Products, laws and numbers here change fast. Checked October 2026.
In plain words
When you ask a chatbot a question, your phone does not do the thinking. Your words travel over the internet to a data center: a warehouse-sized building lined with rows of tall metal cabinets, called racks, each packed with computers.
AI data centers are built around AI chips such as GPUs. Thousands of them are wired together with very fast cables so they can work as one giant machine. That is what it takes to train a large language model and to answer millions of people at once.
All those chips use a lot of electricity and turn it into heat. So a data center is also a power station customer and a giant cooling system: fans, chilled water and sometimes liquid piped right onto the chips. A typical AI-focused data center uses as much electricity as about 100,000 homes.
Everyday example
A city's central kitchen
Some school meal programmes cook for lakhs of children in one central kitchen. Giant stoves, steam cookers and conveyor belts run all day, using huge amounts of gas, water and power, and trucks carry meals out to every school. A data center is the central kitchen for the internet: the heavy work happens in one place, and the results travel to your phone.
Where the example breaks: A kitchen's food physically travels to you. With a data center only information travels, in a fraction of a second, so the building can be in another state or another country.
How it actually works
An AI data center is designed around three limits: how much power it can get, how much heat it can remove, and how fast its chips can talk to each other.
Chips, servers, racks, halls
AI chips sit in servers, servers stack into racks, and racks stand in long rows in large halls. A single top-end AI chip can draw around a kilowatt, so one rack can use as much power as many homes.
A network that makes many chips act as one
To train one model across thousands of chips, every chip must swap results with the others constantly. Data centers use special high-speed links and switches so this traffic does not slow the work. Google's TPU pods, for example, link 9,216 chips directly.
Cooling
Nearly every watt a chip uses becomes heat. Older halls blow cold air; newer AI halls often pump liquid through plates on the chips, because air cannot carry heat away fast enough from such dense racks.
Power and siting
Builders look for places with spare grid capacity, steady and affordable electricity, a way to cool (cold air or water), land, and fast fibre links to users.
- All data centers, 2024
- About 415 TWh of electricity, around 1.5% of the world's use (IEA, April 2025)
- Forecast for 2030
- Around 945 TWh, more than double, roughly what Japan uses in a year (IEA)
- One AI data center
- A typical AI-focused one uses as much electricity as 100,000 households; the largest under construction, 20 times more (IEA)
- Largest known, early 2026
- About 1.1 million H100-equivalent GPUs of compute at one site (Epoch AI estimate, February 2026)
Big cloud companies that run vast fleets of data centers are often called hyperscalers.
Questions people ask
What is an AI data center?
How much electricity do AI data centers use?
A lot, and it is rising. The International Energy Agency estimates all data centers used about 1.5% of the world's electricity in 2024, and that this could more than double by 2030, with AI the main driver. A typical AI-focused site uses as much as about 100,000 homes.
Why do data centers need so much water?
Many cool their chips with evaporative cooling: water absorbs heat and some of it evaporates, the way sweat cools skin. How much depends on climate: in cool places outside air can do the job most of the year. Newer designs pipe water in a closed loop across the chips and reuse it.
Is ChatGPT running in a data center?
Yes. When you type into ChatGPT or any similar app, your message goes to data centers where the model runs on AI chips. Only the reply comes back to your device. That is why these apps need an internet connection, unlike on-device AI.
What is a hyperscaler?
A company that runs cloud computing at enormous scale across many huge data centers, such as Amazon Web Services, Microsoft Azure and Google Cloud. They rent computing power, including AI chips, to other businesses by the hour.