AI Data Center Power Demand Pushes US Electricity to Record Highs

A third straight record, with a fourth on the way

For almost twenty years, American electricity use barely grew. That has now clearly ended. In its October Short-Term Energy Outlook, the U.S. Energy Information Administration said national power consumption will set new records in both 2026 and 2027. It projects demand rising from a record 4,195 billion kilowatt-hours in 2025 to 4,288 billion kWh this year and 4,356 billion kWh next year.3 The agency named data centers built for artificial intelligence and cryptocurrency as the main driver. Homes and businesses replacing fossil fuels with electricity for heating and transportation also add to the load.1

The headline totals hide a change in who uses the power. The commercial sector, where data centers are counted, is on track to overtake residential demand for the first time on record.2 Under the October figures, 2026 commercial sales reach 1,549 billion kWh, compared with 1,541 billion kWh for residential customers. Both numbers beat the 2025 records of 1,493 billion and 1,515 billion kWh.3

The two-year rise works out to about 3.8%.4 That may sound small, but it is a large change for a grid that was planned for years on the assumption of flat demand. A federal estimate from 2026 put data centers at about 4.7% of U.S. electricity use in 2024, or roughly 192 terawatt-hours. The same estimate says that share could reach between 9.5% and 15.3% by 2030.4

The forecasts keep moving, and that matters

The coverage does not settle on one set of numbers. EIA's July outlook put 2026 demand at 4,269 billion kWh and 2027 at 4,399 billion kWh.2 The October revision raised this year's figure and lowered next year's.3 The residential picture changed even more. In July, EIA expected home electricity sales to dip to 1,508 billion kWh in 2026.2 By October it expected a new record of 1,541 billion kWh.5 Another widely shared version of the outlook used different figures again: 4,270 billion kWh for 2026 and 4,349 billion kWh for 2027.8

None of this changes the main conclusion. Every version has demand rising to records in both years. But it shows how hard it is to predict a load driven by private building decisions. The near-term number depends on weather. The 2027 number depends on how many announced data centers actually get built and connected. Regional revisions make the point. EIA cut its 2027 sales forecast for the West South Central region by 4.7% after Texas paused new data center connections to the ERCOT grid until they could be verified and audited.4 One investment-bank analysis from May found that only 50% to 60% of data center capacity scheduled for the next year or two will open on time.4

PJM is where the pressure shows first

The effects are clearest in PJM Interconnection, the grid operator covering the Mid-Atlantic and parts of the Midwest. Its January 2026 long-term forecast projects summer peak demand growing 3.6% a year over the next decade. It expects the peak to reach about 222 GW by 2036 and about 253 GW by 2046. PJM raised its forecasts in 14 utility zones specifically to account for data center load.25 One analysis of that forecast found that large loads, mainly data centers, account for more than all of PJM's expected peak growth over the next five years. Without them, the region's underlying demand would shrink.22

The system has already hit real limits. On July 2, PJM recorded an estimated all-time peak of about 168,158 MW, breaking a record that had stood since August 2006.25 In May, PJM had expected a summer peak of around 156,400 MW and treated anything near 169,100 MW as an unlikely extreme case.29 Real-time prices that day reportedly reached $1,869 per megawatt-hour.25

The capacity market shows the same strain. PJM's July auction for the 2028/29 delivery year cleared at its $325 per megawatt-day price cap. It fell about 6.8 GW short of its reliability target.27 EIA now expects wholesale prices in PJM to rise 41% this year, while prices in the hydro-heavy Pacific Northwest fall 23%.5 That gap is a reminder that a single national average describes very little. Data centers are clustered in specific places, and those places are absorbing the cost.

There is also a new risk to how the grid operates. On July 22, a transmission fault in Northern Virginia reportedly caused about 3,800 to 4,000 MW of data center load to drop off the grid at once and switch to backup generators. That was a large enough swing to disrupt the balance between supply and demand.28 The event suggests grid operators can no longer treat data centers as steady, predictable load. They have to plan for large customers that can disconnect suddenly in response to grid disturbances.

