The hidden cost of
every AI request.
AI feels weightless. It lives in a browser tab, responds in seconds, and leaves no obvious trace. But behind every query is a data centre drawing power, consuming water and generating heat — at a scale most users never see.
Data centres currently consume between 1–2% of the world's total electricity — a figure expected to rise sharply as AI workloads grow.
Training and running large AI models requires significant water for cooling. Some estimates suggest large models consume hundreds of millilitres of water per conversation.
A single AI-generated response can use four times the energy of a standard web search, depending on model size and query complexity.
AI data centre electricity demand is forecast to rival the entire electricity consumption of some mid-sized countries by the mid-2020s.
What a data centre actually is
A data centre is a facility housing thousands — sometimes hundreds of thousands — of servers running continuously, 24 hours a day. These buildings range from the size of a warehouse to the size of several city blocks. They require uninterrupted power, redundant cooling systems, physical security, and constant maintenance.
The servers inside process requests from users around the world in milliseconds — but doing so at global scale means the aggregate energy, water and hardware demands are immense. Major technology companies operate dozens of these facilities across multiple continents, with more being built every year.
AI workloads are particularly demanding. Where a web server might retrieve a stored page, an AI model generates a response from scratch each time — a process requiring significantly more computation, and therefore significantly more energy per request.
Where the impact comes from
Environmental cost accumulates across the full lifecycle of AI infrastructure — not just during operation.
Energy consumption
Data centres require continuous power for servers, networking equipment and the cooling systems that prevent them overheating. Unlike traditional computing tasks, AI inference — the process of running a model to generate a response — is computationally intensive even after training is complete. Every prompt submitted, every image generated, every document summarised draws from this energy pool.
Water usage
Cooling is one of the largest operational challenges for any data centre. Many facilities use evaporative cooling towers that consume millions of litres of fresh water annually. As AI workloads intensify and facilities scale, pressure on local water supplies — particularly in water-stressed regions — is increasing.
Carbon emissions
The carbon footprint of a data centre depends heavily on the energy source powering it. Facilities running on coal or gas-heavy grids produce significant direct emissions. Even renewable-powered facilities carry embedded carbon costs in hardware manufacturing, construction and logistics. The sheer scale of growth in AI infrastructure means absolute emissions continue to rise even as efficiency improves.
Hardware and rare materials
The GPUs and specialised chips that power AI are manufactured using rare earth minerals and metals — cobalt, lithium, tantalum — whose extraction carries its own environmental and humanitarian costs. Hardware lifecycles are short as performance demands increase, generating significant electronic waste.
Local land and infrastructure
Large-scale data centres require substantial physical footprints — often on the outskirts of cities or in regional areas with access to power infrastructure. Their construction impacts local land use, and their demand for power can strain regional electricity grids, sometimes pushing out lower-carbon residential or industrial users.
The growth problem
Efficiency gains in hardware and software are real, but they have historically been outpaced by demand growth — a phenomenon known as Jevons paradox. As AI becomes cheaper and more accessible, usage increases faster than efficiency improves, and total environmental impact continues to rise.
Putting the scale in context
In 2023, the International Energy Agency estimated global data centre electricity consumption at around 240–340 terawatt-hours annually — roughly equivalent to the entire electricity consumption of the United Kingdom. AI's rapid growth since then has accelerated this trajectory considerably.
Training a single large language model can emit hundreds of tonnes of carbon dioxide equivalent — comparable to the lifetime emissions of several cars. Inference — the ongoing process of running the model for users — adds to this continuously at scale.
Water consumption follows a similar pattern. A 2023 study estimated that GPT-3 consumed approximately 700,000 litres of clean freshwater during training alone. For context, that is enough to produce around 370 BMW cars. Operational inference adds ongoing water demand on top of this.
"The AI industry's rapid expansion is creating an unprecedented demand for energy and water resources that could undermine global sustainability goals if left unaddressed."
— Shaolei Ren, University of California Riverside, 2023
What thoughtful usage looks like
Individual actions won't reverse structural trends — but informed, intentional usage does matter, and organisations that think carefully about AI efficiency reduce environmental impact alongside operational cost.
Ask where your data is processed
Cloud and AI providers vary significantly in how much of their infrastructure runs on renewable energy. Choosing providers with credible renewable energy commitments — not just carbon offsets — makes a genuine difference.
Use smaller models where they're sufficient
Not every task requires the largest, most capable model available. Routing simpler queries to smaller, more efficient models reduces compute demand without sacrificing meaningful quality.
Reduce prompt iteration through better prompt quality
Every time a poorly structured prompt fails to produce a useful result and needs to be resubmitted, that's duplicate compute cost. Well-structured, clear prompts that work the first time are more efficient — for you and for the infrastructure running them.
Consolidate and reuse rather than regenerate
Rebuilding the same content or analysis repeatedly across an organisation, because prompts aren't shared or outputs aren't retained, multiplies the computational cost of that work. Shared prompt libraries and documented outputs reduce this duplication.
Be thoughtful about generative media
AI-generated images, video and audio are among the most compute-intensive outputs. Being intentional about when high-quality generative media is genuinely necessary — rather than default — reduces unnecessary load.
Awareness is the starting point.
The environmental footprint of AI is not a reason to avoid it — but it is a reason to use it thoughtfully. Every organisation that understands the cost of compute is better placed to make decisions that balance productivity with responsibility.
This page will be updated as new research and data becomes available.