On the hottest afternoons of summer, when air conditioners across New York State are running full blast, the grid serving 20 million people peaks at roughly 30 gigawatts of electricity demand. That number has long been a benchmark for grid planners. Now, according to estimates from the International Energy Agency and independent grid analysts at Grid Strategies, artificial intelligence data centers in the United States alone are pulling about 29.6 gigawatts, a load that has materialized in a fraction of the time it took New York to build its grid.
The comparison is not hypothetical. It is showing up in utility planning documents, interconnection queues, and the capital budgets of the world’s largest technology companies. And with summer 2026 weeks away, grid operators in several regions are warning that the margin between available supply and projected demand is thinner than it has been in years.
Where the 29.6 gigawatt figure comes from
The U.S. Energy Information Administration defines peak demand as the highest instantaneous electricity draw recorded during a given period, typically driven by cooling loads on the hottest summer days. That definition, outlined in the EIA’s electricity overview, is the yardstick utilities use to plan generation and transmission for the most stressful hours of the year.
No single federal dataset currently isolates AI data center consumption from other commercial and industrial loads. The EIA’s “commercial” sector bundles data centers with office buildings, hospitals, and retail. The 29.6 GW estimate instead comes from third-party analyses that model server density, GPU utilization rates, and cooling overhead across known facilities. The IEA’s Electricity 2025 report flagged data centers as the fastest-growing demand segment globally, and Grid Strategies has tracked U.S. utility interconnection requests showing more than 90 GW of data center capacity in the pipeline as of early 2025. The 29.6 GW figure represents what is already drawing power, not what is planned.
New York’s own grid provides the other half of the comparison. The New York Independent System Operator publishes annual forecasts in its Gold Book, and the state Department of Public Service issues a summer energy outlook each year detailing expected peak demand and reserve margins. Historical summer peaks for the state have ranged from about 29 to 33 GW depending on weather. The AI comparison works as an order-of-magnitude illustration: the computing industry’s electricity appetite now rivals the combined residential, commercial, and industrial load of one of America’s most energy-intensive states at its annual maximum.
Who is building and how fast
The scale of corporate investment makes the demand trajectory concrete. Amazon disclosed plans to spend more than $100 billion on AI infrastructure in 2025 alone, a figure reported by the New York Times in its coverage of the company’s latest cloud expansion. Microsoft, Google, and Meta have each announced comparable commitments. Individual campuses under construction in northern Virginia, central Ohio, and west Texas carry nameplate capacities of 300 to 500 megawatts, loads that would have ranked among the largest industrial customers in any state a decade ago.
Each new facility triggers a chain of grid infrastructure work. Utilities must complete multi-year interconnection studies, build new substations, and in some cases construct dedicated transmission lines. PJM Interconnection, the regional grid operator covering 13 states and Washington, D.C., reported in 2024 that data center requests dominated its interconnection queue, pushing estimated wait times past four years. That backlog means some facilities announced today will not reach full power until 2029 or later, but it also means the 29.6 GW figure is a snapshot of a curve that is still steepening.
The grid stress points heading into summer
Grid operators plan for summer by tallying available generation, forecasting peak demand, and calculating reserve margins. When reserves drop below a threshold, typically around 15 percent, the risk of rolling blackouts during a heat wave rises sharply. The addition of large, constant-draw data center loads complicates that math because, unlike air conditioning, AI workloads do not cycle off when temperatures drop at night. A 400 MW data center runs around the clock, consuming power at 2 a.m. that would otherwise help rebuild reserves for the next afternoon peak.
Natural gas remains the primary fuel for the flexible “peaker” plants that fill gaps between baseload generation and peak demand. The EIA’s weekly storage reports track how much gas is available nationally and by region, and heading into June 2026, storage levels in the East and South Central regions are being watched closely. While no federal report has yet isolated AI-specific demand as a driver of gas consumption, the cumulative effect of tens of gigawatts of new always-on load is already visible in utility procurement filings, where gas-fired capacity contracts have surged.
Some states are responding with policy. Virginia, home to the largest concentration of data centers in the world, passed legislation in 2024 requiring new facilities to meet stricter noise and environmental standards. Local governments in Georgia and Indiana have imposed temporary moratoriums on data center construction while they assess grid and water impacts. At the federal level, the Department of Energy launched a data center energy efficiency initiative in late 2025, but it remains voluntary.
What satellite imagery reveals
Researchers have turned to Copernicus Sentinel-2 imagery, collected by the European Space Agency and distributed through its open-access platform, to independently track the physical footprint of data center construction. The facilities are distinctive from orbit: large, windowless halls with extensive rooftop mechanical equipment and adjacent electrical switchyards. Multispectral imagery can detect cleared land, new concrete pads, and substation construction months before a company makes a public announcement.
The limitation is translating square footage into watts. A satellite can measure a building’s roof area but cannot see how many GPU racks are inside or how hard they are running. Different research groups using the same imagery have produced varying power estimates depending on their assumptions about equipment density and cooling efficiency. Satellite tracking is strongest as a leading indicator of where growth is happening and how quickly sites are being built. It is weaker as a tool for pinning down real-time electricity consumption.
What the numbers still cannot tell us
Several important questions remain open. The 29.6 GW estimate, while grounded in credible modeling, has not been confirmed by metered federal data. Until the EIA or another agency publishes a dedicated data center load series, the figure depends on assumptions that different analysts may reasonably dispute. The pace of growth is also uncertain: if AI model training shifts toward more efficient architectures, or if inference workloads plateau, demand could level off sooner than current projections suggest. Conversely, if every major enterprise adopts AI agents at scale, the curve could steepen further.
New generation is being built, but it takes time. Nuclear restarts, such as the planned recommissioning of Three Mile Island Unit 1 to serve a Microsoft data center, are years from delivering power. Utility-scale solar and battery storage are being deployed faster, but siting and permitting delays remain common. The fundamental tension is between an AI industry that can break ground on a new campus in months and an electricity system where major infrastructure projects take five to ten years from proposal to energization.
For grid operators preparing for summer 2026, the immediate concern is not the long-term trajectory but the next three months. Reserve margins in parts of Texas, the Midwest, and the Southeast are tighter than they were a year ago, and data center load is one reason why. Whether the grid holds through the hottest weeks will depend on weather, plant availability, and whether utilities have accurately accounted for the newest and hungriest customers on their systems.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
