AI & Compute
Where AI Data Centers Get Built And Why
Siting a large computing campus is decided by available electrical capacity, water, land and local permitting, which is why clusters form in a few American regions rather than near users.
Large training and inference campuses are not distributed evenly across the United States. They cluster, and the clustering follows the grid rather than the population.
Power availability is the first filter
A large campus needs an interconnection to high-voltage transmission and a utility willing to commit capacity. Both are scarce, and both take years to arrange.
Developers therefore search for locations where substation capacity already exists or where a retired industrial load has freed it up. Finding spare electrons is harder than finding land.
This is why former aluminum smelters, paper mills and coal plant sites reappear as computing sites: the wires are already there.
Latency matters for some workloads and not others
Serving an interactive application benefits from being physically close to users, because signals take time to travel and each round trip adds delay.
Training does not care. A job that runs for weeks inside one building can sit anywhere the power is, which decouples the largest facilities from the coasts.
The result is a split: smaller inference sites near metropolitan areas, and very large training campuses in places chosen for energy.
Cooling ties the site to water and climate
Heat has to leave the building. Evaporative systems use water and are efficient; closed-loop and air-cooled systems use less water and more electricity.
Dry regions with cheap power therefore force a trade, and communities increasingly negotiate over which side of it a project takes.
Cooler climates reduce the energy penalty of avoiding water use altogether, which has made northern sites attractive despite distance from customers.
Local government decides the timeline
Zoning, noise limits, backup generator permits and tax arrangements are decided county by county. A jurisdiction that has done this before can move quickly.
Places that have not may take longer or refuse, and opposition now commonly focuses on electricity bills and water rather than on the buildings themselves.
Because approval risk is as real as construction risk, developers concentrate where precedent exists, which reinforces the clustering.
Why the pattern is self-reinforcing
Once a region hosts several campuses, it accumulates electrical contractors, transmission upgrades, trained technicians and a permitting process that understands the use.
Each of those lowers the cost of the next project there. The geography of computing ends up looking like the geography of heavy industry, for the same underlying reasons.
