Data Centre Electricity: Reading the IEA Forecasts Without Fooling Yourself

Every month someone publishes a chart about how much electricity AI consumes, and every month the same chart gets quoted with the caveats stripped off. The IEA has now produced two substantial reports on the topic — Electricity 2024 (January 2024) and Energy and AI (April 2025) — that are far more careful than most of the coverage, and also far more quotable out of context than they should be. This article works through what they actually say, so you can use the numbers without inheriting the distortions.

Direct Answer: The IEA's central estimates are that data centres consumed roughly 415 TWh of electricity in 2024 (about 1.5% of global electricity demand), and that consumption is projected to reach around 945 TWh by 2030 in the Base Case — just under 3% of global demand — with a Lift-Off Case above 1,260 TWh. Crucially, the headline risk is local rather than global: data centres account for less than 10% of the growth in global electricity consumption to 2030, but in specific countries they dominate the growth picture. Both the "AI will consume all the electricity" framing and the "it's only 1.5%, relax" framing misread the report.

Two reports, different vintages, different scopes

The two IEA documents answer different questions and use different accounting boundaries. Getting that straight is most of the work.

Electricity 2024 is a market forecast to 2026 with a data-centre spotlight. Its framing is broad: "data centres, cryptocurrencies, and artificial intelligence (AI) consumed about 460 TWh of electricity worldwide in 2022, almost 2 % of total global electricity demand," with total data centre consumption potentially reaching "more than 1 000 TWh in 2026" — described as roughly equivalent to Japan's electricity consumption. Note the category there bundles data centres, crypto, and AI together; the number is not a data-centres-only figure.

Energy and AI (April 2025) is the dedicated report, and it separates the categories. It estimates "electricity consumption from data centres is estimated to amount to around 415 terawatt-hours (TWh), or about 1.5 % of global electricity consumption in 2024," and states that this "has grown at 12 % per year over the last five years." Cryptocurrencies are reported separately: "Cryptocurrencies – primarily from Bitcoin mining – consumed around 125 TWh in 2023 (0.5 % of global electricity)," and the report notes that "Cryptocurrencies and televisions are large energy users often associated with the ICT sector but are technically outside the sector scope according to definitions from the International Telecommunications Union."

So the 460 TWh (2022, combined scope) and the 415 TWh (2024, data centres only) are not a decline. They are different scopes measured a different number of years apart. Anyone presenting them as a series is constructing a trend line from incompatible units. This is the single most common error in secondary coverage of these numbers, and it is entirely avoidable.

The Base Case projection is that "electricity consumption from data centres rises to around 945 TWh by 2030, more than doubling from the 2024 level," reaching "just under 3 % of total global electricity consumption in 2030." The report is explicit about how to weight that:

In the wider context, a 3 % share in 2030 means that the data centre share in global electricity demand remains limited.

And on growth attribution: "Data centres still account for less than 10 % of the growth in global electricity consumption to 2030."

That last sentence is the one most quoted out of context, in both directions. Globally, data centres are a modest share of an electricity system whose demand is being driven by electrification, industrial output, cooling, and transport. A 3 % global share in 2030 does not describe a system being overwhelmed.

But the same report also warns, in the same section, that "the local impacts are far more pronounced." The global denominator is where the "relax" narrative comes from, and it is the wrong denominator for the questions that actually generate conflict.

The local picture is where the risk lives

The report's sub-national numbers are stark. For the United States: "data centres account for nearly half of electricity demand growth between now and 2030," and "In the United States, data centres account for nearly half of electricity demand growth between now and 2030. By the end of the decade, the country is set to consume more electricity for data centers..." The regional concentration is also extreme: "AI-focused data centres can draw as much electricity as power-intensive factories such as aluminium smelters, but they are much more geographically concentrated. Nearly half of data centre capacity in the United States is in five regional clusters."

The earlier Electricity 2024 report gave the European version of the same finding: it forecast that "electricity consumption from data centres in the European Union in 2026 will be 30 % higher than 2023 levels," and noted that "Ireland and Denmark alone make up 20 % of the expected increase in data centre electricity demand to 2026."

Both are the same structural point: electricity systems are built at national and regional granularity, and demand that is 3 % globally can be a third of marginal demand in a specific grid. A national transmission system cannot import electrons from a neighbouring country's surplus in the timescale a data centre build-out needs, so the relevant comparison is against local marginal demand, not global consumption.

Where the projections come from — and where they can go wrong

The report is unusually forthcoming about uncertainty, and readers should take it at its word rather than treating the Base Case as a prediction.

Its stated bases are: "Our Base Case projections for data centre electricity consumption are grounded in the latest industry expectations for server shipments. Three sensitivity cases (Lift-Off, High Efficiency and Headwinds) capture uncertainties in efficiency improvements in hardware and software, AI uptake and energy sector bottlenecks."

Base Case — ~945 TWh by 2030, just under 3 % of global demand.
Lift-Off Case — assumes "stronger AI uptake and limited local constraints on data centre buildout," reaching "over 1 260 TWh by 2030."
High Efficiency Case — "driven by energy savings in both software and hardware."
Headwinds — the case where local build-out constraints bind.

