IEA Scenario: Data Center Power Use Could Double to 1,000 TWh by 2030 – But It Is Not a Forecast

The International Energy Agency (IEA) released the World Energy Outlook 2026 on 19 November 2026, and the headline number circulated quickly in coverage: under the agency's Stated Policies Scenario, global data center electricity…

File photo: Joint IAEA – IEA (International Energy Agency) Seminar on the World Energy Outlook 2011. IAEA Vienna, Austria, 24 November 2011 Copyright: <a href=" http://www.iaea.org/NewsCenter/Multimedia/Imagebank/index.jsp " rel="nofollow">IAEA Imagebank</a> Photo Credit: Dean Calma / IAEA
File photo: IAEA Imagebank, CC BY-SA 2.0 · source This adaptation is licensed under CC BY-SA 2.0.

IEA Scenario: Data Center Power Use Could Double to 1,000 TWh by 2030 – But It Is Not a Forecast

The International Energy Agency (IEA) released the World Energy Outlook 2026 on 19 November 2026, and the headline number circulated quickly in coverage: under the agency's Stated Policies Scenario, global data center electricity consumption doubles to 1,000 terawatt-hours by 2030, with AI-driven investment as a central driver. But the figures come with two important caveats that are easy to lose in citation. First, the report itself – as relayed through write-ups from Unite.AI and Anadolu Agency – stresses that none of the scenarios are forecasts; the 2030 figure is a conditional pathway, not a prediction. Second, the report points out that the real bottleneck for AI expansion is increasingly not the computing power itself, but the power grids, the supply chains and, in some cases, local public opposition.

This article goes through what the numbers actually say, what they assume, and why the scenario framing matters for how they should be used.

What the IEA actually published

The IEA, headquartered in Paris, released the World Energy Outlook 2026 on 19 November 2026, according to Unite.AI's write-up (Unite.AI). Anadolu Agency and Electrek covered the same report the same week, and the three write-ups agree on the core scenario: in the Stated Policies Scenario – the scenario built on policy measures that are formally proposed but not yet adopted – global data center electricity consumption doubles to around 1,000 terawatt-hours by 2030, with AI a substantial share of the increase (Anadolu Agency; Electrek).

It is worth being precise about the sources: none of the available write-ups are first-hand documentation from the IEA itself. All the figures in this article come from three independent media write-ups of the report, and the figures should be read as those outlets' rendition of the IEA's work, not as verified against the report's original text. One figure – that data centers' carbon emissions will rise 70 percent by 2030 – appears only in Anadolu Agency and cannot be cross-checked against the other two write-ups; it is thus a single-source figure and should be cited with that caveat.

Why scenarios are not forecasts

The most important framing for all the numbers below is methodological: the IEA's World Energy Outlook is constructed as alternative development pathways, and the report states, according to Unite.AI's write-up, that none of the scenarios is a forecast. The 1,000 TWh figure for 2030 is a conditional pathway – it materializes only if the scenario's assumptions hold.

The difference between the scenarios is also concrete. According to Unite.AI, the Stated Policies Scenario incorporates policy measures that are formally proposed but not yet adopted, while the Current Policies Scenario looks at today's policy frameworks without further policy. That means even the two main scenarios represent different degrees of political ambition and implementation – and in practice the outcome may land somewhere between them, or entirely outside them, if key assumptions such as AI investment levels, grid access and political will develop differently.

For the reader this has a practical implication: "1,000 TWh by 2030" is a useful anchor for discussions of order of magnitude, but it is analytically sloppy to cite it as a prediction. The report instead tells us something else – and according to the write-ups something more interesting: what would have to happen for such a figure to be reached, and where growth now meets resistance.

The numbers: the grids that carry the growth

It is behind the scenes of the 1,000 TWh figure that the report's most concrete findings sit, according to Unite.AI's walkthrough:

  • Data center capacity approaches three times today's level by 2035, as AI and digital services expand.
  • 26 million kilometers of new power lines must be built toward 2035 in the Stated Policies Scenario – while around 16 million kilometers of existing lines must be replaced or upgraded.
  • Grid investment rises 75 percent over the next decade – even in the Current Policies Scenario, that is, without further political ambition. In other words: grid costs rise sharply regardless of which political direction is chosen, because the investment need is driven by demand, not by the aggressiveness of policy.
  • Peak demand for electric power is calculated to rise 40–50 percent by 2035. This is a systemic figure that matters more than it looks: it is not only the total amount of energy that constrains the system, but how much power must be deliverable in the same moment. Data center load is continuous and concentrated, and peak load is where power grids often first crack.

The point that recurs across the write-ups is that these figures are about infrastructure, not about AI itself. Power grids, supply chains and, in some cases, public acceptance are described as emerging constraints around the investment boom – and Electrek's rendition is explicit that the build-out is "already running into" these constraints.

