A proposed data centre with a large grid-connection request can sound like an immediate claim on the country's electricity supply. In reality, a planning application, a connection reservation, a financed building and an operating facility describe four different stages. They also may refer to different measures of power. The Carbon Trust's 4 November discussion of Britain's AI infrastructure gets to the heart of the confusion: debates about capacity are too often built from quantities that cannot responsibly be treated as if they were identical.[1]
The independent climate organisation set out seven areas for policy and industry attention after convening discussion among stakeholders. Its article covers access to the grid, demand flexibility, on-site energy, the character of AI workloads, planning and community concerns, technical innovation and better standards for measurement. The account describes a discussion conducted under Chatham House Rule. It is not a consensus position attributed to every participant, an independently measured national demand forecast or a new set of legally binding UK rules.[1]
A place in the queue is not electricity consumed
A project may seek a substantial connection before committing to build. Developers need enough promised power to win tenants and financing; operators and grid planners need to avoid reserving the same scarce resource indefinitely for projects that never arrive. A national tally of connection requests can therefore be much larger than the load ultimately connected and used at the same time. Those distinctions make the problem more complicated, not less serious.[1]
The relevant comparison includes requested capacity, approved connection capacity, installed IT capacity and measured peak and average demand. It also needs the expected start date, occupancy and utilisation. Announcing one gigawatt of potential data-centre capacity is not evidence that one gigawatt of computing load is running around the clock, and neither metric directly describes the electrical energy consumed over a year.[1]
Checks that change the debate
| Measure | Value | Context |
|---|---|---|
| Connection request | Not the same as constructed or operating demand | |
| Training workloads | Some may be time-flexible; restrictions apply | |
| Inference workloads | Latency can limit shifting | |
| On-site generation | Fuel and emissions must be accounted for | |
| Publication status | Carbon Trust discussion and analysis, not binding regulation |
Can AI work when and where power is available?
The Carbon Trust raises demand flexibility as an opportunity. Some training jobs, batch analysis and maintenance tasks can be delayed or moved between regions when electricity is expensive or carbon-intensive. Many live applications cannot: a customer waiting for a response needs the service now. Even flexible work may face limits imposed by data locality, hardware reservations, networking and software coordination.[1]
A meaningful flexibility promise should identify the actual shiftable share of workloads, maximum delay, geography, power adjustment achieved and any additional equipment or data-transfer energy needed. If a developer says its campus is flexible but cannot disclose what share of demand will move or shut down during grid stress, planners cannot value that flexibility confidently.[1]
The risks of solving grid scarcity with private generators
On-site gas generation and other behind-the-meter arrangements may let a project start sooner than an upgraded grid connection. That can shift emissions and local air-quality burdens rather than remove them, and it can create dependence on fossil-fuel infrastructure that persists long after a temporary shortage ends. Conversely, well-designed local energy or storage assets may support the wider network. The outcome depends on fuel, emissions, operating hours, interconnection and who pays for the assets.[1]
Water adds a similarly local dimension. Cooling technologies can trade electricity demand against water consumption, but the best choice in a wet coastal region may be unsuitable in a water-stressed catchment. The same applies to waste-heat reuse: a promising diagram is not evidence of a functioning heat buyer, compatible seasonal demand or affordable connection to nearby homes.[1]
Local consent needs an intelligible bargain
Communities may be asked to accept large buildings, construction traffic, land use, noise and new grid equipment in exchange for jobs, tax contributions or the possibility of economic development. The benefits depend on the phases of construction and operation: the labour required to build a campus is not the permanent staffing of a completed facility. An honest appraisal should show both and explain where public infrastructure costs fall.[1]
AI workloads also change over time as chips, cooling methods and model architecture evolve. Forecasts should distinguish training from inference and allow scenarios rather than one confident number. Better reporting standards would disclose the definitions and time horizons behind power figures, alongside actual energy use and relevant water and carbon measures.[1]
The Carbon Trust's seven-part agenda does not answer whether any particular UK project should proceed. It provides a more useful way to ask the question. A credible application should be judged on its stage of commitment, connection and operating timetable, flexibility, total resource impacts and evidence of local benefit. The country needs a comparison of deliverable projects and costs, not a race to announce the biggest theoretical megawatt total.[1]
Sources & further reading
- Powering the AI era Carbon Trust · 4 Nov 2026 · Accessed 5 Nov 2026