Why power could limit the growth of AI

Training and running AI models is increasingly placing new electricity demand at data centres. As new data centres are built, will the bigger challenge be generating enough electricity, or getting reliable power to new sites on time? Whilst investment in AI computing infrastructure has accelerated, the associated electricity networks needed to support this investment are developing at a much slower pace. This introduces concerns that access to reliable and consistent electricity could become the primary determinant in the speed and location of future AI investment.

Demand could almost double by 2030

Figure 1 illustrates historical estimates and projections of electricity use across all data centres worldwide. Estimated electricity consumption increased from about 269 terawatt hours (TWh) in 2020 to 485 TWh in 2025, an increase of around 80% (IEA, 2026b). In its 2025 Base Case, the IEA projected demand of about 945 TWh by 2030, almost double the 2025 level (IEA, 2025). Its 2026 update keeps to that path, at around 950 TWh, and its full range of cases for 2030 runs from about 830 TWh to 1,010 TWh (IEA, 2026b). This growth will place increasing pressure on electricity infrastructure, particularly in regions with high concentrations of data centres. The pressure is likely to be most profound in certain regions where data centres cluster and the local power grid has limited space for new connections (IEA, 2025).

Importantly, sufficient electricity generation at a national level does not necessarily mean that individual regions have the grid capacity to accommodate large new centres, which usually have significant usage demands. These centres are particularly susceptible to this issue because they are often concentrated within existing technology hubs. According to the IEA, around 50% of data centres under development in the US are located within existing large clusters, potentially raising the risk of local bottlenecks (IEA, 2025).

Grids can take far longer than data centres

The disparity between constructing new data centres and the associated power infrastructure is evident. A new centre can be ready to open whilst the local grid still needs an upgrade to handle the extra demand (IEA, 2025). Figure 2 compares how long each can take, and Figure 3 shows how much planned capacity could face connection delays.

A new data centre can take 1 to 3 years, whereas planning, permitting and completing new grid infrastructure can take 5 to 15 years (IEA, 2026a).

Figure 3 shows IEA estimates of grid constraints, indicating that these could delay around 20% of planned global data centre capacity by 2030, or about 24 GW of the 112 GW added in its Base Case (IEA, 2025). This is an estimate of risk, not an observed outcome, and the remaining capacity is not guaranteed to connect on schedule. Grid development can take substantially longer than constructing the facilities it needs to supply. Transmission capacity, substations and connection approvals may become constraints on new data centre investment. Grid delays are not proof that electricity generation is universally sufficient; both generation and delivery constraints need to be examined.

Why delays are costly

For data centre developers, delays in obtaining electricity connections can have significant financial consequences, potentially leaving expensive infrastructure underutilised whilst awaiting access to power. Consequently, developers may increasingly prioritise locations with existing electricity infrastructure over otherwise attractive locations where grid connections might be harder to secure. This does however present investment opportunities in grid infrastructure, including transmission networks and substations. In February 2026 the IEA estimated that annual global grid investment must increase by about 50% by 2030, from $400 billion, in order to meet wider electricity demand, showing the scale of infrastructure investment required (IEA, 2026a). Note that this investment requirement relates to overall electricity demand growth, not just AI. A later IEA report put spending at nearly $450 billion in 2025, with further growth expected in 2026. It said part of that rise reflects higher equipment prices, and that investment remains below requirements (IEA, 2026c).

Generation is a second constraint

Another constraint identified relates to generation; even where the grid has room to connect a new data centre, the power system still needs enough generation to meet that extra demand. The IEA found that gas turbines can often take several years to deliver, which potentially pushes some new power plants’ opening dates beyond 2030 (IEA, 2025). Where meeting additional demand depends on the opening of a new power plant, the plant’s opening date can matter as much as the data centre’s connection date.

How developers can respond

If the local grid is the bottleneck, developers can look for sites where the network has room to connect a large new customer sooner (IEA, 2025). Comparatively, where connections are slow or generation is short, developers might pursue onsite gas generation; it should be noted however that long waits for turbines can limit how quickly it can supply power (IEA, 2026b). The practical question for developers is which investment can bring the site online sooner. Another practical solution is the deployment of onsite battery storage, allowing operators to reduce their reliance on the grid during periods of high demand (IEA, 2026b). These measures could help developers make more efficient use of existing electricity infrastructure, but they cannot eliminate the need for extra generation and grid investment.

Power may matter as much as computing

The rapid growth of AI driven electricity demand has exposed a disparity between the speed of data centre construction and the development of the infrastructure needed to power it. Whilst grid connection delays appear particularly significant in regions experiencing concentrated data centre expansion, insufficient generation capacity presents a clear bottleneck. Overall, the future expansion of AI infrastructure may depend as much on investment in electricity networks and power generation as on advances in computing technology.

References

International Energy Agency (IEA) (2025). Energy and AI. Paris: IEA. iea.org/reports/energy-and-ai

International Energy Agency (IEA) (2026a). Electricity 2026. Paris: IEA. iea.org/reports/electricity-2026

International Energy Agency (IEA) (2026b). Key Questions on Energy and AI. Paris: IEA. iea.org/reports/key-questions-on-energy-and-ai

International Energy Agency (IEA) (2026c). World Energy Investment 2026. Paris: IEA. iea.org/reports/world-energy-investment-2026