Historical pace and contractual-economic adjustments

Another set of figures, relayed by Unite.AI as a finding from an earlier IEA analysis (which is not identified in the write-ups, and whose identity therefore cannot be stated here), puts the scenario in context: data centers' electricity demand grew 17 percent in 2025, against 3 percent growth in global electricity consumption overall. That is a gap that illustrates why data centers now dominate the debate on power planning – they are growing in a different league from the rest of the system.

The same earlier analysis stated, according to Unite.AI, that the volume of conditional agreements in which data center operators commit to buying power from so-called small modular reactors (SMRs) grew from 25 gigawatts at the end of 2024 to 45 gigawatts when the analysis was published. That is nearly a doubling in a short time, but it is worth noting that these are conditional agreements – that is, commitments on paper, not completed built capacity. The growth in agreement volume points to how hard technology companies are trying to secure their own supply outside the public grid, but it says something about intent, not delivered power. The write-ups give no date or title for the analysis beyond this, so it is uncertain exactly which IEA publication the figures come from.

The write-ups also mention a mechanical point relevant to the debate about AI's energy efficiency: efficiency gains per individual AI task, according to the rendition, are "eaten up" by increased use at the system level. That is a well-known dynamic (often called the rebound effect): when unit costs fall, usage rises, and total consumption can increase even if each individual operation becomes more efficient. It is the same logic that means pointing out that models are getting cheaper per token is not enough – the total ledger is determined by the volume of use.

Emissions and regional spread

Emissions are where the sources diverge the most. Anadolu Agency is the sole source for the claim that the IEA projects data centers' carbon emissions rising 70 percent by 2030, driven by growing AI demand. The figure does not appear in Unite.AI or Electrek in the available passages, and must therefore be cited as a single-source rendition.

More interesting than the total is the regional spread in energy mix that Anadolu Agency relays:

  • In the United States, natural gas is expected to account for around 60 percent of data centers' electricity consumption by 2030.
  • In China, coal is expected to account for roughly half of the electricity consumption of AI-driven data centers.
  • In the EU, renewable energy is expected to cover around 70 percent of data centers' power needs by the end of the decade.

This spread matters analytically because it shows that "AI's climate footprint" is not one global number, but three very different stories. In the US, AI growth is in practice tied to gas expansion. In China, a large share of AI growth ties to coal power, with the emissions consequences that entails. In the EU, by contrast, a different picture emerges, where data centers' power needs are largely expected to be met by renewables. One and the same technology – AI training and inference – thus gets very different emissions profiles depending on where it physically runs, and the emissions trajectory will in practice be the sum of regional decisions about siting and power purchasing, not a global technological given.

The constraints: where the bottleneck actually sits

Perhaps the most important finding in the report, as the write-ups relay it, is not the consumption figures but the diagnosis of where growth now meets resistance. Three constraints are named explicitly by both Electrek and Unite.AI: power grids, supply chains and public acceptance.

  • The grids: The figures above (26 million km of new lines, 16 million km needing upgrades, 75 percent higher grid investment) point to implementation as a gigantic logistical task. Grid companies must both build new and maintain and replace existing lines, and peak load grows faster than average consumption.
  • The supply chains: Electrek and Unite.AI both name supply chains as a growing constraint around the investment boom. The write-ups do not go into detail about which specific components are involved – in principle it could range from transformers and cables to turbines and semiconductors – so it should be kept at the level of abstraction the sources allow.
  • Public acceptance: Unite.AI and Electrek point out that local opposition in some cases is a real constraint. It is a politically realistic observation: data centers need physical space, water and grid connections, and such projects in practice encounter local politics.

The overall impression from the write-ups is that the report's focus has shifted: from "how much power does AI use?" toward "can the power and grid sectors deliver the pace that AI investment assumes?". In this telling, the bottleneck is no longer computing power.

Open questions

It is worth being honest about what the available sources cannot answer, because this affects how solid each claim is:

  • No first-hand source. None of the write-ups are the IEA's own report, so all the figures are media renditions. For central uses, the figures should be verified against the report's original text.
  • The +70 percent emissions figure is single-sourced. It comes only from Anadolu Agency and cannot be cross-checked against the other write-ups in the available material.
  • The earlier IEA analysis is not identified. The figures on 17 percent growth in 2025 and the SMR agreements from 25 to 45 gigawatts are attributed to "an earlier IEA analysis" by Unite.AI, without the write-ups stating which one. It is therefore not possible to look it up or date it precisely.
  • Concrete mechanisms are not specified. The write-ups mention supply chain constraints and local opposition, but do not go into which specific components or which places are hardest hit.

What this means

The World Energy Outlook 2026 gives the debate about AI's energy footprint a concrete, official numerical anchor: 1,000 TWh by 2030 in the Stated Policies Scenario, confirmed by three independent write-ups. But the report's own framing – that these are scenarios, not forecasts – and its diagnosis of where growth now meets resistance should travel with the figure when it is cited. The big story in the report's numbers is not only that AI could double data centers' power consumption, but that what determines whether it happens lies in the pace of grid construction, the capacity of supply chains and the willingness of local communities to host the infrastructure – conditions no AI model can optimize its way around.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.