AI data centers drive US power demand to record highs two years running
AI data centers are pushing America’s power grid into record territory.
The US Energy Information Administration forecasts electricity demand will hit new records two years in a row: from 4,198 billion kilowatt-hours in 2025 to 4,256 billion in 2026 and 4,364 billion in 2027, per the EIA’s July 2026 outlook. Data centers built for AI and crypto are named as a main driver, alongside the electrification of heating and transport.
The milestone hiding in the numbers
One crossover inside the forecast says more than the totals. Commercial electricity sales are set to pass residential demand in 2026 for the first time ever: 1,550 versus 1,508 billion kilowatt-hours.
Put plainly, the buildings running servers will officially consume more electricity than the homes of the people they serve. For the entire history of the US grid, homes came first in that ranking. The AI buildout, together with the broader commercial load it anchors, is ending that.
How steep the curve gets
The forward-looking estimates suggest the records are just the start. Goldman Sachs expects US data center power demand to more than double by 2027, from 31 to 66 gigawatts.
Doubling a load of that size in roughly two years is the kind of demand shock grids are not designed to absorb casually. Generation, transmission and interconnection queues all move on multi-year timelines; data center construction now moves faster than any of them.
Every model has a meter running
The uncomfortable arithmetic behind every AI product is that inference is electricity. Every model you run has a meter running somewhere, and the EIA’s forecast is what those meters add up to at national scale.
That is why the policy question is arriving now rather than later: should AI companies pay for grid upgrades? The demand they are adding is measurable, the upgrades it forces are expensive, and someone’s bill will reflect them. The forecast makes one thing certain: the grid is now part of the AI story, records and all.
Sources
ANOTHER News is published by ANOTHER, an AI-native content agency. Daily coverage also runs on Instagram.
Related
DeepSeek is building a gigawatt of compute in Inner Mongolia, Bloomberg reports
Bloomberg reported on July 30, 2026 that Hangzhou-based DeepSeek is adding one gigawatt of compute in Ulanqab, Inner Mongolia, about 350 kilometers northwest of Beijing, building its own campus while leasing extra capacity, with part expected online by end of 2027 or start of 2028. The chips are undecided: Nvidia, Huawei, DeepSeek's own future silicon, or a mix. It would be the largest AI facility any Chinese company operates.
Meta discloses $279 billion in future AI data center lease commitments
In a filing for the quarter ended June 30, 2026, Meta disclosed $279 billion in future lease obligations, mostly AI data centers, up 53% from $183 billion a quarter earlier, with another $68 billion added in July alone. Under US accounting rules these commitments stay off the balance sheet until services begin. Shares fell roughly 7%.
Musk's Davos warning that AI will run out of electricity is aging into consensus
Speaking at the World Economic Forum on January 22, 2026, Elon Musk said electrical power is fundamentally the limiting factor for AI deployment, warning that with US generating capacity growing only 3 to 4 percent a year, the industry could soon produce more chips than it can turn on. Months later, data-center power deals and grid bottlenecks dominate the AI infrastructure conversation.