AI Data Center Power Use Could Hit 295 TWh by 2030, Study Finds

A new estimate, and what it counts

A peer-reviewed study in Communications Sustainability puts a number on AI's electricity appetite. Data centers run by six large technology firms could use between 239 and 295 terawatt-hours a year by 2030, up from about 118 TWh in 2024. That would be roughly 1% of global electricity demand.1 The authors are Danbo Chen, Zijun Zhou, Yongyang Cai and Lei Chen, and the paper was published on September 21, 2026.4 The six firms are Amazon, Microsoft, Google, Meta, Oracle and Apple. The authors estimate that together they account for about 70–75% of global hyperscale and cloud-linked data-center electricity demand.1

The method is unusual. The team used large language models to pull siting and expansion signals out of financial filings, sustainability reports and corporate announcements. It then fed those signals into scenario-based electricity projections.1 It modeled three growth paths for new AI data-center capacity: 15%, 25% and 35% a year. In all three, baseline data-center consumption grows 10% a year while the share of AI-intensive workloads rises.1 The resulting compound annual growth rate of roughly 13–17% is, by the authors' comparison, like adding a medium-sized national electricity market every two years, or the annual use of about 27 million US households.4

The headline needs a careful read. The 1% figure is not a full count of AI electricity use worldwide. It covers the whole data-center footprint of six companies, adjusted for the growing share of AI work. It is not a census of every AI chip on the planet.1 Coverage of the paper has mostly repeated the framing as given. One trade summary did say the projection concerns data centers "run by six leading technology firms".6

How it compares with the IEA's numbers

The study sits comfortably inside existing forecasts. That is its strength, and also a sign of how modest it is. The International Energy Agency's base case has global data-center electricity use roughly doubling to about 945 TWh by 2030, just under 3% of world consumption.2 The authors cross-checked against that figure. They assumed hyperscalers use 70% of data-center power and that these six firms make up 40% of hyperscale activity, which gives an IEA-implied 2030 figure of about 265 TWh, close to the middle of their range.1

The figures start to diverge on the definition of "AI". Our World in Data, summarizing IEA figures, reports that AI-focused data centers used about 155 TWh in 2025 and could reach 465 TWh by 2030. That would put them at about half of all data-center demand.10 On that measure, AI-dedicated facilities alone would use well over the 1% share the new paper reports. Brookings, citing other estimates, says data centers could account for 3% to 4% of global electricity by 2030, with some forecasts as high as 9%.7 Our World in Data also notes that some analysts consider the IEA among the more conservative forecasters.10

Our reading is that the 1% figure is best treated as a floor for how much power the biggest AI builders will use. It is not a ceiling for AI overall. Smaller cloud providers, colocation operators, telecoms and national operators are excluded, mainly because they disclose less. The authors also acknowledge that regions such as Africa and Australia may be underrepresented.1

Why a small global share still causes problems

One percent sounds small, and in global terms it is. Carbon Brief's analysis of IEA data finds data centers account for only a little over 1% of electricity demand today. By 2030 their growth would make up about 8% of the total increase in world demand.9 The IEA itself says data-center growth is less than 10% of global demand growth through 2030 in its base case.2

The study's more important finding is about where the load lands. More than 90% of projected computing capacity is in North America, Western Europe and Asia-Pacific.1 Fewer than ten regions account for nearly two-thirds of projected demand. Oregon, Virginia, Iowa, Ohio and Ireland each exceed 15 TWh a year in the central scenario.4 The authors find that Oregon, Ireland and Iowa face more pressure relative to the size of their power systems, while a larger system such as Texas absorbs new load more easily.4 Brookings reports similar concentration elsewhere: data centers could take 42% of local demand in Frankfurt and nearly 80% in Dublin.7

The authors also flag a limit that matters a lot. Their demand-pressure index may understate local stress where transmission is constrained, because connection queues can block new large loads even when there is enough supply overall.4

The grid is the real constraint

Industry reporting from 2026 suggests that access to the grid, more than total generation, is what slows AI expansion. According to Lawrence Berkeley National Laboratory's 2026 Queued Up analysis, about 8,200 projects seeking interconnection represent 1,312 GW of generation. Projects completed in 2025 took a median of more than five years from queue to operation.14 Only 13% of interconnection requests filed between 2000 and 2019 were operating by the end of 2024.17

