Data Centre Electricity Demand Rose 17% in 2025 as AI Load Surged

The number behind the post

In August, the International Energy Agency used its X account to push a message it first published in the spring. Global electricity use by data centres jumped in 2025, AI was the main reason, and demand is on track to almost double by 2030. The agency expects AI-focused facilities to triple their consumption over the same period.37 The figures come from Key Questions on Energy and AI, a follow-up released in April to the IEA's 2025 Energy and AI report.1

The headline numbers are clear. Data centre electricity demand rose 17% in 2025, while global electricity demand grew about 3%.1 Consumption at AI-focused facilities grew about 50% in a single year.2 The IEA puts total data centre use at roughly 485 TWh in 2025 and expects about 950 TWh in 2030, which would be around 3% of the world's electricity.2 Our World in Data, working from the IEA figures, splits the 2025 total into about 155 TWh for AI-focused sites and 330 TWh for everything else. By 2030, AI and non-AI facilities would each use roughly half.8

The main point is that the agency did not have to revise its outlook. It says 2025 growth matched what it projected a year earlier, and its central forecast stays close to the 2025 path.2 The surge was large, but it was expected.

Efficiency is real, and so is the rebound

The report pairs fast efficiency gains with rising total consumption. The IEA says the energy used per AI task has fallen by at least a factor of ten each year recently. A simple text query now uses less electricity than running a television for the same length of time.2 If every ordinary web search became a simple AI text query, the total would be under 4 TWh a year, which is less than 1% of current data centre use.2

The growth comes from heavier workloads. Video generation, reasoning models and agentic tasks can use hundreds or thousands of times more energy per query than basic text generation.2 The full report says directly that the Jevons paradox applies: efficiency gains drive wider adoption, so total AI electricity use still rises in the near term.6 The arithmetic also raises an awkward point. If text inference at huge scale is only a small slice of demand, most of the planned capacity must be meant for training, generative media and autonomous agents.6

Money is the real driver. Capital spending by five of the largest technology and data centre companies passed $400 billion in 2025 and is expected to grow another 75% in 2026.4 The IEA warns that spending at this scale now needs capital markets, not just company balance sheets. That makes the buildout sensitive to investor sentiment and financing conditions.2

Where the estimates diverge

The direction of travel is not in dispute. The size of the numbers is. Gartner puts 2025 data centre consumption at 447 TWh, below the IEA's figure. It expects 26% growth to 565 TWh in 2026 and more than 1,200 TWh by 2030, which is well above the IEA's central case.24 Energy Institute and S&P Global figures put 2025 near 788 TWh, but they include crypto mining, which the IEA leaves out.8 Our World in Data notes that some critics consider the IEA among the most conservative forecasters of AI demand.8

The opposite risk is also documented. A Congressional Research Service review cites an analysis arguing that projections overstate how many data centres will actually be built. It also cites a May 2026 investment-bank estimate that only 50–60% of capacity scheduled for the next year or two will arrive on time.23 NERC reports that many US grid operators lowered their summer 2026 demand forecasts because large loads were connecting more slowly than expected. NERC still projected peak demand about 11 GW higher than in 2025.23

The IEA's 2030 figure looks to me like a reasonable middle estimate. The agency itself says near-term growth is limited by physical bottlenecks, and it sees possible upside after 2030 if those bottlenecks ease.6 The biggest uncertainty is not demand. It is whether power can actually be delivered.

The grid is the binding constraint

On this point the coverage is unusually consistent. The IEA reports tighter supply chains over the past year for gas turbines, transformers and advanced chips. It also says the flood of data centre applications is overloading planning systems and slowing grid connections.1 Industry reporting describes transformer lead times of two to four years and transmission lines that can take more than a decade to permit and build.28 In Northern Virginia, the largest data centre market in the world, waits for new grid connections can exceed seven years.26

TrendForce gives the starkest forecast. It estimates global data centre power demand capacity at 161 GW in 2026 and expects grid supply to fall behind from this year onward. By 2030 it sees a gap of 268 GW, including more than 170 GW in the US.29 TrendForce notes that this figure leaves out on-site generation that operators build themselves.29

That on-site generation is now widespread. The IEA used satellite tracking to monitor US data centre projects with their own gas generation and found that many are still at an early stage.1 The CRS cites a separate analysis finding that only about 2.2% of announced behind-the-meter capacity was running by mid-2026.23 Running a data centre off the grid gets announced far more often than it gets built.

