Every enterprise AI conversation starts with the same questions. Which models are you using? How many GPUs do you need? Which accelerators? These are important questions to ask, but it’s becoming clear that they’re not what separates companies getting real value from AI from those stuck running pilots in perpetuity.

The gap is data. Not how much data a company has, but whether an agent or model can actually put it to good use. A GPU cluster is only as productive as what reaches it, and for most companies, what reaches it is scattered across systems. Much of that data is unstructured and half-prepared at best.

There are several compounding factors that stop enterprises from getting the most from their data. The first factor is data fragmentation. Customer records in one system, machine telemetry in another, documents and images scattered across a dozen more, and none of it talking to each other. That part’s not new. It’s the same problem enterprises have had with structured data for twenty years, the stakes are now much higher with the combination of structured and unstructured data.

Less obvious is that most enterprise data isn’t AI-ready, even if you can find it. IDC says that 80 percent of enterprise data is unstructured and most of that is invisible to AI systems until someone searches for it and transforms it into something useful for AI models. Each data type needs its own kind of processing, which is part of why this process takes longer than planned.

There’s also an operational gap. Many organizations can get data prepared and still can’t run AI as a repeatable process across teams, so GPUs sit underutilized. Some suggest that a single storage system can solve this problem, but it’s not how the physics work. Extreme-performance flash is wasted on cold archival data. General-purpose file storage can’t keep up with thousands of GPUs retrieving data in parallel during training.

Enterprises should treat storage to play three distinct roles, not one. The first is broad, everyday AI workloads that need the simplicity of enterprise file storage. The second is extreme-scale training and large-scale inferencing clusters, which have to keep thousands of GPUs fed without interruption. The third is massive unstructured datasets, which call for a single searchable namespace instead of scattered buckets.

Quite simply, storage exists to deliver data at the rate GPUs consume it, not the other way around.

Getting to the data isn’t the same as being able to use it. Before any of it is useful to a model, it has to be queried, transformed, and indexed. That layer gets skipped in most infrastructure conversations because it’s less glamorous than GPUs.It comes down to four capabilities working together in practice. Customers need a way to query data where it already sits, a processing engine that can handle transformation at real volume, a search layer that makes enterprise data usable for retrieval and agentic workflows, and a security governance layer that only delivers data the users and agents are entitled to.

What’s new is where the GPU acceleration is landing. For years, it lived almost entirely in training and inferencing. Early performance gains are now showing up in the data processing layer itself, though it varies by workload. I’d treat any specific multiplier you see quoted with some skepticism until you’ve tested against your own data.

A few patterns that I’m starting to see in production are worth mentioning. Banks are using this architecture for real-time fraud detection, screening transactions against hundreds of risk factors in the time it takes to approve a purchase. Systems integrators building sovereign AI infrastructure for government and enterprise clients are processing data volumes in minutes versus days. And more than one large IT organization has eliminated a chunk of its ETL pipeline by querying data where it lives.

Unglamourous? Perhaps, but it’s the type of change that adds up over the course of a year.

None of this holds together without orchestration, and that’s the part that gets the least attention. Storage and data preparation can be fast but still add up to a slow pipeline if data isn’t moving through ingestion, preparation, retrieval, governance, and inference as one coordinated flow. This is where AI initiatives die, not from bad models, but from nobody owning the plumbing between steps.

The compute question, at this point, is mostly answered. Enterprises know how to buy GPUs. The companies that solve the data challenge will get real value out of AI versus spending another year in pilot purgatory.