A founder running a twenty-person consultancy in Manchester recently described a question she kept putting aside. She had seen headlines connecting AI to the energy consumption of small countries, and she wanted to know whether switching her team to AI-assisted workflows would create a sustainability liability she hadn’t planned for. What she actually needed was the business-level answer, not the global figure.
What is AI’s actual energy footprint?
A single AI chat query uses roughly ten times the energy of a standard Google search, around 0.0029 kWh compared with 0.0003 kWh per query. Generating one AI image uses approximately the same energy as charging a smartphone fully. Training a large model is far more intensive: OpenAI’s GPT-3 training consumed an estimated 1,287 MWh, producing roughly 502 metric tonnes of CO₂, the annual equivalent of emissions from around 112 petrol cars.
These figures are real, but they need context. Once a model is trained, running it day-to-day, the process known as inference, can account for up to 60% of a model’s total energy consumption over its life. Queries run continuously at scale across millions of users. The training cost is a one-time event. The ongoing running cost compounds quietly in the background.
At the aggregate level, global data centres account for roughly 1-2% of global electricity demand, broadly comparable to the airline industry. The International Energy Agency projected in its 2024 Electricity report that demand from data centres, AI, and cryptocurrency could double between 2022 and 2026. Machine learning and AI currently account for less than 0.2% of global electricity demand and less than 0.1% of global greenhouse gas emissions, but both figures are rising as adoption grows.
Why does AI’s energy use matter for your business?
For a business using cloud-based AI tools from Microsoft, Google, or OpenAI, the electricity runs through the provider’s data centres, not your office meter. You see the energy cost as a subscription fee or usage charge. The indirect routes it takes to your business are through cloud pricing, regional wholesale electricity costs, and, for firms subject to extended ESG reporting, Scope 3 emissions disclosures.
To illustrate the scale in business terms: if fifty staff collectively ran 10,000 AI queries per working day, the energy total at the quoted per-query rate would be roughly 7,000 kWh per year if that compute ran on the firm’s own hardware. That is comparable to the annual electricity consumption of a small office. In practice, that compute runs in the cloud, so it doesn’t touch the office meter. It shows up in the cloud bill.
The indirect electricity price effect matters too. EirGrid, Ireland’s grid operator, reported that data centres already account for around 18% of Ireland’s national electricity demand, projected to rise to 28-30% by 2031. The UK’s National Grid ESO has flagged data-centre and digital demand as a material driver in grid planning out to 2035. These are infrastructure pressures on business energy costs, not costs that arrive via a cloud invoice, but they are real.
Where will you actually encounter AI’s energy costs?
Owner-managed businesses tend to meet AI’s energy story in three places. On your cloud bill, AI workloads are priced per token, per image, or per user licence. On your electricity tariff, grid-level pressure from data-centre growth can influence what UK businesses pay for power over time. And in your ESG reporting, if your firm discloses carbon emissions, cloud AI may need to appear as a Scope 3 line item.
The cloud cost is the most immediate. The major providers charge by usage rather than by kilowatt-hour, which means the energy relationship is invisible at the point of purchase. What you see is the bill. What you can control is query frequency, the choice between a smaller and a larger model for a given task, and whether you run queries in batches or one at a time.
The UK’s Streamlined Energy and Carbon Reporting framework requires many medium-sized businesses to report their electricity use and greenhouse gas emissions, including Scope 3 where material. The Science Based Targets initiative and similar frameworks are beginning to treat cloud AI workloads as a specific reporting category. Ask your reporting advisers whether your current usage volume meets the materiality threshold for your framework, rather than assuming it sits comfortably below it.
When should you factor this in, and when can you set it aside?
For a firm running standard cloud AI tools across a small team, the energy component is unlikely to dominate your cloud bill or appear on your office meter. Active attention is warranted when usage grows significantly, your ESG reporting scope expands, or a client starts asking supply-chain questions about your carbon footprint. The signal to watch for is a noticeable and sustained jump in cloud spend as AI usage scales.
The regulatory picture currently puts the spotlight on providers rather than on the businesses using their tools. The ICO’s AI guidance focuses on fairness, transparency, security, and accountability rather than energy thresholds. The NCSC’s secure AI guidelines require you to understand your cloud and hardware dependencies, connecting to data-centre energy and supply-chain resilience. The CMA’s 2023 review of foundation models flagged high energy requirements as a barrier to entry in AI infrastructure, pointing to concentration risk at the hyperscaler level.
The EU AI Act introduces transparency and energy-efficiency documentation requirements for systemic general-purpose AI models with very high compute usage. The UK is not bound by it, but UK businesses serving EU clients or using EU-hosted AI may encounter those requirements through supply-chain obligations. Client-driven requirements are likely to arrive before formal regulatory ones for many owner-managed businesses.
What else sits alongside AI’s energy costs?
Three areas tend to come up in the same conversation. AI governance: the ICO expects Data Protection Impact Assessments for high-risk AI use, and proportionality checks can include the environmental cost of processing. Vendor selection: your provider’s renewable energy commitments determine whose infrastructure you rely on. Model efficiency: smaller, task-specific models use materially less energy than large general-purpose ones, making tool choice as important as usage frequency.
On vendor selection, the major cloud providers all publish sustainability disclosures and renewable energy commitments. The picture is one of rising absolute emissions alongside renewable energy targets: infrastructure build-out for AI has increased total energy consumption even as renewable percentages have grown. Comparing provider sustainability disclosures before committing to a platform is standard procurement practice for businesses with ESG commitments, and it takes an afternoon rather than a project.
On model efficiency, MIT analysis suggests that smarter scheduling and more efficient model selection can reduce data-centre electricity demand by 10-20%. The practical implication is that choosing a smaller language model for a routine task, rather than defaulting to the most capable option available, reduces both the cost and the energy overhead of that query.
The founder in Manchester decided to treat AI energy use the way she treats cloud security: background due diligence rather than a reason to pause. She asked her cloud provider for their sustainability disclosure, noted the renewable energy commitments against her own carbon targets, and flagged AI cloud spend as a line to monitor in her next ESG review. It took an afternoon. If the same question has landed for you, that is a reasonable place to start. If you want to think through what this looks like for your specific operation, Book a conversation.