The range between ~945 TWh and ~1,260 TWh in 2030 — a spread of roughly a third — is the honest uncertainty band for a six-year projection of a fast-moving sector. There is also a definitional problem the report names directly: "combined with the lack of common definitions (Masanet, Lei and Koomey, 2024), this results in widely divergent estimates, even for historical consumption." If the historical number has a wide band, be careful with the forecast.

And the report is candid about the efficiency assumption underneath: "The operational efficiency gains of accelerated servers may be reaching their limit due to high utilisation and limited scope for further idle power reductions." A forecast that depends on hardware efficiency improving faster than demand grows is one assumption away from a different answer.

A concrete arithmetic sanity check

You can do a quick plausibility test on any claim you encounter. The report gives enough anchors: 415 TWh in 2024, ~12 % annual growth over the prior five years, and ~945 TWh in the Base Case by 2030.

Compound 415 TWh forward at 12 % for six years and you get roughly 820 TWh; the Base Case at ~945 TWh implies something closer to 15 % annual growth over the forecast window. Both are consistent with the report's own framing of "around 15 % per year" for 2024–2030 and "more than four times faster than the growth of total electricity consumption from all other sectors." So the Base Case embeds continued above-trend growth, and the gap between it and the Lift-Off Case is not noise — it is the difference between the 15 % path and a materially faster one.

For a scale anchor, the Electricity 2024 report gives global totals: global electricity demand rises from about 27,500 TWh in 2022 to about 30,700 TWh in 2026 in that forecast. A 1,000 TWh data-centre figure against a ~30,000 TWh global system is roughly 3 % — consistent with the later report's Base Case framing.

Emissions: the framing that changes the number

The Energy and AI report's emissions estimate depends entirely on which electricity mix you assume:

Emissions from electricity use by data centres grows from 180 million tonnes (Mt) today to 300 Mt in the Base Case by 2035, and up to 500 Mt in the Lift-Off Case.

It then puts this in context: "While these emissions remain below 1.5 % of the total energy sector emissions in this period, data centres are among the fastest growing sources of emissions." And it adds the counterfactual that is usually missing from headlines: "The widespread adoption of existing AI applications could lead to emissions reductions that are far larger than emissions from data centres – but also far smaller than what is needed to address climate change."

Note the divergence between a 2030 horizon for electricity and a 2035 horizon for emissions — the two are not aligned, and quoting them together as if they were a single forecast timeline is sloppy.

How to use these numbers

Practical rules that follow from the documents themselves:

Check the scope label before the number. "Data centres" alone, "data centres + crypto," and "ICT sector" are three different quantities. The IEA separates them deliberately.
Never chart the two reports' headline figures as one series. 460 TWh (2022, combined scope) and 415 TWh (2024, data centres only) are incomparable.
Quote the case, not just the Base Case. ~945 TWh by 2030 (Base) and >1,260 TWh (Lift-Off) are both IEA numbers, and so are the High Efficiency and Headwinds cases.
Lead with the local framing when the local framing is the point. "3 % of global electricity" and "nearly half of US electricity demand growth" are both true and both from the same report; they answer different questions.
Treat efficiency gains as an assumption with a named author. The report flags that accelerated-server efficiency gains "may be reaching their limit."

For the infrastructure-side view of the same problem — where requests physically terminate and how capacity behaves — see Technical SEO at Scale: URL Inventory, Capacity, and Observability and CDN Caching and Googlebot: Avoid Origin Overload.

References & Further Reading

International Energy Agency, Electricity 2024 — Analysis and forecast to 2026, January 2024 (revised version January and May 2024). Supports the 460 TWh 2022 combined-scope figure and its "almost 2 %" share, the "more than 1 000 TWh in 2026" projection and Japan comparison, the EU +30 % versus 2023 forecast, the Ireland and Denmark 20 % share of the EU increase, the global demand figures from roughly 27,500 TWh (2022) to 30,700 TWh (2026), and the count of more than 8,000 data centres globally with about 33 % in the United States, 16 % in Europe and close to 10 % in China.
International Energy Agency, Energy and AI, April 2025. Supports the ~415 TWh 2024 data-centre estimate and 1.5 % share, the 12 % annual growth since 2017, the Base Case ~945 TWh by 2030 at just under 3 %, the Lift-Off Case above 1,260 TWh, the High Efficiency and Headwinds case descriptions, the United States 45 % / China 25 % / Europe 15 % shares of 2024 data-centre consumption, the "less than 10 % of the growth" global attribution, the "nearly half" US demand-growth figure, the five-cluster US concentration, the ITU-T scope treatment of cryptocurrencies with 125 TWh in 2023, the ~15 % annual growth in the Base Case, the emissions figures of 180 Mt today rising to 300 Mt (Base) or 500 Mt (Lift-Off) by 2035, the acknowledgement of divergent estimates from a lack of common definitions citing Masanet, Lei and Koomey (2024), and the observation that accelerated-server efficiency gains may be reaching their limit.