Data-centre electricity demand is set to more than double by 2030 to around 945 terawatt-hours — roughly Japan's entire consumption today — driven mainly by AI. The surge strains power grids and creates large, long-dated investment exposures across energy and infrastructure for asset owners.
For two decades, electricity demand in most advanced economies barely grew. Efficiency gains roughly offset rising consumption, and utilities planned around a flat horizon. Artificial intelligence has ended that calm. The data centres that train and run AI models are now among the fastest-growing sources of power demand in the world, and the consequences reach far beyond the technology sector — into the power systems, infrastructure assets and energy commodities that long-term asset owners hold across their portfolios.
How much power will data centres need?
The clearest authoritative estimate comes from the International Energy Agency, which projects that global data-centre electricity demand will more than double by 2030, reaching around 945 terawatt-hours. To make that figure tangible: it is slightly more than the total electricity consumption of Japan today. An entire G7 economy's worth of new demand, concentrated in a single category of building, arriving inside a decade.
Artificial intelligence is the engine. The IEA identifies AI as the most significant driver of the increase and expects electricity demand from AI-optimised data centres to more than quadruple by 2030. Within that, the servers running AI workloads — predominantly inference, the day-to-day task of answering queries rather than training models — are projected to grow their electricity consumption by roughly 30% a year. This single category is expected to account for almost half of the net increase in global data-centre electricity consumption between 2024 and 2030.
The United States illustrates the steepness of the curve. US data-centre grid-power demand is forecast to rise about 22% in 2025 to roughly 61.8 gigawatts, then climb to around 75.8 gigawatts in 2026 and as high as 134.4 gigawatts by 2030. That is more than a doubling in five years, and individual AI campuses are now being planned at scales exceeding a gigawatt apiece — each one comparable to the load of a large city.
Why does AI use so much more electricity?
The leap in demand is not incremental; it reflects a different kind of computing. Traditional data centres are built around general-purpose servers handling web, storage and enterprise workloads. AI computing depends on dense racks of specialised accelerators — graphics processing units and similar chips — packed tightly together and run at high utilisation. These systems draw far more power per rack, and they generate enough heat that cooling becomes a major load in its own right.
Training a frontier model is an enormous, concentrated burst of computation. But the larger and more persistent driver is inference at scale: once a model is deployed, every query it answers consumes power, and as AI is embedded into more products used by more people, that steady-state demand compounds. The result is a demand profile that is both very large and, unlike a one-off construction boom, structurally ongoing.
Can the grid actually handle it?
This is where ambition meets physics. Electricity grids in many regions are already strained, and adding tens of gigawatts of new, concentrated demand is not something the system can simply absorb. The IEA estimates that around 20% of planned data-centre projects could be at risk of delay unless grid constraints and supply bottlenecks are addressed — meaning the binding limit on AI's growth may turn out to be power, not chips or capital.
The problem is made harder by geography. Data centres cluster where land, connectivity and incentives align, which means new demand often lands on a small number of regional grids rather than spreading evenly. A single cluster can request more power than a local utility had planned to supply for years, forcing expensive and slow upgrades to generation and transmission. Interconnection queues — the waiting lists to connect new load and new generation to the grid — have lengthened accordingly. The mismatch between how fast a data centre can be built and how slowly transmission can be permitted and constructed is now one of the central tensions in the entire AI build-out.
What does this mean for long-term asset owners?
For a universal owner — a pension fund, sovereign wealth fund, endowment or insurer holding a broad slice of the global economy — the data-centre power surge is not a niche technology theme. It is a structural shift in the demand outlook for an entire sector they already own.
The most direct exposure is to power generation and the grid. A sustained increase in electricity demand supports the case for new generation of every kind — renewables, natural gas, and the renewed interest in nuclear, including small modular reactors — as well as for the unglamorous but essential infrastructure of transmission lines, transformers and grid equipment. Companies that make and install this equipment face order books shaped by AI demand. For asset owners building direct infrastructure portfolios, this is creating a large pipeline of long-dated, contracted assets of exactly the kind they seek.
The second exposure is through energy commodities and prices. If demand growth outpaces new supply, wholesale power prices can rise, with consequences that ripple from utility valuations to inflation expectations — a variable that matters acutely to funds managing long-dated liabilities. The third is the equity market itself: the technology companies driving the build-out and the energy and industrial companies supplying it are large index constituents, so the theme is embedded in passive holdings whether an investor seeks it or not.
There is also a risk that runs the other way. If AI revenue grows more slowly than the build-out assumes, or if more efficient models reduce the power needed per unit of useful output, some of the capacity now being planned could arrive ahead of the demand to use it — raising the prospect of overbuilt assets and disappointed returns. The same uncertainty that shadows AI infrastructure projects such as Stargate shadows the power demand built to serve them.
How should an asset owner respond?
The measured response is to treat the data-centre power surge as a genuine, multi-decade demand shift while underwriting each individual exposure on its own merits. The aggregate direction — more electricity demand, more pressure on grids, more need for generation and transmission — is about as well-evidenced as any forward-looking energy trend can be. That supports a strategic tilt toward power and grid infrastructure as a long-term theme.
What it does not support is uncritical enthusiasm for every data-centre or power asset on offer. The dispersion of outcomes is wide: location, grid access, power-purchase terms, tenant credit quality and the durability of underlying AI demand will separate the assets that compound from those that disappoint. The IEA's own analysis adds a constructive note — AI can help optimise grids, integrate renewables and improve energy efficiency — so the technology is both a source of demand and a potential tool for managing it. For the long-horizon owner, the task is to capture the demand trend through well-structured, well-located assets, while staying alert to the possibility that the build-out, like the AI boom that drives it, runs ahead of itself.