AI's Energy and Electricity Demand: The Data-Centre Power Surge

Executive summary

AI's energy footprint is dominated by the electricity that data centres consume. The International Energy Agency (IEA) estimates that global data centres used about 415 terawatt-hours (TWh) of electricity in 2024, equal to roughly 1.5% of world electricity consumption. In its 2025 Energy and AI report, the IEA projects this roughly doubling to about 945 TWh by 2030 in its Base Case — just under 3% of global electricity — with artificial intelligence as the key driver. Electricity demand from AI-optimised data centres is projected to more than quadruple over that period. IEA Executive Director Fatih Birol summarised the scale: data-centre demand is set to 'more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today.' The load is geographically concentrated — the United States alone made up about 45% of 2024 consumption, and the US plus China account for nearly 80% of projected growth. At the facility level, single sites have crossed the gigawatt threshold: Epoch AI tracks Colossus 2 in Memphis at about 946 MW of IT power. The strain is not only on generation but on grids, permitting, water for cooling, and emissions, which the IEA expects to peak near 320 Mt of CO2 from data-centre electricity by 2030. Figures labelled 2030 are projections.

Key findings

How much electricity does AI use? (question H2)

Almost all of AI's energy use shows up as electricity consumed by data centres. The IEA estimates that data centres worldwide used about 415 terawatt-hours (TWh) of electricity in 2024 — roughly 1.5% of global electricity consumption. AI is not the whole of that figure (data centres also run cloud storage, streaming and general compute), but AI-optimised servers are the fastest-growing slice.

For scale, 415 TWh is more than the annual electricity consumption of most countries. The IEA notes data-centre electricity has grown about 12% per year since 2017, more than four times faster than total electricity demand across all other sectors combined.

A single large AI-focused data centre consumes electricity comparable to about 100,000 households, and the largest facilities under construction can draw roughly 20 times that — power in the range of an aluminium smelter, per the IEA. That concentration is what makes AI's demand a grid-planning problem rather than a rounding error.

Why is AI driving the surge? (training vs inference and GPU power)

AI workloads run on power-dense accelerators — GPUs and custom AI chips — packed into racks that can draw tens of kilowatts each, far above the density of conventional servers. The IEA attributes the fastest growth to these 'accelerated servers,' whose electricity use it projects rising about 30% per year in the Base Case.

The demand splits into two phases. Training a frontier model concentrates enormous power into one cluster for weeks or months. Inference — serving the model to millions of users — is individually smaller but runs continuously and scales with adoption, so as AI products spread, inference becomes the dominant and steadily rising load.

The IEA projects electricity demand from AI-optimised data centres to more than quadruple by 2030. That is why AI, rather than traditional cloud computing, is treated as the key driver of the overall data-centre power curve.

What does doubling by 2030 mean? (the IEA projection)

The IEA's headline projection is a rough doubling of data-centre electricity demand from about 415 TWh in 2024 to about 945 TWh by 2030 in its Base Case — just under 3% of global electricity. Executive Director Fatih Birol framed it directly: demand is 'set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today.'

All 2030 figures are projections and depend on chip efficiency, model demand and how fast power can be built. The table below compares the 2024 baseline with the IEA Base Case for 2030.

The IEA also stresses a nuance often lost in headlines: even at 945 TWh, data centres account for less than 10% of total global electricity demand growth to 2030 — significant and concentrated, but not the largest single driver of world electricity growth.

Where is AI's power demand concentrated? (US and the biggest sites)

Data-centre demand is highly concentrated by geography and by facility. The IEA puts the United States at about 45% of global data-centre electricity in 2024, China at about 25% and Europe at about 15%. Looking forward, the US and China together account for nearly 80% of projected growth to 2030.

In the US specifically, the IEA finds data centres drive nearly half of national electricity demand growth to 2030, and by the end of the decade the country is set to use more electricity for data centres than for producing aluminium, steel, cement, chemicals and all other energy-intensive goods combined.

At the facility level, the frontier has crossed into the gigawatt range. Epoch AI's AI Data Centers dataset ranks the largest known sites by IT power capacity as follows.

What are the grid and siting constraints?

The binding constraint is increasingly the grid rather than the chips. Gigawatt-scale sites need firm, around-the-clock power and high-voltage transmission that can take years to permit and build, so developers cluster where interconnection and generation already exist. The IEA reported in 2025 that tightening bottlenecks were driving a scramble for solutions, from on-site generation to grid upgrades.

Because a single AI campus can rival the load of a mid-sized city, utilities face lumpy, fast-arriving demand that is hard to plan around. This has revived interest in dedicated generation — gas turbines, nuclear restarts and long-term nuclear contracts — sited next to data centres to bypass congested grids.

