A figure now doing the rounds in the business press is easy to misread. The Wall Street Journal, drawing on the US Energy Department, has reported that data centres could consume as much as 12% of United States electricity by 2028, up from about 4.4% in 2023. The underlying study, prepared for the Department of Energy by Lawrence Berkeley National Laboratory, did not offer a single number. It set out a range of roughly 6.7% to 12% by 2028, depending on chips, how hard they are run, and how the buildings are cooled. It is a scenario for the American grid. It is not a forecast for a substation in the West Midlands.
What the 12% figure is, and is not
The top of the range is the number headlines keep. The bottom of the range is still a large rise on 2023. Both ends are projections. They can move if hardware gets more efficient, if orders slow, or if the buildings are used less intensively than the high case assumes. Quoting 12% as something that will happen, or as a share of British electricity, overstates what the study said.
The Energy Department's own summary puts the 2023 figure at about 4.4% and the 2028 estimate at 6.7% to 12%. A 2025 update to the same study has since moved the horizon to 2030, with a central estimate of 11.8% and a range of 9.5% to 15.3%. The press is using the older headline number. For a UK reader, the honest translation is narrower either way. The AI features you switch on are often running in those American, or similarly large, data centres. If that demand strains power supply, the effects that reach you are price, capacity and how casually vendors add always-on features.
Where a small business actually meets the cost
You will not see a line on your electricity bill labelled "model training". You may see it in a software invoice, a hosting plan that jumps a tier, or a feature that is cheap in a trial and metered once staff use it every day. Training a large model is the heavy spike. Everyday questions, often called inference, are the ongoing load. A company that sends every page view through a large model pays for that load whether or not the answer needed it.
That is a design choice. A page of clear service information does not need a model. A search across your own documents, or a chatbot that answers from pages you already wrote, uses far less than asking a general model to invent the answer. AI development that starts with one job is also the lighter build. An assistant connected to every inbox, running whether or not anyone asked it something, is the heavier one.
Hosting is part of the same decision
The website and the software still have to live somewhere. Cloud hosting on platforms such as Vercel, AWS or Azure is a practical choice for UK businesses that have outgrown a single cramped server. It is not a reason to move every background task into the largest, always-available machine. A review of what you run, how often it runs, and what can stay as an ordinary page, cuts waste before anyone debates a national energy strategy.
- Prefer published pages for facts that rarely change.
- Use a model where the input is messy: a free-text enquiry, a document, a question your site does not already answer.
- Switch off features nobody used after the first month.
- Ask the host what you are paying for when traffic is quiet, not only at launch.
None of this shrinks American data-centre demand by itself. It stops a small firm buying computing it does not need, and it keeps the site fast for the visitor, which is the performance problem UK customers actually notice.
Ask the host a plainer question
Energy strategy is a poor brief for a web project. A better brief is what this site costs to run on a quiet Tuesday, and what changes if you add a chat box to every page. Ask whoever hosts it how usage is metered, what happens at a traffic spike, and whether a background job keeps running when no one is looking. Those answers tell you more than a national percentage.
If the current host cannot say where the site runs, or a renewal date is a surprise every year, fix that before you add a model. A move to a platform you can actually operate is often the saving, not a larger plan. The electricity story is the backdrop. The decision in front of you is still which features earn their keep.
Read the forecast, then choose a smaller tool
Energy reporting is a reason to be specific, not a reason to avoid software. A booking form, a well-hosted site and a narrow assistant can all be justified. A plan to train your own general model, because a competitor mentioned one, usually cannot. The 12% case is what happens if the high-growth assumptions hold in the United States. Your decision is whether your next feature belongs in that growth or on a page you already control.
After the choice, someone still has to watch the hosting, the certificates and the features that creep in. Maintenance and support covers that ordinary upkeep. Web Works Rise is a UK-led team, with the office in Birmingham and a development hub in Tunisia that adds delivery capacity. If you want a straight view of what your site and any AI feature are running on, contact the team. Bring the current host and the job you hope the software will do.