The IEA Expects Data Centres to Use About 950 TWh of Electricity Worldwide in 2030, Berkeley Lab Puts the United States Alone at 649 TWh, and the Two Disagree on the US by More Than 200 TWh

The International Energy Agency (IEA) says data centres used 485 terawatt-hours (TWh) of electricity in 2025 and will use about 950 TWh in 2030, around 3% of world demand. A terawatt-hour is a billion kilowatt-hours. The figure is the central case in the agency’s April 2026 report Key Questions on Energy and AI, and it sits almost exactly where its 2025 report had it, at 945 TWh. For the United States alone, the two main public forecasts have moved apart: Lawrence Berkeley National Laboratory’s June 2026 report gives 649 TWh for 2030, and the IEA’s data table gives 426 TWh.

What the IEA counts worldwide

The IEA’s 2025 report put 2024 consumption at 415 TWh, 1.5% of the world’s electricity, with the United States at 45%, China at 25% and Europe at 15%. The April 2026 report says consumption grew 17% in 2025. It says AI-focused data centres, the ones packed with accelerator chips, grew 50%. In the agency’s Base Case that group triples to about 465 TWh by 2030.

Data centres still account for slightly less than 10% of global electricity demand growth to 2030 in that case, the IEA says, although some regions are much more affected. For 2035 the 2025 report gave a Base Case of about 1,200 TWh and a range of 700 to 1,700 TWh across its cases.

Energy per AI task has dropped by at least an order of magnitude a year in recent years, the IEA reports. Our piece on what one prompt costs covers the per-prompt figures and the water.

Two American estimates more than 200 TWh apart

Berkeley Lab’s first report, in December 2024, counted 176 TWh in US data centres in 2023, which was 4.4% of US electricity. It projected 325 to 580 TWh in 2028, or 6.7% to 12.0%.

The June 2026 update, by Sarah Smith, Arman Shehabi and six co-authors, revises 2024 to 192 TWh, or 4.7%. Its Reference Case is 464 TWh for 2028 and 649 TWh for 2030, which it puts at 11.8% of the US electricity that the North American Electric Reliability Corporation forecasts for that year. The scenario range for 2030 runs from 521 to 843 TWh, or 9.5% to 15.3%. Reference Case use rises 22% in 2025 and 29% in 2026. If NERC’s forecast includes data centres, as the lab assumes, they account for 33% of US load growth between 2024 and 2030.

The IEA’s data table lists the United States at 183 TWh in 2024, close to the lab’s 192, and at 426 TWh in its 2030 Base Case. The Berkeley Lab report names the gap: it calls the IEA’s 2030 figure, which it gives as approximately 430 TWh, “substantially lower” than its own, describes the IEA’s method as broadly consistent with its own, and blames the divergence on assumptions about server and accelerator shipments.

The IEA’s April report, two months earlier, says its Base Case “continues to provide a balanced assessment” of the momentum in AI and the financial and physical frictions that limit near-term growth.

What the modellers say they cannot see

The Berkeley Lab authors say they do not model breakthroughs, market shifts or changes in supply chains, regulation or grid availability, because doing so “would require speculation”. Their 521 to 843 TWh bounds stack every sensitivity at once, which they call a “stress test”, and they say it is unlikely that all factors would hit their extremes together. The biggest single swing is how busy AI servers are when answering queries: raising their idle power and use lifts the 2030 figure to 782 TWh, more than 20% above the Reference Case.

Converting energy into power plant capacity is a second gap. At an assumed 50% use of connection capacity, 649 TWh equals 148 gigawatts (GW) of grid connections in 2030. The lab says current use of that capacity “is not well documented”.

The IEA lists uncertainty over the pace of AI adoption and over model and hardware efficiency. It says bottlenecks across the value chain make aggressive near-term scenarios less likely, and that clearing them could produce a higher case after 2030.

Who is signing for the power

Meta announced nuclear agreements on 9 January 2026 that it says support up to 6.6 GW by 2035. They include 20-year contracts with Vistra for more than 2.1 GW from two Ohio plants plus uprates, meaning equipment upgrades that raise output, totalling 433 MW. They also include an Oklo campus of up to 1.2 GW in Pike County, Ohio, and funding for two TerraPower Natrium units of up to 690 MW, with delivery as early as 2032.