Epoch AI's data shows why planners are anxious: the power of leading AI supercomputers has roughly doubled every 13 months, and single-site compute capacity has doubled about every 7 months, outpacing normal grid build-out cycles.

How much water do AI data centres use for cooling?

Beyond electricity, cooling those chips consumes water. Industry and IEA estimates put a typical 100 MW data centre's water use at roughly 2 million litres per day when using evaporative cooling — comparable to a town of tens of thousands of people. Because frontier AI sites are far larger than 100 MW, their cooling water demand scales accordingly.

Water intensity varies widely by cooling design and climate: air- and closed-loop liquid-cooling systems use far less water but more electricity, creating a direct water-for-power trade-off. Siting in hot, dry regions — often chosen for cheap land and solar — raises both cooling load and local water stress.

The IEA and independent researchers flag water as a growing local-resource constraint, even where it is small relative to national totals, because withdrawals are concentrated in specific watersheds around large campuses.

What about carbon emissions from AI?

Emissions depend on how the extra electricity is generated. In the IEA's analysis, CO2 emissions from electricity generation for data centres rise and then peak at around 320 Mt of CO2 by 2030 in the Base Case, before declining as grids decarbonise — a level comparable to the emissions of a mid-sized industrialised country's power sector.

The near-term concern is that fast, firm demand can be met with fossil generation, particularly natural gas, where clean supply cannot be built quickly enough. The IEA notes AI-driven demand is expected to support additional gas as well as renewables and nuclear.

The offsetting case is that AI can also cut emissions elsewhere — optimising grids, industry and buildings — though the IEA treats those savings as potential rather than guaranteed, and dependent on deployment.

Can efficiency keep up? (efficiency gains vs Jevons paradox)

Per-computation energy use keeps falling. Chip efficiency and model optimisation mean each query or training step costs less energy over time, and hyperscale data centres run at low power-usage-effectiveness (PUE) overheads for cooling and distribution.

But efficiency gains have so far been swamped by growth in usage — a textbook Jevons paradox, where cheaper, more efficient compute drives so much more of it that total energy use rises. That is precisely why the IEA's absolute projection still doubles despite steady efficiency improvement.

The practical question is not whether AI gets more efficient per task — it does — but whether efficiency can outrun demand. On current trajectories, the IEA's Base Case says it does not through 2030.

What it means

AI's energy story is a power-system story. The credible anchor is the IEA's 2025 Energy and AI Base Case: about 415 TWh in 2024 rising to about 945 TWh by 2030, roughly a doubling, with AI as the key driver and the load concentrated in the US and China. Those are the numbers to cite; the 2030 figure is a projection.

For grids and policymakers, the challenge is speed and concentration — gigawatt sites arriving faster than transmission, generation and permitting can adapt, with water and emissions as secondary constraints. For the AI industry, power availability is becoming the true bottleneck on scaling, ahead of chips or capital.

The open questions are how much of the new load is served by clean firm power, whether efficiency can bend the curve, and how local water and grid stress are managed. Until those resolve, 'how much power does AI use' has a moving but well-bounded answer: about 1.5% of world electricity today, heading toward 3% by 2030.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

Figures are drawn from primary sources. Data-centre electricity totals, projections, growth rates, regional shares and emissions come from the IEA's Energy and AI report (2025) and its executive summary; each 2030 value is the IEA Base Case and is labelled as a projection, not measured data. Individual data-centre power-capacity figures come from Epoch AI's AI Data Centers dataset (accessed 2026-07-30), which ranks facilities by IT power (MW). Water-use figures are order-of-magnitude estimates for evaporative cooling and vary by design and climate. No figures were interpolated or invented; where the IEA and Epoch AI report ranges or rounded values, the rounding is preserved.

Data dictionary

Frequently asked questions

Glossary

TWh (terawatt-hour)

A unit of energy equal to one billion kilowatt-hours; used to measure annual electricity consumption at national and sector scale. ↗

Accelerated server

A server built around AI accelerators (GPUs or custom AI chips); the fastest-growing and most power-dense category of data-centre hardware, per the IEA. ↗

IT power capacity (MW)

The electrical power drawn by a data centre's computing equipment, in megawatts, excluding cooling and infrastructure overhead unless stated as total facility power. ↗

Jevons paradox

The observation that improving the efficiency of a resource's use can increase, rather than reduce, total consumption because it makes the resource cheaper to use more of. ↗

Sources

• IEA — Energy and AI (2025)
• IEA — Energy and AI, Executive summary
• IEA — Energy demand from AI
• IEA News — AI set to drive surging data-centre electricity demand (Fatih Birol)
• Epoch AI — AI Data Centers dataset

Related reports & data

• The Biggest AI Data Centers
• AI Compute Index