AI Data Center Energy Demand Hits 565 TWh as Grid Queues Swell

The short version

Artificial intelligence has moved the data center industry's main constraint away from chips and real estate. The limit now is the electrical grid, and recent forecasts and regulatory moves show how fast that happened. Gartner expects global data center electricity consumption to reach 565 terawatt-hours (TWh) in 2026, up from 447 TWh in 2025, a 26% jump in one year.1 Power demand, measured as capacity, is projected to rise from 104 gigawatts (GW) to 132 GW this year and to about 290 GW by 2030.1

AI accounts for most of that growth. Gartner estimates that AI-optimized servers will use 175 TWh in 2026, an 84% increase that makes up 31% of all data center electricity. It expects AI servers to use more power than conventional servers in 2027.1 Conventional server consumption, by contrast, is barely moving: 193 TWh in 2025 and 195 TWh in 2026.1 Put simply, nearly all of the new demand comes from AI.

Where the forecasts agree, and where they don't

The main forecasters point the same way but differ on how steep the climb is. The International Energy Agency (IEA) puts global data center use at about 485 TWh in 2025 and roughly 950 TWh by 2030, so its baseline starts higher than Gartner's but grows more slowly.8 Gartner has warned that consumption could pass 1,200 TWh by 2030 if current growth continues, which is above the IEA's base case.7

One number needs a warning label. Several commentary sites report that data centers will use more than 1,000 TWh in 2026, comparing the figure to Japan's entire electricity use. Some credit the IEA and others credit Gartner.2610 That claim doesn't match either organization's published numbers. Gartner's own 2026 forecast is 565 TWh, and the IEA's 950 TWh figure is for 2030.18 One of those sites even notes that its 2026 number would already exceed the IEA's 2030 base case and calls this an upward revision.10 The more likely explanation is that the figure got garbled as it was passed along. Readers should use the 565 TWh estimate as the best guide to this year and treat the 1,000 TWh claim as a 2030 scenario.

The U.S. forecasts are steeper. Goldman Sachs Research expects U.S. data center power demand to rise from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027.79 By its estimate, that would raise data centers' share of U.S. peak summer demand from 4.1% to 5.3% within a year.10 Bank of America analysts expect data centers to add about 125 GW of U.S. load between 2026 and 2030, and they note that large gas turbines are essentially sold out through 2030.7

The grid is the bottleneck

Demand forecasts only matter if power can actually be delivered, and that is where things are tightest. Lawrence Berkeley National Laboratory's 2026 Queued Up analysis counts about 8,200 active projects waiting to connect to the grid, representing 1,312 GW of generation and 749 GW of storage. Projects that came online in 2025 waited a median of more than five years.25 Historically, only about 13% of queued projects reach commercial operation, and roughly three-quarters drop out.2530

It matters what those queues measure. They track power plants and batteries trying to connect, not data centers trying to draw power, so a large queue doesn't mean power is available to new facilities.25 Data centers wait in their own lines for load connections, and those lines are growing even faster.

Texas shows this most clearly, though the reported numbers have changed so quickly that they conflict depending on the date. ERCOT's large-load queue grew from 63 GW to 226 GW between late 2024 and late 2025.23 Later figures put it at about 410 GW in April 202629, more than 438 GW by June28, and 474 GW by September, with about 90% of requests coming from data centers.26 ERCOT received 198 GW of large-load applications in the first quarter of 2026 alone.29 Many of those requests are speculative and will never be built. Still, when applications run this far ahead of what the grid can deliver, it shows how far demand has outrun physical capacity.

Regulators are now responding. In June, the Public Utility Commission of Texas approved rules letting ERCOT review a first group of large-load requests, called "Batch Zero."28 The same day, the Federal Energy Regulatory Commission (FERC) found that most grid operators' rules for connecting large loads were inadequate and told them to write region-specific interconnection rules for data centers.28 One report says Governor Greg Abbott paused new Texas data center grid connections in August 2026 pending an audit.26 Only one outlet has reported this so far, so it should be read with caution.