On 6 October 2026 Google and Constellation Energy announced two agreements. A 20-year contract funds uprates adding 890 MW at 11 Constellation nuclear units in Illinois, Pennsylvania and New Jersey, with the first due by 2028. A 15-year supply agreement covers a further 2,700 MW from Constellation’s existing PJM fleet. The release reports more than $4.3 billion in new investment by Constellation and does not say what Google pays. Our daily roundup on 7 October carried the announcement.

Constellation says Crane Clean Energy Center, the former Three Mile Island Unit 1, will restore 835 MW and expects it ready for service in 2027. The NRC issued the final environmental assessment on 18 September 2026 and lists licensing requests as still under review. The National describes the restart as part of a deal involving Microsoft.

At Palisades in Michigan, a fuel assembly tipped over during loading at the end of August, The National reports. An NRC spokeswoman told the paper there was “no radiological release”, and the paper gives no new restart date.

Gas has its own queue. The IEA reports a 70% jump in gas turbine orders in 2025 and cites Wood Mackenzie for delivery times of around five years for turbines and two to three years for transformers. It estimates 15 to 27 GW of onsite gas could power data centres by 2030, mostly in the United States.

The grid operators’ waiting lists

PJM, the grid operator for 67 million people in 13 states, buys capacity through an annual auction held about 23 months ahead of the delivery year. The price per megawatt-day was $28.92 for 2024/25, $269.92 for 2025/26 and $329.17 for 2026/27. The 2027/28 auction on 17 December 2025 cleared at the cap of $333.44 and left PJM 6,623 MW short of its reliability requirement. PJM said its peak load forecast was 5,250 MW above the one used for the previous auction, nearly 5,100 MW of it data centres. Executive vice president Stu Bresler said the auction “leaves no doubt that data centers’ demand for electricity continues to far outstrip new supply”.

The 2028/29 auction, reported on 14 July 2026, cleared at its $325.00 cap. Total procurement of 149,181.6 MW was 6,831.3 MW below the requirement, a reserve margin of 14.7% against a 20% target. PJM raised the requirement from 152,400 MW to 156,013 MW on a 1,374.5 MW forecast load increase “largely due to additional Large Loads”. PJM says a shortfall means slimmer reserves, not necessarily that it cannot serve customers. In December 2025 PJM said about 57 GW of generation projects had completed its study process, though many were slowed by local opposition, permitting, supply chains or financing.

ERCOT, the Texas grid operator, said in a presentation dated 29 July 2026 that it was tracking about 474 GW of large loads seeking to connect as of June, roughly 90% of them data centres. Its queue of generation projects stood at 462,785 MW on 30 June. About 205 GW of the load requests qualified for its first study batch, called Batch Zero, because they already had studies. After Governor Greg Abbott’s 3 August directive, ERCOT’s 9 September notice says it is verifying and auditing data centre projects, and a load that fails is excluded from Batch Zero.

India’s 1.57 GW and the 17 GW figure

The Press Information Bureau, the Indian government’s information agency, published a note on 14 September 2026. It says installed data centre power capacity was around 375 MW in 2020 and 1.57 GW as of August 2026, and is projected to reach nearly 8 GW by 2030. It attributes to the Central Electricity Authority a figure of 17 GW of demand by 2031 to 2032, and does not describe how the capacity figures were counted. A gigawatt running all year uses 8.76 TWh, so these capacity figures cannot be compared directly with the energy totals above.

Cushman and Wakefield, a property consultancy, counted 1,789 MW operating in India’s top seven markets in the first half of 2026. Mumbai had 890 MW, Chennai 214 MW and Hyderabad 151 MW, with pipelines through 2030 of 1,726 MW, 368 MW and 710 MW. The firm’s total pipeline is 3,860 MW. Its city rows add up to 1,755 MW operating, and its note does not explain the gap to 1,789 MW.

Google said on 14 October 2025 that it would invest about $15 billion over 2026 to 2030 in a Visakhapatnam AI hub with AdaniConneX and Airtel, including a data centre campus “capable of gigawatt-scale compute capacity”. Google said it would work with partners on transmission lines, clean generation and storage in Andhra Pradesh.

Reliance chairman Mukesh Ambani told shareholders on 29 August 2025 that Reliance Intelligence would build “gigawatt-scale, AI-ready data centres, powered by green energy” and that work had begun at Jamnagar. On 10 June 2026 Reliance announced a 168 MW data centre there for Meta, to be delivered within two years with an option to scale, powered by renewable energy and cooled with desalinated seawater.