Table of Contents
Research reveals the environmental cost of AI
EVER wondered what happens environmentally every time you ask an artificial intelligence (AI) chatbot a question or generate an AI image?
Typing a prompt into a chatbot feels almost weightless. The answer appears in seconds, as if it materialised out of thin air. But behind that instant response sits a vast network of servers, data centres and specialised computer chips working constantly to generate the reply.
Each prompt may feel small but at scale, those requests add up.
Not just ‘the cloud’
AI systems do not run in the cloud in the literal sense. They run inside massive data centres filled with rows of servers and processors designed to handle heavy computational workloads.
When you ask a chatbot a question, those machines perform calculations to interpret the prompt and generate a response. This process known as “inference” requires electricity every single time the system is used.
Researchers estimate a single AI chatbot query can use about five times more electricity than a typical web search.
That difference might seem minor, but millions of people are now using AI tools daily. Multiply that extra energy use across billions of prompts and the demand becomes significant.
Data centres already consume huge amounts of electricity worldwide. Their energy use reached roughly 460 terawatt hours in 2022 and is expected to more than double within a few years as AI services expand.
Cost of training AI
Before AI can answer questions, it must first be trained.
Training a large AI model involves feeding enormous datasets through thousands of specialised processors over days or weeks. The computing power required is immense.
One estimate suggests training a large language model consumed around 1,287 megawatt hours of electricity, easily producing about 552 tonnes of carbon emissions.
And because AI models are frequently replaced with newer versions, the training process happens repeatedly across the industry.
A 2025 MIT analysis examining generative AI’s environmental impact found AI training clusters can require seven to eight times more energy than typical computing workloads.
Water behind every answer
Electricity is only part of the environmental footprint.
The servers generating AI responses produce large amounts of heat and keeping them cool requires significant water resources.
Many data centres rely on chilled water cooling systems that circulate water through facilities to absorb heat from the hardware.
Researchers estimate roughly two litres of water may be required for every kilowatt hour of electricity used in a data centre.
That means large AI computing clusters can consume substantial amounts of water, particularly in regions already facing water stress.
A 2025 study published in Nature Sustainability projected expanding AI infrastructure could generate a water footprint of hundreds of millions of cubic metres annually while also producing tens of millions of tonnes of carbon emissions.
Hidden hardware footprint
The environmental cost of AI also extends to the machines themselves.
Running modern AI systems requires powerful chips known as GPUs, which are manufactured through complex semiconductor fabrication processes. Producing these chips consumes energy and involves chemicals and raw materials that must be mined and processed.
Demand for these processors has surged as companies race to build more AI computing capacity. Millions of GPUs are now shipped to data centres every year to support AI workloads.
Transporting and instaling this hardware adds another layer of emissions.
AI’s sustainability challenge
None of this means AI is inherently harmful. In fact, AI is already helping scientists model climate change, improve renewable energy systems and accelerate research.
The challenge lies in balancing those benefits with the growing environmental footprint of the technology itself.
More efficient chips, smarter software and renewable-powered data centres could significantly reduce the impact.
But as AI becomes embedded in search engines, apps and everyday digital tools, one reality is becoming clear.
That quick AI response on your screen may only take seconds to generate, but the infrastructure behind it is running constantly, consuming energy and resources long after the prompt is answered.
READ MORE:
Why AI agrees too much, according to research
Vox machina: When AI takes over your voice
