AI's Real Energy Constraint Is the Grid Queue, Not the Gigawatt Headline

Two numbers, two stories

The case that AI is about to eat the world's electricity supply rests on one trend line. In its 2026 update, the International Energy Agency put growth in data-centre electricity consumption at 17% for 2025, against 3% growth in global electricity demand, and projected that data-centre consumption doubles by 2030 while consumption from AI-focused facilities triples.

A peer-reviewed study published in September in Communications Sustainability lands somewhere else. Looking at the AI-specific build-out of six leading operators, it estimates electricity use of about 118 TWh in 2024 rising to between 239 and 295 TWh by 2030, or roughly 1% of global electricity demand.

Both cannot be the headline at once, and both are defensible. One measures a fast-growing sector. The other measures the share of the whole system that sector will occupy. The distance between them is where the real story sits.

Why the global share stays small

The IEA's own accounting explains the gap. Data centres consumed about 415 TWh in 2024, around 1.5% of global electricity, rising on the agency's base case to about 945 TWh by 2030 and 1,200 TWh by 2035. That is a large number in absolute terms, a little more than Japan's total consumption, and a modest one as a share of a system that grows every year anyway.

Efficiency is doing more work than the headlines credit. The agency finds that power consumption per AI task is falling at a rate it describes as unprecedented in the history of energy use. The catch is volume: more users, and more energy-intensive uses such as agents, keep pulling the total up even as each individual query gets cheaper.

For intuition, the IEA compares a typical AI-focused data centre to 100,000 households, with the largest facilities now under construction drawing 20 times that. Facilities of that size are not evenly spread, which is the part that matters to a grid engineer.

How the estimates are built

The two projections are not really competing. The IEA works from a new global model and dataset of data-centre electricity demand, then applies scenarios for utilisation, efficiency and the mix of AI and conventional workloads. The Communications Sustainability team combines firm-level expansion scenarios for six operators with projected AI workload intensity and a baseline trend for the rest of the sector. One set of numbers describes every rack; the other isolates the racks built for AI.

Scope does the rest of the work. A facility built for ordinary cloud workloads draws power whether or not anyone is training a model, and only part of its load is inference or training at any moment. As AI's share of that load rises, AI-specific consumption can climb fast while total data-centre consumption climbs more slowly. That is precisely the pattern the two studies report: a sector growing at double digits, inside a system where it still holds a low single-digit share.

Measurement is the weak link in all of it. UNECE notes that regulators lack sufficient information about how large facilities use electricity in real time, which makes forecasting, siting decisions and enforcement harder than the public debate assumes.

The constraint that actually binds

If the global share is small, why do utilities sound alarmed? Because the load arrives faster than the wires. The IEA estimates that around 20% of planned data-centre projects could be delayed unless grid constraints are addressed. New transmission lines take four to eight years to build in advanced economies, and lead times for transformers and cables have doubled in three years.

The UN Economic Commission for Europe made the same point in September: a large facility can be built and connected in two to five years, while expanding transmission and other grid infrastructure takes more than 10 years. Ireland has imposed connection restrictions in Dublin; the Netherlands has constrained where data centres can be built at all.

Nor is the pressure spread evenly. Data centres account for about one-tenth of global electricity demand growth to 2030, but more than 20% of that growth in advanced economies, where demand had been flat for two decades.

The mismatch has a second-order effect that is easy to miss. A project that cannot get a connection does not disappear; it relocates. Developers chase spare headroom into jurisdictions that can energise sooner, which is why a queue in one country shows up as a construction announcement in another.

Fatih Birol, the agency's executive director, put the dependency plainly: "there is no AI without energy".

Geography is the story

Concentration explains the politics. The United States took 45% of data-centre electricity consumption in 2024, China 25% and Europe 15%, and nearly half of United States capacity sits in five clusters. The September study finds more than 90% of projected AI computing capacity in North America, Western Europe and Asia-Pacific.

That turns a national statistic into a local fight. The study singles out Oregon, Ireland and Iowa as places where one load can dominate the local system, while Texas, with a larger and more interconnected grid, absorbs new demand more easily. A country can look comfortable on average and still have a county where the queue is the entire story.

Who pays for the upgrades is unresolved. UNECE notes there is no consistent framework for splitting connection costs between developers, utilities and billpayers, and warns that without one the necessary work slips.

The regulatory perimeter is also widening. Rules are starting to cover water use, emissions and local resource impacts rather than electricity consumption alone, but UNECE describes those rules as fragmented — which means two projects of the same size can face very different conditions depending on which side of a border they land.

What would change the curve

Three levers show up in both bodies of work. The first is flexibility. Data centres carry spare server capacity, and onsite batteries or backup generation can be run to suit the grid rather than only the site. The IEA argues that with the right incentives those assets make a data centre easier to integrate, not harder.

The second is supply. Technology companies signed around 40% of corporate renewable power purchase agreements in 2025, and conditional offtake agreements with small modular reactor projects grew from 25 GW at the end of 2024 to 45 GW. Onsite gas is filling the gap in the United States while those projects mature.

The third is efficiency, and it is where the range widens. In the IEA's High Efficiency case, data-centre demand in 2035 is 20% lower than in the base case; across all of its cases the 2035 figure spans 700 to 1,700 TWh. A forecast that wide is not a prediction. It is a statement about how much engineering and policy still matter.

There is also a version of this story in which AI is the supplier rather than the customer. The IEA finds that proven AI applications could cut energy costs in energy-intensive industries by 3 to 10 percentage points, and that operators blocked by slow connections are pairing onsite gas generation with batteries, because AI load swings are too fast for a turbine to follow on its own. Satellite tracking suggests many of those onsite projects are still early-stage, which is a warning against treating the workaround as a substitute for the grid.

Bottom line

The useful question is not whether AI will double the world's electricity use. On current evidence it will not. The question is whether the grids that host it can connect load faster than the load arrives, and that is a permitting, transmission and cost-allocation problem decided locally, one queue at a time.

Which makes the queue an instrument of industrial policy. Emerging and developing economies outside China hold about half the world's internet users and less than 10% of global data-centre capacity, and the IEA is explicit that reliable, affordable power is the precondition for changing that. A country that cannot schedule a connection will not host the compute, whatever its AI strategy says.

PulseGrid will keep tracking the gap between announced capacity and energised capacity. The second number is the one that does any work.