The pressure is showing up in prices and policy. In the PJM market, capacity prices rose from $28.92 to $329.17 per megawatt-day in about two years.13 ERCOT's large-load interconnection queue grew from 63 GW to 226 GW in a single year. Power transformers now take an average of 128 weeks to arrive.17 Regulators are responding. FERC has issued show-cause orders to six regional grid operators over how large loads connect.14 Virginia has created a large-user rate class that requires data centers to pay for minimum shares of their contracted demand.17

The market-research firm TrendForce goes further. It estimates global data-center power demand capacity at 161 GW in 2026 and 490.7 GW by 2030. Against that, it estimates only 222.6 GW of grid capacity available to data centers, a gap of 268 GW.18 TrendForce notes that on-site generation could close part of that gap, and that this generation is not counted in its grid figure.18 These are capacity numbers, not annual energy, so they cannot be compared directly with the study's terawatt-hours. Still, they point the same way: demand is growing faster than the grid can deliver it.

Nuclear: plenty of deals, little new power yet

Hyperscalers have responded by signing up nuclear power, and how fast it arrives is often overstated. Nuclear News Network's tracker, as of October 2026, finds that about 8.5 GW of US reactor output is contracted to Amazon, Meta, Google and Microsoft. Most of it comes from existing plants, and newer deals increasingly fund power uprates and license renewals.33 Examples include Amazon's power-purchase agreement for up to 1,920 MW from Talen's Susquehanna plant and Meta's 20-year agreements with Vistra in Ohio.33

New reactors are still mostly plans. Only one data-center-linked project, Kairos Power's Hermes 2 for Google via TVA, is under construction, with first power expected in 2030.33 Conditional agreements between data centers and small modular reactor developers reached 45 GW. That figure is not a tally of construction.33 Meta's commitment of up to 6.6 GW is the largest single pledge, but it is aimed at 2032 to 2035.32 Trackers also disagree on its details: one puts the TerraPower share at 2.8 GW across eight Natrium units, another at 4.0 GW.3238

The restart of Microsoft's Three Mile Island Unit 1, renamed Crane, is the nearest new source. One report says a June 2026 FERC transmission waiver moved its timeline to the second half of 2027.32 Another analysis estimated in July 2026 that only 19.6% of committed hyperscaler nuclear capacity was actually delivering power.36

The timing problem is plain. The study's demand ramp runs through 2030, but most new nuclear arrives in the 2030s. For the rest of this decade, existing reactors, gas and the grid will have to carry the load.

Cooling: water savings, limited electricity savings

Cooling is the one part of the problem that operators can change quickly. The study counts cooling and power-distribution losses as overhead in its facility-level estimates. It says better cooling designs or more efficient chips could push energy intensity below its assumed 1–3% annual efficiency gain.1

The move to liquid cooling is now mainstream. IDC forecasts that 90% of large, performance-intensive computing deployments will use direct liquid cooling by 2027.24 Nvidia says its newest server designs are fully liquid-cooled with a closed loop that needs no new water. It estimates a 50-megawatt facility could save more than $4 million a year in cooling energy and water costs.27 Industry benchmarks put direct-to-chip cooling at a typical power usage effectiveness (PUE) of 1.15–1.30, and immersion as low as 1.03–1.08. Conventional air-cooled halls typically run between about 1.3 and 1.95.30

The limit is that cooling only reduces overhead. It does nothing about the power the chips themselves draw, which keeps rising with demand. Water also complicates the picture. Bluefield Research estimates that about 72% of data-center-related water use by 2030 will happen at power plants, not on site.29 Cutting water use inside the building does not remove the water and grid costs of the electricity it buys.

What it adds up to

The 1% headline is both reassuring and misleading. The share of world electricity is modest. But the load is concentrated in a few grids, and the infrastructure meant to serve it, including transmission, transformers and new reactors, runs years behind the buildout. By the study's own description, AI infrastructure is becoming a structural part of how power systems work.4 The practical tests over the next few years will be interconnection queues, capacity prices and whether restarts like Crane come online on schedule.