There is a data center going up on the Columbia River Plateau where the nearest town has a single stoplight. The site was chosen not despite its remoteness but because of it: cheap land, cold air for free cooling, a hydroelectric grid that can deliver clean, stable power at scale, and fiber routes that planners quietly threaded through the region years before anyone imagined needing this much bandwidth. The building is not yet finished. The compute cluster it will house is already spoken for.
This is the new geography of intelligence. For decades, the default assumption was that data centers clustered near major metro areas — close to undersea cable landings, corporate campuses, and the engineers who maintain the hardware. That logic has not disappeared, but it is being overwhelmed by a different imperative: the sheer energy appetite of frontier AI training and inference is so large that it can no longer be satisfied inside the places where most people live and work. The power is out there, in the places that have rivers, wind corridors, geothermal resources, and room to build.
The numbers reframe the imagination. A single large-scale AI training cluster today can draw 100 to 200 megawatts continuously. The gigawatt-class campuses now under active planning — multiple sites across the American interior, the Scottish Highlands, Scandinavia, and the Atacama plateau — represent a scale of dedicated industrial infrastructure that rivals aluminum smelting operations. Hyperscalers and their AI infrastructure partners are signing power purchase agreements for output that won’t physically exist for three or four years, betting that the demand will still be there and will almost certainly be larger.
What makes this moment genuinely different from the last round of data center expansion is the density. Earlier generations of web-scale compute could be distributed across many modest facilities. Training a frontier model, and increasingly running inference for the systems that come out of it, wants to be in one place, or a small number of tightly coupled places, because the interconnect bandwidth between accelerators is the bottleneck. You cannot split a 100,000-GPU training run across sites that are hundreds of kilometers apart and expect to keep the interconnect fabric anywhere close to saturation. So the buildings get bigger, and they need more power in one spot than a small city might draw.
This concentration is producing some spectacular feats of infrastructure. Microsoft, Google, and several newer entrants are investing in direct agreements with nuclear operators — existing plants in some cases, advanced reactor designs in others — specifically to provide the kind of always-on, carbon-minimal baseload that renewables alone cannot guarantee at this density. The xAI Colossus facility in Memphis drew attention partly because of its scale and partly because of how fast it was assembled: a massive liquid-cooled GPU cluster built out in months, not the years that major industrial projects typically require. That pace is becoming the expectation rather than the exception.
The thermal engineering involved is quietly extraordinary. Liquid cooling, once a specialty approach, is now standard at the frontier. Immersion cooling — submerging entire server racks in dielectric fluid — is moving from exotic to mainstream because air cooling simply cannot keep up with the heat density of the newest accelerators. Some facilities are experimenting with closed-loop systems that capture waste heat and redirect it to district heating networks, turning what would be an environmental liability into a resource.
The infrastructure being laid down right now is not just serving the models of today. It is the physical substrate for whatever comes next — the scaffolding inside which the systems of the late 2020s will take shape. That remote stoplight town on the Columbia Plateau is not really building a data center. It is building the foundation of something we don’t have a name for yet.