duplicate copies. The same dataset copied to several places "just in case", each copy on expensive storage, none of them tracked.
- Manual movement. Engineers running copy jobs by hand, checking them by eye and cleaning up when they remember. (See rsync, MPI or a data mover?)
- Search time. Researchers and data scientists looking for data the organisation already owns, or recreating it because they cannot find it.
- Compliance exposure. Sensitive data moved without a record of who moved it, where and when.
Each item looks small on its own. Together, they decide how fast an AI investment pays back.
What to budget for instead
Our view is simple: treat data movement and data management as part of the AI infrastructure, not as an afterthought to it. In budget terms, that means planning for four capabilities from day one.
- Visibility. A catalogue of the unstructured data you already own, across every storage system, searchable by metadata and not only by path.
- Policy-driven movement. Rules that stage datasets onto fast storage before a job and move them off afterwards, without someone running a script.
- Verification and audit. Every transfer checked and recorded, so you can prove what data a model was trained on.
- Tiering and preservation. Cold data moved to lower-cost object storage, and recalled when needed, so the fastest tier holds only what is in use.
These are not exotic requirements. They are what any mature storage team already does for backup and archive. The difference is that AI pipelines need them to run continuously, at the speed of the compute, across every storage system the organisation uses.
Questions to ask before the next AI budget is approved
If you are a CIO or IT director reviewing an AI infrastructure plan, a few questions will tell you quickly whether the data side has been thought through:
- Where does the training data live today, and how many storage systems does it span?
- Who moves it onto the GPU tier, how long does that take, and is it automated?
- Can we say which version of which dataset each model was trained on?
- What happens to staged data after a job ends? Who deletes or archives it?
- What share of our fast storage holds data nobody has touched in months?
If the answers are vague, the bottleneck is already there. It just has not been named yet.
Make data a first-class part of the AI factory
The term "AI factory" is useful because it reminds us that AI is production, not a one-off experiment. Factories plan for their supply chain. In AI, the supply chain is data, and most of it is unstructured. The organisations that get the most from their accelerators will be the ones that budget for moving, finding and governing that data with the same seriousness as the compute itself.
We will share more about how we approach this over the coming weeks. In the meantime, if your AI plans are running into data movement or data management questions, read our Data Mover for the AI factory white paper or talk to a NODEUM expert about your environment.
