Every chatbot answer, AI image and video recommendation runs on servers in a data centre somewhere. As AI use has exploded, so has demand for electricity and cooling. This is one of the biggest engineering stories of the decade, affecting power grids, water supplies and electricity prices from Virginia to Mumbai.
What is inside an AI data centre
A data centre is a large, secure building full of server racks. AI workloads run on specialised chips, mostly graphics processing units, packed densely together. A single modern AI rack can draw as much power as dozens of homes. Around that core sit transformers, backup generators, battery systems and large cooling plants.
How much electricity?
The International Energy Agency estimated that data centres used roughly 1.5% of the world's electricity in 2024, and projected that this could more than double by 2030, driven largely by AI. In some regions with many data centres, they already take a much larger share of local supply, which is why utilities are building new power plants and transmission lines.
Training a large AI model takes a huge burst of energy over weeks. But because hundreds of millions of people use AI every day, the energy spent answering questions, called inference, now adds up to a large share too.
Why data centres use water
Almost all the electricity a chip uses ends up as heat. Many data centres remove that heat using evaporative cooling towers, which work like sweating: water evaporates and carries heat away. That water is lost to the air. Data centres also use water indirectly through the power stations that supply them.
Newer designs reduce water use with liquid cooling that runs coolant directly to the chips in a closed loop, or with air cooling in cooler climates. The trade-off is often slightly higher electricity use.
How engineers are tackling the problem
- More efficient chips that do more calculations per watt.
- Liquid and immersion cooling, where servers sit in special non-conductive fluid.
- Clean power deals with solar, wind and, increasingly, nuclear plants, including plans for small modular reactors.
- Heat reuse, piping waste heat into district heating networks in some European cities.
- Smaller models that run on phones and laptops instead of in the cloud.
What it means for you
For ordinary users, a single chatbot query uses a small amount of energy, often compared to running an LED bulb for a few minutes. The concern is scale. Billions of queries, plus model training, add up to the consumption of entire countries. Engineering efficiency, not giving up AI, is where most of the solution lies.
Frequently asked questions
Is India building AI data centres?
Yes. Mumbai, Chennai and Hyderabad are major hubs, with large new campuses announced by Indian and global companies. Access to reliable power and water is a key factor in where they are built.
What is PUE?
Power usage effectiveness compares a data centre's total power with the power used by its computers. A PUE of 1.2 means 20% extra goes to cooling and other overheads. Lower is better.