AI Is Moving Infrastructure Closer to the Data. That's the Right Call.

For twenty years, the model was simple: centralize compute in big regional data centers, connect everyone with faster networks, done. It worked. It’s still not wrong.

But AI changes the math.

Training a large model needs massive GPU clusters, no argument there. Running that model in production, what actually happens every time it does its job, is a different problem entirely. That’s inference: the moment AI reads a sensor, answers a claim, flags a defect, approves a transaction. Inference needs to move fast, stay secure, and sit close to wherever the data actually gets created. That’s not a training problem. That’s a distance problem, and distance gets solved with three things at once: a place to put compute, a network fast enough to get data there and back, and cloud flexible enough to run whatever sits on top.

So the question I hear from customers is shifting. It used to be “where should we host this?” Now it’s “where does this actually need to run, and can we get data there fast enough to matter?” Different question, different answer.

The Edge Just Got Bigger

“Edge” used to mean one thing: get the app closer to the user, cut the latency. AI stretches that definition, and it puts real weight on the network underneath it.

The next wave of AI workloads isn’t staying inside hyperscale regions. It’s moving onto manufacturing floors, into hospital systems, alongside trading platforms, through logistics networks, on enterprise campuses. Wherever a decision needs to happen in milliseconds, that’s where the compute has to live, and that compute is worthless if the fiber connecting it can’t keep up with the same millisecond math.

I don’t think the answer is one enormous AI data center anywhere. I think it’s a network of regional infrastructure, tied together by fiber we control end to end, that shows up wherever the workload demands it. That’s the bet we’re making at US Signal, and it’s not a hedge. It’s the whole thesis.

Why This Actually Matters

Not every AI workload wants the same thing. Some need a massive centralized GPU cluster for training. Some need inference at the edge, answering in milliseconds, no round trip to a distant region. Some customers have compliance and data residency requirements that make “just send it to the cloud” a non-starter.

We’re watching this play out with real customers right now. Insurance companies want inference on claims data without shipping it cross-country. Manufacturers want quality control decisions made on the floor, not three states away. Neither of those works without three things lined up: compute sitting close enough to matter, fiber fast and resilient enough to move the data without adding its own latency, and cloud services that can orchestrate the model without forcing the data to travel to get there. That’s not a future scenario. That’s this year.

Over time, AI workloads spread out. They don’t consolidate into one place. Infrastructure, and the network connecting it, has to be built for that reality now, not retrofitted for it later.

Build Ahead of the Curve, Not Behind It

Infrastructure has always had to move first. The companies that built virtualization platforms before virtualization was the obvious move were ready when the market caught up. Same story with cloud. Same story with the fiber networks that made cloud actually usable at scale. Most people don’t think about that part, but none of it works without the network underneath it.

AI is that same inflection point, running again. A lot of providers are chasing today’s GPU shortage. Fair enough, that’s real demand. But we’re also building for what’s next: a distributed AI fabric, fiber, regional data centers, and cloud working as one system, that doesn’t force every workload back to one massive facility.

I’ll go deeper on this in the next two pieces: what this infrastructure actually looks like, why AI density breaks the old rules on cooling and power, and how to start planning your next move before you’re forced into one.

The future of AI infrastructure isn’t bigger data centers. It’s smarter distribution, connected by a network built to move inference at the speed it actually happens. We’d rather be early on that call than reacting to it.