The local effects are already visible. One industry summary links PJM capacity prices to data centre growth and reports an increase of 833% between the 2024–25 and 2025–26 delivery years.21 In the UK, about 140 proposed data centre schemes could need 50 GW combined, which is more than the country's current peak demand.9 Regulators are reacting. In June 2026, FERC required grid operators to justify or revise their rules for connecting large loads. In July, it told NERC to develop reliability standards for computational loads.23

Flexibility and batteries

The IEA also makes a less obvious technical point. AI data centres can swing by more than half their rated load within a second.2 On-site gas plants struggle to handle swings like that, so on-site battery storage is becoming essential. The IEA says that, with the right incentives, these facilities could become useful to the grid rather than just a burden on it.1

Analysis in the US supports this. NERC found that ERCOT's summer 2026 peak forecast was 3.7 GW lower than the 2025 peak, partly because more data centres can now be curtailed during emergencies.23 An ITIF review argues that modest curtailment by data centres could let the grid absorb 76–126 GW of new demand without building more capacity.27 In my view, flexibility is the cheapest solution available and the least used. The IEA's recommendation that flexibility can speed up grid connections and protect affordability points the same way.4

Nuclear: contracts now, new reactors later

Nuclear is the most visible response, and the timelines are important. The IEA says the pipeline of conditional agreements between data centre operators and small modular reactor projects grew from about 25 GW at the end of 2024 to 45 GW at the end of 2025.6 The agency also says the first of those projects are not expected to come online until around 2030.6

Deal tracking shows two very different speeds. About 8.5 GW of output from existing US reactors is now contracted to Amazon, Meta, Google and Microsoft. Newer deals increasingly pay for uprates and licence renewals, such as Amazon's Calvert Cliffs agreement signed on September 30.12 Microsoft's restart of the reactor formerly known as Three Mile Island is on track for 2027.12 Meta's January deals cover 2.1 GW from Vistra's operating plants plus new-build agreements with Oklo and TerraPower.13 Kairos Power's Hermes 2 reactor for Google has been under construction since April and is targeting 2030.12 For now, no data centre runs on a new reactor built for it.12

Popular totals vary depending on what they count. Several trackers cite about 9.8 GW of hyperscaler nuclear commitments across 13 deals.1416 Those totals mix binding power purchase agreements with frameworks and options. The IEA's 45 GW figure counts conditional SMR offtake agreements.6 My conclusion is that nuclear supports the 2030s outlook but does little about the 2026–2029 shortfall. In that period, gas, batteries and existing reactors will carry most of the load.

Cooling: the heat problem

The IEA's figures on power density explain why cooling is now an energy issue as well as a facilities issue. AI server power density rose elevenfold between 2020 and 2025 and is expected to quadruple again by 2027. By then a single refrigerator-sized rack could draw as much peak power as 65 households.2 Gartner expects cooling and other infrastructure to use 195 TWh in 2026, about as much as all conventional servers.24

Liquid cooling is the industry's answer. One researcher estimates that air cooling adds about 30% to a facility's power draw, while liquid cooling brings that down to about 10%.35 Nvidia says its 45°C liquid-cooling design can run without chillers in suitable climates. It claims this cuts cooling water use from roughly 2.6 million gallons per megawatt per year to near zero.31 Microsoft and Oracle are making similar closed-loop claims for new builds. Critics point out that Microsoft has announced no programme to retrofit its existing data centres.3336

The skeptics have a fair point. Water used to generate the electricity and to make chips can double or triple a facility's total water footprint. That means on-site fixes address only about a quarter to a third of the problem.38 Water consumption is also still rising. Google reported a 34% increase in 2025, to 10.9 billion gallons.34 Better cooling helps, but the cleanliness of the power supply matters more for the overall environmental footprint.

The bottom line

The IEA is no longer describing a hypothetical. Demand rose sharply in 2025, as the agency forecast, and AI facilities are taking a growing share of it.2 The questions that matter for the rest of the decade are about delivery: transformers, interconnection queues, flexibility rules and reactor timelines. The agency also argues that, with the right policies and investment, rising demand does not have to push up power prices.1 Whether that holds will depend on how fast grids and regulators move, not on AI demand alone.