I've worked around cloud infrastructure long enough to remember when geographic redundancy, colocation and disaster recovery were the questions everyone asked. They already demanded substantial power, cooling and connectivity. So why does an AI data centre sometimes look like a different industrial project altogether? The simplest answer is that certain AI workloads concentrate an extraordinary amount of computing demand in one place, and companies are trying to deploy it quickly.
A traditional enterprise facility may contain many types of servers doing relatively independent jobs: databases, websites, storage and business software. Large AI training systems often rely on dense groups of accelerators exchanging vast quantities of data through high-speed networks. That changes the electrical and thermal design. A rack that holds particularly demanding AI equipment can require far more power than a conventional rack, though there is no single figure that describes every deployment. An announced AI campus may also be designed to expand in phases, so its proposed final scale can be much larger than its present workload.
All the electricity used by computing equipment eventually becomes heat. The cooling system has to carry that heat away. This is one reason liquid cooling has become more important: it can bring cooling closer to very dense equipment. But liquid cooling doesn't automatically mean a data centre will consume vast amounts of fresh water. Some designs recirculate coolant and reject heat without substantial evaporative losses; others rely on water-intensive systems. The actual design and measurements matter more than the label.
The speed of development is another difference. If multiple large campuses seek connections in the same utility region, the challenge may be transmission capacity, new generation or the availability of transformers and switchgear rather than the buildings themselves. Local governments then face decisions about land, water, noise and financing of supporting infrastructure. Some projects can absorb their own costs; others need regulated agreements that deserve close examination.
This doesn't make AI data centres inherently better or worse than older cloud facilities. It makes comparisons harder if we only count buildings or announced megawatts. Our Observatory tracks whether a facility is proposed or operating, what electricity and cooling it actually uses, and how its costs and benefits are allocated. Those details allow us to compare the existing cloud industry with the AI buildout on evidence rather than assumption.
These explainers separate announced plans, recorded outcomes and unresolved questions. Find individual facility sources and submit corrections in the Observatory.
General background: IEA, Energy and AI. For site-specific claims, consult each facility's source records.