Inside the Data Centers That Train the World’s Biggest Models

A building designed for artificial intelligence training looks less like an office than like a substation with a very large computer attached, and the people designing it are electrical engineers far more often than software ones. The dominant constraints are electrical and thermal, and they shape everything about the design — floor load, ceiling height, the direction the building faces.

One rack, one house

A single rack of the latest accelerators can draw more power than a large suburban house. A training cluster containing tens of thousands of them draws as much as a mid-sized town. Delivering that power reliably requires dedicated substations, high-voltage transmission upgrades and, increasingly, on-site generation.

The consequence is that site selection is now driven by grid interconnection queues, not by real estate prices or fibre availability. Regions with surplus hydroelectric capacity, or with fast permitting for new gas plants, have become the industry’s preferred locations, and land in those places has been repriced accordingly.

Air cooling gives up at 20 kilowatts

• Air cooling works to roughly 20 kilowatts per rack and struggles beyond it.
• Direct-to-chip liquid cooling handles far higher densities and is becoming standard for new builds.
• Immersion cooling is efficient but complicates maintenance and hardware supply.
• Water consumption for evaporative cooling has become a local political issue in dry regions.

The optimisation target is the interconnect

Training a large model is a communication problem as much as a computation one. Thousands of chips must exchange gradients continuously, and the interconnect determines what fraction of the hardware’s theoretical throughput is usable. This is why high-speed fabric and optical interconnect have become as important a supply chain as the chips themselves.

A handful of balance sheets

The capital intensity of these facilities is such that only a handful of organisations can build them without partners. That concentration has consequences: it determines who can train the largest models, and it makes the availability of power a national policy question rather than a corporate one.

Two to three years, but not for the walls

A conventional office building takes two to three years. A large AI campus can be built faster in structure and much more slowly in everything that makes it functional, because the long-lead items are not walls but transformers, generators, switchgear and cooling equipment.

Equipment lead times of eighteen months to two years are common for high-capacity electrical gear, and the manufacturers are booked out for most of that window. This is why announcements about new capacity should be read as construction schedules rather than availability dates.

Capital is easy, engineers are not

Operating a facility at this density requires a different skill set from running a traditional data centre. Liquid cooling, high-voltage distribution and thermal management at rack scale are specialised disciplines, and the supply of experienced engineers is far smaller than the demand.

The consequence is that the binding constraint has shifted from capital to people. An organisation can raise money for a campus far more easily than it can hire the team to run one, and that asymmetry is shaping which projects actually complete.