The data center buildout race

A GPU sitting in a warehouse does no work. To turn chips into useful AI compute, you need a data center: a building with dense power delivery, heavy cooling, fast networking, and round-the-clock operations. As demand for AI has surged, the shortage has shifted from just chips to the places that can actually run them.

That is why a global buildout is underway. Operators are racing to add capacity because the bottleneck is increasingly physical space with the right power and cooling, not only the hardware itself. A company can hold a stockpile of accelerators and still be unable to deploy them without somewhere to put them.

The race has a clear logic. Whoever brings working capacity online first can serve demand that competitors cannot reach yet, and demand for AI services has been arriving faster than buildings can be finished.

Building a modern data center is a multi-year project. It needs land in the right place, a grid connection that can deliver large and steady power, water or other cooling capacity, and skilled teams to operate it. Each of these can become the limiting factor, and they rarely line up neatly.

Power is often the hardest. Lawrence Berkeley National Laboratory projects U.S. data center electricity could reach 325 to 580 TWh by 2028, and connecting that much new load to the grid takes planning and time. A site can have approvals and hardware and still wait on power for a year or more.

These constraints compound. A facility might secure land quickly but wait on a substation upgrade, or finish construction but lack the skilled operators to bring it fully online. The slowest piece sets the pace for the whole project.

Because the right mix of power, land, cooling, and connectivity is rare, capacity tends to concentrate in specific regions. The IEA notes the United States accounted for about 45 percent of global data centre electricity in 2024, a sign of how clustered the buildout has become.

This concentration means access matters. Having a place inside a well-run, well-located facility is its own advantage, separate from owning the chips. As prime locations fill up, the value of an existing spot with secured power only rises.

Concentration also brings friction. When capacity clusters in a few regions, those areas face rising pressure on local grids, water, and community patience, which can slow new approvals. The result is a push to find new locations with spare power, sometimes far from existing hubs, and a premium on facilities that already have their power and permits in hand.

With so much pressure to bring capacity online, there is a temptation to build fast and worry about the details later. That rarely pays off. A facility rushed into service with undersized cooling or thin operations can throttle hardware, waste power, and suffer downtime that erases the advantage of arriving early.

The operators who do well tend to balance speed with durability. They secure power and cooling that match the hardware, design for the density AI demands, and staff for round-the-clock operations from the start. Building it right the first time avoids costly retrofits and keeps expensive hardware productive.

This is why the buildout is as much about engineering discipline as raw construction speed. The goal is not just a building, but a facility that can run dense AI hardware reliably for years, which is a harder and more valuable thing to deliver.