A few weeks into deploying AI tools across your team, someone raises the question at a Thursday stand-up: isn’t all this going to cost the earth in energy? You give a half-answer, because two figures you’ve come across in the news do not match. One source puts an AI query at roughly the same as a Google search. Another puts it at ten times that. Both figures can be accurate. They’re answering slightly different questions.
AI energy consumption is one of those topics where headline numbers vary so widely that it is easy to dismiss them as noise. This post explains where the variation comes from and what it means for a business like yours.
What is AI energy consumption?
AI energy consumption is the electricity used at each stage of an AI system’s life. Building a model from scratch, the training phase, uses significant power and is done once, by the provider. Running the model for users, the inference phase, uses power every time someone sends a prompt. For a business that has bought access to a third-party model rather than built one, inference is the relevant part.
Published estimates for inference vary considerably. One UK-facing summary puts a GPT-4o text query at around 0.3 watt-hours. Jisc’s National Centre for AI, in an explainer published in May 2025, uses a rougher benchmark of 2.9 watt-hours per LLM request. The tenfold gap between those two figures is not an error. It reflects differences in model size, prompt length, output length, hardware generation, data-centre cooling, and what exactly counts as a single request.
That variability is also why AI energy figures appear contradictory when you read coverage from different sources. Closed-model providers do not publish enough technical detail to allow a universal per-query number. Any confident claim about exactly how much electricity a specific AI query uses should be read as an estimate. The range is real, and it narrows as infrastructure improves and as models become more efficient over time.
Why does it matter for your business?
The electricity your AI tools use is consumed at data centres operated by your provider, not in your office. That distinction matters because the environmental footprint of a business using commercial AI sits almost entirely in the supplier’s infrastructure. How efficiently that infrastructure is run, where it is located, and what energy source powers it are decisions the provider makes, not you.
Scale is not trivial, though. Analysis from the UN Regional Information Centre, drawing on IEA data, puts global data-centre electricity use at about 1.5% of global supply in 2024, with a projection to approach 3% by 2030. The UK’s AI sector has grown to more than 5,800 companies with estimated revenue of around £23.9 billion and gross value added of approximately £11.8 billion, according to the UK Government’s 2024 AI Sector Study. As commercial AI use grows, data-centre demand grows with it.
As a business running AI tools, you are contributing to that demand. That does not by itself call for anxiety, and the direct contribution is modest at the scale typical of an owner-managed business. What it does mean is that energy transparency is becoming a reasonable question to put to any AI supplier, in the same way you might ask about data residency or security practices. Over time, as providers face increasing pressure to publish more detail, those conversations will become easier.
Where does AI energy use actually show up?
For the typical owner-managed business, AI energy use surfaces indirectly: slightly higher cloud usage, more API calls, extra vendor processing overhead. A team using a chatbot for a few hours each day draws a small fraction of what the office heating consumes. Jisc’s National Centre for AI puts one hour of Netflix HD at 0.077 kWh; a single LLM query at roughly 0.003 kWh.
By that measure, dozens of daily queries from a small team add up to roughly the energy of a short video call or a brief streaming session. The UK Government’s 2022 AI Activity in UK Businesses report found around 15% of UK businesses, roughly 432,000 companies, had adopted at least one AI technology at the time of the survey. Adoption was far higher in larger firms: 68% of large companies had adopted at least one AI technology compared with 15% of small ones.
Owner-managed businesses currently sit at the lower end of AI intensity. The tools are in use, but typically for short text tasks rather than continuous background processing. Where energy overhead becomes more noticeable is when AI is doing something heavier: generating images at volume, processing large documents across a team, or running recurring automated workflows. Understanding which of those tasks your business is actually running, and at what frequency, is where the energy picture starts to become genuinely worth tracking.
When should you pay attention to it?
When choosing AI tools or managing how your team uses them, the variable worth tracking is model size relative to task. A large general-purpose model applied to routine work, summarising a report, drafting a reply, extracting key figures, costs more in API fees and more in electricity than a smaller, task-specific model. The gap matters at scale. The right model is the smallest one that gets the job done adequately.
UK regulators have not issued energy-specific compliance rules for business AI adoption. The ICO’s focus is data protection, lawfulness, and transparency under UK GDPR. The FCA’s AI guidance covers governance, controls, and operational resilience for regulated firms. The NCSC advises on secure deployment, prompt injection risk, and supplier assurance. None of these demand an energy audit, but all of them expect you to understand what tools you’re deploying, who operates them, and what governance is in place. That is a reasonable standard for any supplier relationship involving personal data or business-critical processes.
In that context, the energy question is one signal among several that tells you how much a supplier understands and can account for in their own infrastructure. A vendor who can answer questions about data residency, security practices, and energy sourcing is likely running a more considered operation than one who cannot. The energy question works as a useful proxy for supplier maturity.
What other questions does this connect to?
AI energy use sits inside a wider conversation about data centre demand, vendor transparency, and sustainability disclosure. The EU AI Act introduces documentation requirements for AI providers, including technical disclosures that may, over time, include energy-related detail. UK businesses buying AI tools from EU-facing vendors may start seeing that information surface in procurement documentation and vendor contracts.
The Brookings Institution’s analysis of global energy demands and the AI regulatory landscape notes the tension between AI growth and energy infrastructure planning. As governments look at data-centre demand more closely, transparency expectations on providers are likely to increase gradually. Voluntary disclosure will give way to structured reporting for at least some categories of AI system.
For owner-managed businesses, the practical posture is straightforward: use AI for clear, specific tasks; avoid unnecessary prompting loops; prefer smaller models where they are adequate; ask suppliers what they can share about their infrastructure; and fold AI into your ordinary procurement, security, and governance reviews. The energy question is worth having a working handle on. For businesses running AI at current scale, it is rarely the most urgent item on the list. Accuracy, data handling, and governance typically sit above it. If you have those three in reasonable shape, the energy question looks after itself.
