The question tends to arrive buried in a client questionnaire. Somewhere below the data protection section sits a line asking you to describe the environmental impact of any AI tools used in delivering the work. If your team uses ChatGPT, Copilot or Gemini every day, that line is awkward to answer. The tools feel weightless, yet every prompt runs through a data centre that draws electricity and, at many sites, evaporates fresh water to stay cool.
I have watched founders react to this in two ways. Some dismiss it as greenwash paperwork. Others feel a vague guilt and say nothing. Both responses miss the useful middle ground, which is knowing roughly what the numbers say and having a proportionate answer ready.
What are the hidden energy and water costs of AI tools?
AI tools such as ChatGPT, Copilot and Gemini run in large data centres that consume significant electricity and use fresh water for cooling. Training a large model takes an enormous one-off dose of energy, and every query afterwards, called inference, uses far more power than an ordinary web search. Because providers rarely publish per-query figures, the footprint stays hidden from the businesses using the tools.
The training numbers grab the headlines. Building GPT-3, the model behind the first version of ChatGPT, is estimated to have used about 1,287 MWh of electricity and produced roughly 502 tonnes of CO2, comparable to the annual emissions of 112 petrol cars. An earlier academic study put the training emissions of one large language model at around 300 tonnes.
The everyday numbers are smaller but they never stop. One AI query is commonly estimated to use around ten times the electricity of a standard web search. Water follows the electricity, since cooling consumes roughly 1.7 litres per kWh at a typical data centre. Researchers behind one widely cited 2023 paper estimated that a session of 10 to 50 chatbot responses indirectly consumes about 500 ml of fresh water, a small bottle’s worth. None of these figures is precise, because the providers publish so little, but the direction is consistent.
Why does this matter for your business?
The footprint matters less for its size today than for the questions it is starting to generate. Larger customers, lenders and public sector tenders increasingly ask suppliers about the environmental impact of their IT, and AI use is joining that list. Meanwhile the aggregate numbers are growing fast enough that regulators have started demanding transparency from the model providers themselves.
The scale explains the regulatory attention. The International Energy Agency projects that data centres, AI and cryptocurrency together could draw more than 1,000 TWh of electricity by 2026, roughly double the 2022 figure and close to what Japan uses in a year. Google reported that its data centres used about 5 billion US gallons of fresh water in 2022, a fifth more than the year before. Academic projections put global AI-related water use at 4.2 to 6.6 billion cubic metres a year by 2027, more than half of the UK’s annual water withdrawal.
Regulation is following the numbers. The EU AI Act requires providers of powerful general purpose models to document energy and resource use across the lifecycle, and the CMA’s foundation models review flagged environmental transparency as information businesses need to make informed choices. You are a user of these models rather than a provider, so neither obligation lands on you directly. The questions still flow downstream, because your larger clients need answers from their supply chain to meet their own reporting commitments.
Where will you actually meet these costs?
You will meet them in paperwork before you meet them in a bill. Supplier questionnaires from larger customers, ESG sections in tenders, and lender due diligence are where the topic first lands on an owner-managed business. Behind the paperwork, the footprint itself sits upstream, in your providers’ data centres, which is also where nearly all of your influence over it lives.
For a firm of five to fifty people sending a few hundred prompts a day, the office kit barely registers. Your influence runs through which tools you chose, how heavily your team uses them, and where the workloads run. Day-to-day use adds up more than people expect. Inference, the running of queries rather than the training of models, can account for up to 60 per cent of an AI system’s lifetime energy.
Usage patterns separate the light from the heavy. A few staff drafting emails and summarising documents sits at the light end. Automated image or video generation at scale, or an always-on chatbot handling thousands of queries a day, sits at the heavy end and deserves scrutiny before you build it. If you run workloads directly on a cloud platform, both Microsoft and Google publish tools that estimate the carbon impact of specific services and let you pick data centre regions with cleaner grids.
When should you act, and when can you let it go?
Act when your use is heavy, when clients start asking, or when you next renew a major software contract. Let it go, for now, if your use is light and occasional, because a writing assistant used by two staff adds a small footprint next to travel, heating and general IT. Proportionate attention beats both denial and guilt.
There are honest counterweights to the alarm. Electricity grids in the UK and Europe are decarbonising, so the carbon cost of each query should fall even as usage grows. Hardware and model efficiency keep improving. And AI that removes business travel, trims materials waste or replaces an inefficient process can be a net environmental gain, a point the UK government’s own decarbonisation programme makes when it funds AI for exactly that purpose.
The Monday move is small. Add one line to your ESG or responsible business policy covering purposeful AI use and supplier sustainability checks. At the next contract renewal, ask your main software vendors what data or commitments they can share on the energy and water use of their AI features. Prefer UK or EU hosting regions where the grid is cleaner and client data rules allow. And point staff at the smaller, cheaper model tiers for routine tasks, keeping the biggest models for work that genuinely needs them. Efficiency and economy point the same way here, which makes this one of the easier sustainability questions to act on.
What related terms will you hear alongside this?
Five terms cover nearly every conversation on this topic. Training and inference describe the two phases of an AI model’s energy life. Scope 3 emissions is the accounting category your cloud and AI use falls into. Grid mix explains why the same query costs more carbon in one country than another. And edge AI describes the move to run smaller models locally.
Training is the one-off build of a model and inference is every use of it afterwards; when a vendor talks about efficiency gains, ask which phase they mean. Scope 3 covers the emissions that happen in your supply chain rather than your own building, which is exactly where AI’s footprint sits for an owner-managed business, and it is the category larger clients will ask about. Grid mix matters because a data centre running on renewables produces a fraction of the carbon of the same workload on a fossil-heavy grid, which is why region choice is one of your few real levers. Edge AI, running small models on laptops and phones rather than in hyperscale data centres, will reach you through packaged products rather than anything you deploy yourself, and it trims both the data transfer and the cooling load.
The tools can keep their place in your business. What changes is that you now have an answer ready for page four of the questionnaire, which is more than many of your competitors will manage.
