How Much Electricity Do AI Data Centres Use? The Official Numbers Explained

Headlines about artificial intelligence and electricity tend to swing between two extremes: AI is about to break the grid, or the fuss is overblown. The official numbers sit somewhere more interesting. Data centres still use a small slice of the world's power, but that slice is growing fast, it is concentrated in a handful of places, and AI hardware is changing what a single building draws.

This guide sets out what the main official estimates actually say, where the electricity goes inside a data centre, and how to read the next big projection you see without being misled by it.

Key facts at a glance

• Global use today: the International Energy Agency (IEA) estimates data centres consumed about 415 terawatt-hours (TWh) in 2024, around 1.5% of the world's electricity.
• Growth rate: that consumption has grown by about 12% a year since 2017, more than four times faster than total electricity use.
• The 2030 outlook: the IEA's base case has data centre demand more than doubling to around 945 TWh by 2030, slightly more than Japan uses today.
• The US picture: a Lawrence Berkeley National Laboratory report published by the US Department of Energy found data centres used about 4.4% of US electricity in 2023 and could reach 6.7% to 12% by 2028.
• Where it is concentrated: the United States accounted for 45% of global data centre electricity use in 2024, China 25% and Europe 15%.

Who this guide is for

This explainer is for engineers, IT buyers and planners who need to reason about AI infrastructure, and for anyone trying to make sense of competing claims about AI's energy footprint. It does not assume an energy background. If you are choosing a cloud region, planning on-premises GPU capacity or simply evaluating a vendor's sustainability claims, the same basic numbers apply.

Where the power goes inside a data centre

A data centre is mostly a building full of computers plus the equipment needed to keep them running. The IEA breaks the load into a few parts:

Source: IEA, Energy and AI (2025).

The wide range for cooling is the most useful number in that table. It means two buildings with identical servers can have very different total bills. The industry's standard way of expressing this is power usage effectiveness, or PUE: total facility energy divided by the energy used by the IT equipment. A PUE close to 1.0 means almost all the power reaches the computers.

Why AI changes the picture

Two things make AI data centres different from the ones that ran email and web hosting.

Accelerated servers use more and grow faster. The IEA projects electricity use by accelerated servers, mainly driven by AI, to grow by about 30% a year in its base case, compared with about 9% a year for conventional servers. Accelerated servers account for almost half of the projected net increase in global data centre consumption.

Power density is rising. Racks of AI hardware pack far more computing into the same floor space. NVIDIA's GB200 NVL72, for example, connects 72 Blackwell GPUs and 36 Grace CPUs in a single rack-scale, liquid-cooled design. Liquid cooling becomes necessary because air struggles to remove that much heat from one cabinet. NVIDIA says the design also saves floor space, and claims 25 times more performance at the same power than its earlier air-cooled H100 infrastructure.

Scale is industrial. The IEA notes that AI-focused data centres can draw as much electricity as power-intensive factories such as aluminium smelters. Global investment in data centres has nearly doubled since 2022, reaching about half a trillion dollars in 2024.

The US numbers in detail

The US is where the growth is most visible. According to the Berkeley Lab estimates released by the Department of Energy in December 2024, total US data centre electricity use climbed from 58 TWh in 2014 to 176 TWh in 2023, and could reach between 325 and 580 TWh by 2028. The department said data centre load growth had tripled over the past decade and was projected to double or triple again by 2028.

The IEA reaches a similar conclusion from a global angle. It expects US data centres to account for almost half of the country's growth in electricity demand to 2030, and says that by the end of the decade the US is set to use more electricity for data centres than for producing aluminium, steel, cement, chemicals and all other energy-intensive goods combined.

Elsewhere the share of growth is lower but still significant. In advanced economies as a whole, which have seen decades of flat demand, the IEA says data centres account for more than 20% of demand growth to 2030. Its executive director, Fatih Birol, has said data centres will account for more than half of demand growth in Japan and as much as one-fifth in Malaysia.

The real bottleneck: the grid

The global percentage hides the local problem. The IEA found that nearly half of US data centre capacity sits in five regional clusters, so a few utilities feel the pressure far more than the national average suggests.

Building a data centre can take two to three years, but building the energy system around it takes much longer. The IEA says new transmission lines can take four to eight years in advanced economies, and estimates that around 20% of planned data centre projects could be at risk of delay unless grid risks are addressed.

On supply, the agency expects renewables to meet about half of the growth in data centre demand, adding over 450 TWh of generation by 2035. Natural gas expands by about 175 TWh, notably in the US, with nuclear contributing about the same amount of additional generation.

How to check a headline claim

When you see a new figure about AI and electricity, run through these questions:

1. Energy or capacity? TWh measures electricity used over a year. Gigawatts (GW) measure the maximum a site or fleet can draw at one moment. A "1 GW campus" is a capacity figure, not annual consumption.
2. All data centres or just AI? Most official totals cover every data centre, including cloud, streaming and enterprise sites. AI is the fastest-growing part, not the whole.
3. Which scenario? The IEA publishes a base case plus Lift-Off, High Efficiency and Headwinds cases. By 2035 its range spans from 700 to 1,700 TWh, against a base case of about 1,200 TWh. Quoting only the top end overstates the certainty.
4. Global, national or local? A 1.5% global share and a strained local grid can both be true.
5. What year is the baseline? Figures for 2023 and 2024 are estimates published with a lag, so check which year a number refers to.

What if the numbers seem to conflict?

They often will, and that is not necessarily a sign that one source is wrong. The Berkeley Lab range for 2028 is wide, from 6.7% to 12% of US electricity, precisely because it depends on how quickly AI hardware is deployed and how efficient it becomes. The IEA's High Efficiency case, for example, ends up with data centre demand 20% lower in 2035 than its base case.

If two estimates differ, compare their scope, units and assumptions before comparing the headline figures. And treat single-company projections or investor presentations as claims rather than measurements.

The official view

Both the IEA and the US Department of Energy frame data centre growth as manageable but demanding. "Global electricity demand from data centres is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today," Birol said when the IEA report was released. US Energy Secretary Jennifer Granholm, launching the Berkeley Lab report in December 2024, said, "We can meet this growth with clean energy." The department listed onsite generation and storage, reuse of retired coal sites, new rate structures and technologies such as advanced nuclear and long-duration storage among its strategies.

Takeaways

1. Small share, fast growth: data centres use about 1.5% of global electricity, but consumption is rising about 12% a year and could more than double by 2030.
2. AI is the main driver: accelerated servers are growing about three times faster than conventional ones in the IEA's base case.
3. Cooling efficiency matters: cooling can take anywhere from about 7% to over 30% of a facility's power, so design choices change the total.
4. The constraint is local: grid connections and transmission, not global supply, are the likeliest source of delays.
5. Read projections carefully: check units, scope, scenario and baseline year before repeating a number.