The case that the crunch is overstated

The most important disagreement in recent coverage is about whether the projected shortage is real. PJM's independent market monitor told federal regulators in September that 4,016.9 MW of forecast data center load had been canceled or delayed. That is 42% of the growth PJM had projected between the 2026/27 and 2028/29 delivery years. Remove that load, and the monitor calculates the 6.8 GW shortfall shrinks by 55%, to about 3 GW.30 The monitor's concern is that each utility zone's share of new capacity costs will be fixed to PJM's 2026 forecast through 2042/43. If that forecast proves too high, customers could pay for years for capacity that serves load that never arrives.30

Other evidence points the same way. After PJM began requiring transmission service agreements and financial commitments from large customers, its high-probability data center forecast reportedly dropped from about 18 GW at the end of 2025 to about 11 GW in the second quarter of 2026.28

In our reading, this does not mean the demand boom is fake. EIA's actual consumption records and PJM's July peak are measured facts, not forecasts. What it shows is that interconnection queues have been full of speculative requests. The shortage is real at the margin. It is probably smaller than the worst projections, and it falls unevenly on the regions where planners and ratepayers are most exposed to forecasting mistakes.

Regulators shift risk onto data centers

Policy is moving toward making large new customers carry more of their own risk. In August, PJM proposed that new loads of 50 MW or more either bring their own power supply or accept emergency curtailment when the grid is under strain.28 At the federal level, FERC opened a proceeding in June on whether grid operators' interconnection rules can handle very large loads. In July, it directed NERC to draft reliability standards for computing loads, with proposals due by the end of 2026.4 The administration's "Ratepayer Protection Pledge" asks tech companies to pay for their own energy and infrastructure costs. It is voluntary and cannot be enforced, which leaves state regulators to decide how costs are split.4 States are also acting on their own. New York temporarily paused environmental permits for new data centers in July, citing concerns about electricity bills and reliability.4

As a result, developers increasingly build power plants alongside their data centers. Natural gas turbines and engines lead among announced on-site generation projects, followed by nuclear and fuel cells.4 Across the U.S. grid as a whole, natural gas will still supply about 40% of electricity in 2026. Renewables are expected to rise from 24% in 2025 to 27% in 2027.6

Cooling becomes a way to stretch scarce power

Inside the buildings, the power squeeze is changing how data centers are cooled. A single Nvidia GB200 NVL72 rack draws roughly 120 to 130 kW. Traditional air-cooled halls were built for racks of 8 to 12 kW.16 Average rack density reportedly rose from about 16 kW in 2025 to 27 kW in 2026, and Nvidia's Vera Rubin platform can push a single rack to 246 kW.18 Industry roadmaps now look toward racks that use a megawatt or more, with individual chips approaching 4,000 watts.15

The industry agrees on where cooling is heading, but estimates of the pace differ. Goldman Sachs projects that liquid-cooled systems will make up 76% of AI servers in 2026, up from 15% in 2024.14 TrendForce counts AI chips rather than servers and puts liquid cooling at 53% this year and nearly 60% in 2027. It names Google as the most aggressive adopter, using liquid cooling in more than 80% of its AI servers.17 The two estimates measure different things, but both describe a shift happening within a few years.

Cooling matters to the grid story because it affects how much of a site's power reaches the chips. Older air-cooled facilities often run at a power usage effectiveness (PUE) of about 1.55 to 1.67. In other words, roughly a third of the electricity coming in goes to supporting equipment rather than computing. Direct-to-chip liquid cooling commonly gets PUE down to around 1.10 to 1.20.18 For an operator with a fixed power allocation from its utility, every megawatt saved on cooling can run more computing without waiting years for a new grid connection.18

That benefit has limits. Lower overhead per server does not reduce overall demand when operators simply use the savings to install more servers. Liquid cooling helps each grid connection deliver more computing. It does not shrink the national total. The consumption records EIA is projecting look likely to hold. The open questions are how much of the announced building actually gets connected, and who ends up paying for the grid capacity it needs.