The NewsRupt

Published forecasts for data-centre electricity demand over the next five years differ by roughly a factor of five. That spread is the most informative thing about them, and it is usually the first detail dropped when a forecast is reported.

The disagreement is not mainly about arithmetic. It is about assumptions that the forecasts do not share and often do not state prominently: how quickly announced capacity actually gets built, how much of it is AI-specific rather than general cloud growth, what average utilisation will be, how fast efficiency improves at the chip and cooling level, and whether demand growth continues at recent rates or saturates. Change two of those inputs within defensible ranges and the headline number moves by several hundred terawatt-hours.

Some things are reasonably well established. Data centres currently account for a low single-digit percentage of global electricity consumption, and rather more than that in a handful of concentrated markets — Ireland, parts of the Nordics, Virginia, Singapore — where the local grid effect is large even though the global one is not. Consumption is growing, and growing faster than it was five years ago. Individual AI training and inference clusters are far more power-dense per rack than the facilities that preceded them, which creates transmission and cooling constraints that are local and immediate regardless of what the global total does.

Several things are widely reported and much weaker than they sound. Estimates of energy per individual query circulate with two significant figures and are, in most cases, derived from assumptions about hardware, model size, batching and utilisation that the authors could not have known; providers do not publish per-query figures, and the true number varies by orders of magnitude between a small model answering a short prompt and a large one performing extended reasoning. Announced capacity is a statement of intent, not construction: a substantial share of announced projects are delayed, downsized or cancelled, frequently because grid connection queues run to years. And comparisons between AI energy use and national consumption are usually constructed to be arresting rather than to be informative.

The constraint that is actually binding in most markets is not generation but interconnection. In several of the largest data-centre regions, the wait for a grid connection of meaningful size is now measured in multiple years, and transmission build-out is slower than both generation build-out and demand growth. This is why operators are signing long-term power agreements, exploring on-site generation and, in some cases, siting facilities according to grid availability rather than proximity to customers. It is also why the local politics — water use for cooling, ratepayer cost allocation, land — is intensifying faster than the national-level energy debate.

Efficiency is the countervailing force and it is genuinely substantial. Performance per watt at the chip level has improved sharply across successive generations, inference-serving techniques have reduced the compute required per unit of output, and smaller task-specific models are displacing large general ones in exactly the high-volume workloads that dominate aggregate consumption. Whether efficiency gains outrun demand growth is the central open question, and the historical record on this — where efficiency has often enabled more consumption rather than less — does not favour optimism.

The limitations of any assessment here are severe and should be stated rather than buried. Operators publish little granular data; most public estimates rest on modelled assumptions rather than measurement. Reporting boundaries differ, with some figures including cooling and networking and others not. Regional mixes vary enormously, so the same terawatt-hour carries very different emissions depending on where it is drawn. And forecasts of a fast-moving industrial buildout have a poor historical track record in both directions.

The defensible position is therefore an uncomfortable one for anybody wanting a number. Data-centre electricity demand is rising, the local effects in concentrated markets are already material and already political, the global effect remains modest relative to industry and transport, and the five-year figure is not knowable to the precision at which it is routinely quoted. A forecast presented without its assumptions is not a forecast; it is a position, and the spread between positions is currently wide enough to accommodate almost any argument somebody wants to make.