Why Regional AI Data Centers Matter for Inference

Why Regional AI Data Centers Matter for Inference

By Brad Alexander, Chief Technology Officer, DartPoints

The Next Chapter of AI Runs Closer to Home

Walk any conference floor this year, and you'll hear one word over and over: gigawatts. Gigawatt campuses, gigawatt power deals, gigawatt pipelines. I understand why. Those sites train the foundation models everyone else builds on, and nobody in our industry should downplay that work.

I keep coming back to a different question, though. What happens to the 10 MW and 20 MW enterprise data centers that carried corporate IT for the last two decades? And as the industry's focus moves from training to inference, where does all that AI actually run?

My answer, and the one we've built DartPoints around, is that a large share of it will run much closer to home than the headlines suggest.

Start with where the workloads already are

Uptime's 2025 survey put roughly 45% of IT workloads in corporate facilities. Separate research from Foundry and CoreSite found that 98% of IT leaders surveyed had adopted or planned to adopt hybrid IT. Those companies have contracts, compliance rules and years of systems in place. They aren't going to rip all of it out because AI showed up.

So AI has to work with what they already have. I think of enterprise AI as a placement problem before anything else. Where does the data sit? How fast does the application need to respond? Does the workload have to stay under the company's control, or inside a particular jurisdiction? How soon can capacity actually come online? Answer those four questions honestly, and the right location usually becomes obvious.

I'm from South Carolina, where manufacturing is a big part of life. Think of a German car manufacturer or a tire manufacturer with a major operation here. When AI shows up on a production line like theirs, it may need an answer in a millisecond or two. A round trip to a hyperscaler in Ashburn or Chicago won't cut it. Some of that processing may need to stay right in the plant. Other workloads may fit a nearby data center. Hospitals working with patient scans and banks checking transactions for fraud have their own requirements, but they face the same placement question.

That's part of the gap the enterprise AI node fills. It's a regional facility built for private inference and everyday operational workloads, sitting in the markets where companies actually operate. We use the model ourselves, too. Across our 12 data centers in 11 markets, AI agents help us run and tune our own facilities.

Density changes more than the rack

Cabinets that used to draw 10 to 25 kW may now need to handle 60 to 120 kW. We see it in our own buildings and across the market. Most of the industry conversation stays on the rack itself, which makes sense. That's where the new cooling hardware and the eye-popping power figures live.

What worries me more is everything feeding that rack: utility service, substations, electrical distribution and the thermal path out of the building. Then there are harmonics. Total harmonic distortion rarely comes up at conferences, yet it can become a problem as equipment and electrical loads change, and plenty of older facilities were never designed for those conditions. Plan an upgrade around the cabinet alone, and you'll find those problems later at a much higher price.

Why we redevelop existing sites

We don't build gigawatt campuses. We find available power near dense business markets, take an existing building or operating campus, bring in connectivity and convert it.

The attraction of an existing data center is that you're not starting from an empty field. There's infrastructure to work with. We still have to sort out the power, permits and what the building can actually support. From there, we can plan zones for AI and others for traditional enterprise workloads, giving customers both under one roof. Once a zone works, we repeat it in the next one.

Speed is why this matters. Everybody in this business knows how hard power is to get right now. A ground-up project can take three to six years. In some cases, we can bring a brownfield site to market in under 12 months, with converted space supporting 120 to 150 kW per cabinet and liquid cooling at the floor.

For a customer, the useful question is simple: Can you support my workload, and when can I move in?

Earn the community's trust early

Data center operators have to earn their welcome now. People have real questions about power, water, noise and what a project will mean for their town. We owe them straight answers. We're building 30 MW and 50 MW sites to serve local demand, a very different animal from a gigawatt campus, but the size of the project doesn't make those questions go away.

So community engagement goes into our business plan on day one. That covers regulatory awareness and giving something back through education programs, innovation and research centers, and outreach to middle and high school students. Getting the conversation started early matters. So does following through when people raise concerns. Utilities have come around as well. Because our power partnerships can return capacity to the grid, many of them now want us in their territory.

Where this goes

The megacampuses will keep setting the pace on scale, and the industry needs them. The businesses, hospitals, factories and towns that use AI every day need capacity that fits the work they're doing. For some, that means keeping it nearby, and enterprise nodes can meet that need. I expect we'll see far more of them as inference grows.

As I see it, our opportunity is to take infrastructure enterprises already know and trust and turn it into the capacity they'll need for what comes next.

Your next AI deployment could be closer than you think. Talk with DartPoints about capacity in the markets where your business operates.