Other states are taking different approaches. Ohio's AEP adopted terms that require data centers to post collateral and pay minimum bills. Virginia created a large-user rate class that charges data centers for a minimum share of their contracted demand. West Virginia created "certified microgrid districts" that pair data centers with their own on-site generation, letting them skip the queue entirely.23

Hardware lead times and ratepayer costs

Even a project that wins grid approval can stall waiting for equipment. Power transformers now take about 128 weeks to arrive, and generator step-up units take about 144 weeks.23 Another analysis gives a range of 90 to 130 weeks for transformers and breakers. It also warns that developers who wait for a signed interconnection agreement before ordering equipment end up running two multi-year timelines back to back.22 Of the 16 GW of U.S. data center capacity targeted for 2026, only 5 GW had entered active construction, and 30% to 50% of the rest is expected to slip to 2027 or later.23

The most crowded markets are close to full. Northern Virginia's colocation vacancy rate was reported at 0.72%. A Google official testified that one utility estimated it would take 12 years to evaluate a single interconnection request.23 In PJM, the grid operator covering Virginia and much of the Mid-Atlantic, the largest AI campuses face waits approaching seven years.2629 One analysis concludes that a 2026 announcement with a 2028 target date in PJM is more likely to be a 2031 facility.29

Households are also paying for the buildout. In PJM, data center growth is blamed for a $9.3 billion increase in capacity costs for the 2025–26 delivery year, or about $16 to $18 a month on residential bills in states such as Ohio and Maryland. Capacity auction prices rose from $28.92 to $269.92 per megawatt-day in one year.26

As a result, big developers are increasingly paying for their own power. Microsoft's planned 2-GW campus in Pecos, Texas, will come with company-funded energy infrastructure.28 xAI's Colossus facility in Memphis opened with about 150 MW of computing load but only about 8 MW of initial utility connection, and it made up the gap with generation on site, outside the grid connection.30

Cooling: real progress, partial answer

The same chips that drive power demand are also changing how data centers are cooled. Nvidia's GB200 NVL72 racks give off more than 120 kilowatts of heat each, and Nvidia requires liquid cooling for them. If coolant falls out of spec, the chips throttle automatically and performance can drop by as much as 60%.17 One industry guide says 2026 is the year liquid cooling stopped being a supercomputing specialty and became the standard way to cool frontier AI systems.13

The biggest announcement came in June. Nvidia released a reference design called DSX that sends coolant to every chip at 45°C. Because the coolant is that warm, outdoor dry coolers can shed the heat without evaporating water, and the company says this can cut cooling water from about 2.6 million gallons per megawatt per year to near zero in favorable climates.11 The coolant is a 75/25 mix of water and propylene glycol that is filled once and recirculated in a closed loop.1618 Nvidia estimates that a 50-megawatt facility could save more than $4 million a year on cooling energy and water.14 University of Chicago professor Andrew Chien explained why the warm temperature matters: hotter coolant can reject heat outdoors without running air conditioning. He added that truly zero water use is unrealistic.14

The disagreement in the coverage is about what Nvidia's claim actually covers. Nvidia's chief sustainability officer said the water problem for data centers is "largely solved."12 TechCrunch countered that water used to generate electricity and make chips can double or triple a facility's total footprint, so the design addresses only a quarter to a third of the problem.12 Tech Times went further. Citing a Xylem and Global Water Intelligence analysis, it reported that on-site cooling accounts for only about 4% of the extra water AI will need by 2050, compared with roughly 54% for power generation and 42% for chipmaking.18 The IEA expects natural gas and coal to supply more than 40% of the new electricity data centers need through 2030.12

The engineering advance is real. Vantage's closed-loop campus in Port Washington, Wisconsin, is designed to withdraw about 22,000 gallons a day at peak, compared with about 5 million for a similar facility cooled by evaporation.20 But the main issue is power, not water. Cooling and other infrastructure are forecast to use 195 TWh this year, up 22.6%.1 Better cooling cuts that overhead and frees up grid capacity for computing. It doesn't reduce the amount of electricity the chips themselves need.

The takeaway

The coverage points to one conclusion: AI's growth will be paced by how fast transformers, transmission lines and grid connections can be built, not by how fast chips can be made. Efficiency gains in cooling help at the margins. The bigger factors are queue reform, the mix of energy sources that ends up powering these campuses, and who pays for the new grid.