The world’s AI data centers now pull roughly 29.6 gigawatts of electricity from power grids, according to estimates aligned with the International Energy Agency’s latest assessment of energy and AI. To put that in perspective: New York State’s grid operator, the NYISO, projects a statewide summer peak demand of 31,471 megawatts, the maximum draw when every air conditioner in every apartment from Montauk to Buffalo is running at once. AI infrastructure alone is now consuming nearly that much power, not for a few sweltering afternoons, but around the clock, every day of the year.
That gap between a seasonal spike and a constant industrial load is exactly what has grid planners, utility executives, and state regulators scrambling. Building a new power plant or transmission line takes five to ten years. Training and deploying a new large language model takes months. The mismatch is already showing up in interconnection queues, rate cases, and emergency reliability reviews across the country.
Where the numbers come from
Two institutional sources anchor the comparison. New York’s 2025 summer energy outlook, built on NYISO data, sets the state’s projected peak at 31,471 MW and total available capacity at 40,983 MW. The roughly 9,500 MW reserve margin is designed to keep the lights on during heat waves. It was not designed for a world in which a single hyperscale data center campus can request 1,000 MW or more of new load.
The IEA, in a companion public statement, warned that AI is “set to drive surging electricity demand from data centres” while also offering tools that could reshape energy operations. The agency’s detailed report treats AI as the primary accelerant behind rising data center power consumption globally, with training runs and inference workloads growing far faster than efficiency gains can offset.
The 29.6 GW figure itself circulates widely in energy and technology reporting and is consistent with the IEA’s broader findings. The agency’s published materials do not isolate that exact number in a single labeled table, so it should be understood as an informed estimate drawn from the IEA’s data rather than a single audited line item. The scale, however, is not in serious dispute. Even conservative readings of the IEA’s scenarios place global AI-related data center demand in the tens of gigawatts and climbing.
Who is driving the buildout
The demand is not abstract. Microsoft, Google, Amazon, and Meta have collectively announced or begun construction on data center campuses that, if fully built, would add tens of gigawatts of new load to grids in Virginia, Texas, Georgia, Iowa, and beyond. Meta’s planned campus in Richland Parish, Louisiana, alone could draw more than 2 GW. Microsoft has signed power purchase agreements with nuclear operators to secure carbon-free baseload. Amazon has invested in small modular reactor developers for the same reason.
These commitments show up in interconnection queues managed by regional grid operators like PJM Interconnection, which oversees the grid across 13 states and the District of Columbia. As of early 2025, PJM’s queue contained more than 260 GW of proposed generation and storage projects, many of them driven by data center demand in Northern Virginia, the densest data center market on Earth. The queue is so backlogged that new projects face wait times of four years or more just to get a grid connection study completed.
Why constant load changes the math
The comparison between 29.6 GW and New York’s 31,471 MW peak is striking, but it can also be misleading if taken at face value. New York hits that peak for only a handful of hours on the hottest summer days. For most of the year, statewide demand sits well below 20,000 MW. Data centers, by contrast, run at or near full capacity 8,760 hours a year. A single gigawatt of constant data center load consumes roughly three to four times more total energy annually than a gigawatt of peak demand that only materializes during summer afternoons.
That distinction matters for fuel procurement, emissions accounting, and infrastructure investment. A utility planning for a summer peak can rely on peaker plants that sit idle most of the year. A utility serving a 24/7 data center load needs baseload generation that never stops. If that generation comes from natural gas, the emissions implications are enormous. If it comes from renewables, the storage requirements are enormous. Either way, the capital costs are real, and they flow through to ratepayers.
What grid operators and regulators are doing
Responses are emerging at multiple levels. In Virginia, Loudoun County officials and Dominion Energy have faced public pressure over the pace of data center approvals, with residents raising concerns about noise, water use, and the strain on local substations. Georgia’s Public Service Commission has scrutinized large load requests from data center operators seeking to connect to Georgia Power’s system. At the federal level, the Department of Energy has launched initiatives to accelerate transmission buildout, recognizing that generation capacity means little if the wires to deliver it do not exist.
Policy discussions are coalescing around three broad strategies. The first is to fast-track investment in low-carbon generation, particularly solar, wind, and battery storage, so that new AI demand does not simply extend the life of aging fossil fuel plants. The second is to push AI developers and chip manufacturers to prioritize energy efficiency, not just raw performance, in model design and hardware architecture. The third is to reform siting and interconnection rules so that large data centers are steered toward regions with surplus clean power rather than wherever land and tax incentives are cheapest.
What remains unresolved
Several important questions do not yet have clear answers. The IEA’s report does not break out how much of the 29.6 GW falls on U.S. grids versus facilities in Europe, Asia, or the Middle East. NYISO’s filings do not separate AI-related data center load from other commercial and industrial consumption within New York. And the difference between contracted capacity and actual metered consumption remains significant: a data center campus that has signed a power agreement but has not yet energized its servers does not strain transformers today, even if it will in two years.
Technological uncertainty cuts both ways. Advances in chip efficiency, liquid cooling, and workload scheduling could slow demand growth. But the rapid expansion of AI into real-time video generation, autonomous systems, and always-on software copilots could multiply inference volumes far beyond current projections. The IEA acknowledges both possibilities without predicting which will dominate.
What this means for electricity bills
For households and businesses that share a grid with large data centers, the practical stakes are straightforward. Every new gigawatt of data center load competes for the same generation, transmission, and distribution infrastructure that serves existing customers. If new supply does not keep pace, grid operators face a short menu: curtail industrial users, fire up expensive peaker plants, or accept reliability risks. All three options can raise costs, and those costs show up on monthly bills.
The 29.6 GW figure, whatever its precise sourcing, captures something real and urgent. AI workloads are arriving on the grid faster than new infrastructure can be permitted, financed, and built. The institutions responsible for keeping the lights on are aware of the problem. Whether they can move fast enough to stay ahead of it is the question that will define energy policy for the rest of this decade.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
