The cloud is a place
We talk about the cloud as if it were made of air. In truth, every message you send, every video you stream, and every AI answer you receive travels through a very physical world of buildings, cables, and machines. When people say a service runs in the cloud, they mean it runs in a data centre — a large, carefully controlled building filled with computers.
These buildings are usually plain and windowless, tucked away near power lines and fibre routes rather than city centres. Yet they are the beating heart of the modern internet, and the recent surge in artificial intelligence has made them more important, and more power-hungry, than ever before.
What is actually inside
Step inside a data centre and you find long aisles of tall racks. Each rack holds dozens of servers — stripped-down computers with no screens or keyboards, designed only to compute and store data. They are stacked close together to save space and connected by a dense web of cables. Every facility is really four systems working together:
- Power — high-voltage feeds, backup generators, and battery banks that keep everything running without a flicker.
- Cooling — chillers, fans, and increasingly liquid-cooling loops that carry heat away from the chips.
- Compute — tens of thousands of servers, many packed with GPUs and other accelerators that do the AI work.
- Networking — the switches, routers, and fibre links that move data in and out at enormous speed.
A single large facility can hold tens of thousands of servers. Together they consume as much electricity as a small town, which is why location, power supply, and cooling are decided long before a single machine is switched on.
Why AI changed the game
Traditional web services — email, websites, online shopping — are relatively light on computing power. Artificial intelligence is not. Training a large model involves running trillions of calculations across thousands of specialised chips for weeks at a time. Even after training, answering questions at scale demands a steady stream of heavy computation.
To meet this demand, companies have shifted from ordinary processors to graphics-style chips, known as GPUs and other accelerators, that can perform many calculations in parallel. Racks packed with these chips generate far more heat and draw far more power than the servers of a decade ago, reshaping how data centres are designed. Training large language models is one of the most demanding jobs these buildings now do.
To put numbers on it: a single large AI data centre can draw hundreds of megawatts of electricity — roughly 100MW to 500MW for one site, with the largest planned facilities approaching a full gigawatt. That is the appetite of a city, not a building. The compute is concentrated too: xAI's Colossus cluster in Memphis, Tennessee packs 100,000 NVIDIA Hopper GPUs into a single training system, one of the largest AI training clusters ever built.
Every clever AI answer rests on an unglamorous foundation of concrete, copper, silicon, and electricity.
The heat problem
Computers turn electricity into work and heat, and packed together they produce a lot of heat. Left unmanaged, that heat would quickly destroy the equipment. Cooling is therefore one of the largest costs and engineering challenges in the whole industry.
For years, air conditioning was enough. As chips grew hotter, operators turned to more advanced methods: directing cold air precisely where it is needed, piping cool liquid directly to the hottest components, and even submerging servers in special non-conductive fluid. Some facilities are built in cold climates or beside rivers to use the natural environment as a giant heat sink.
Measuring efficiency: PUE
The industry scores efficiency with a ratio called PUE — Power Usage Effectiveness: total power drawn by the building divided by the power that actually reaches the computing equipment. A perfect score would be 1.0, meaning no overhead at all. The global average, tracked by the Uptime Institute, sat at 1.54 in 2025 — meaning that for every watt spent computing, more than half a watt again goes to cooling and infrastructure overhead. Every tenth of a point matters when a site burns through hundreds of megawatts, which is why operators obsess over airflow, liquid cooling, and cool climates.
Power, water, and responsibility
The scale of modern data centres raises real questions about resources. They consume enormous amounts of electricity, and some cooling designs use significant water. As more of them are built to support AI, communities and regulators are asking sensible questions about where that power comes from and how the environmental cost is managed.
Many operators have responded by buying renewable energy, improving efficiency, and reusing waste heat to warm nearby buildings. Progress is real but uneven, and the honest picture is that demand is rising faster than efficiency alone can offset. How the industry balances growth with responsibility will be one of the defining infrastructure stories of the decade.
The geography of all this is changing too. Because data centres need cheap, reliable power and cool conditions, operators increasingly build them in specific regions — near hydroelectric dams, in cold northern climates, or beside major fibre routes. Northern Virginia, nicknamed "Data Center Alley" and centred on Ashburn in Loudoun County, is the world's largest data centre market, while newer AI campuses such as xAI's Colossus have landed as far afield as Memphis, Tennessee. This has turned some quiet rural areas into unexpected hubs of the digital economy, bringing investment and jobs but also straining local power grids and water supplies. Governments now weigh these projects carefully, balancing the economic benefits against the demand they place on shared resources. It is a striking sign of how central these facilities have become that their location is now debated in council meetings and national energy plans alike, rather than settled quietly by a single company.
Why this matters to you
You will probably never visit a data centre, yet your digital life depends on them completely. Understanding that the cloud is a physical place helps make sense of the news — why chip shortages slow down AI, why energy policy and technology are now linked, and why a company's choice of where to build affects real communities.
The next time an app answers instantly or a model writes a paragraph in seconds, it is worth remembering the hidden machinery behind it: rows of humming servers in an anonymous building, kept cool and powered around the clock, quietly turning electricity into the intelligence we increasingly take for granted. One response is to move some processing out to smaller sites near users, an approach known as edge computing.
