Somewhere between the first ChatGPT query of the morning and the last image generated before midnight, the world’s artificial intelligence data centers are now pulling 29.6 gigawatts of electricity. That is roughly the amount of power New York State consumes at the absolute peak of a brutal summer heat wave, when every air conditioner from Buffalo to Brooklyn is running full blast.
The figure comes from the 2026 AI Index Report, published by Stanford University’s Institute for Human-Centered Artificial Intelligence. Released earlier this year, the report has become one of the most cited benchmarks in the growing debate over AI’s physical footprint. And the number it puts forward is not a forecast or a worst-case scenario. It is a measure of capacity that already exists.
Putting 29.6 gigawatts in perspective
Numbers this large lose meaning without a frame of reference. The United States has roughly 1,300 gigawatts of total installed generating capacity, according to the U.S. Energy Information Administration. AI data centers, at 29.6 GW globally, now account for a slice that would have been unthinkable five years ago.
The New York comparison is especially vivid because the state’s peak demand is a recognized stress test for grid reliability. The New York Independent System Operator (NYISO) recorded its all-time peak of nearly 34 GW on July 19, 2013. More recent summer peaks have landed in the 29 to 31 GW range. During those episodes, wholesale electricity prices spiked sharply across the Northeast, a sign that the system was straining against its limits. The New York Department of Public Service uses those peaks as planning benchmarks in its seasonal energy outlook, assessing whether generation and transmission can handle extreme heat without rolling blackouts.
When Stanford’s researchers chose that same yardstick, they were making a deliberate point: the tech industry’s appetite for electricity now rivals a scenario that already pushes one of the nation’s largest grids to its edge.
Where the strain is showing up
The 29.6 GW figure is global, spread across hundreds of facilities in dozens of countries. But the pressure is not distributed evenly. In the United States, the pain points are concentrated in regions where data center construction has outpaced grid upgrades.
Northern Virginia, which hosts the densest cluster of data centers on Earth, has forced Dominion Energy to repeatedly revise its load-growth forecasts upward. PJM Interconnection, the grid operator covering 13 states and Washington, D.C., has seen its interconnection queue swell with data center requests, creating yearslong backlogs for new projects trying to connect to the grid. In Texas, ERCOT has flagged data centers as a major driver of demand growth in its long-term planning documents, even as the state’s grid remains vulnerable to extreme weather events.
For utility customers in these regions, the consequences are tangible. New transmission lines must be built. Substations must be upgraded. And the question of who pays for that infrastructure is becoming politically charged. Regulators in Virginia and several other states are actively debating whether the cost of serving massive AI facilities should be spread across all ratepayers or borne directly by the companies signing the contracts.
What the data does and does not tell us
The Stanford figure is the strongest anchor in this discussion. The AI Index Report is an annual research program that government agencies, major media outlets, and academic institutions treat as authoritative. But it is not without limitations. The report’s landing page provides the 29.6 GW headline; the full methodology, including which facilities count as “AI” versus general-purpose data centers, has not been broken down in publicly available appendices as of June 2026. Without that granularity, independent researchers cannot yet verify whether the number captures only GPU-heavy training clusters or also includes inference workloads and mixed-use facilities.
The New York comparison, meanwhile, works as an analogy rather than a direct equivalence. AI data centers drawing 29.6 GW are scattered worldwide. New York’s peak demand occurs on a single interconnected grid over a few sweltering hours. The two are not competing for the same electrons. The real collision happens locally, where a single large data center can consume as much power as a small city and overwhelm transmission lines that serve surrounding communities.
A Brookings Institution analysis compiled references to Department of Energy, Lawrence Berkeley National Laboratory, and International Energy Agency data on data center electricity use, alongside Congressional testimony on AI power projections and industry estimates. The analysis pointed in one direction: demand is growing faster than new supply can be built. But Brookings is a secondary source. Readers looking for the hardest numbers should follow its citations back to the underlying government reports, especially when weighing claims about multi-decade demand trajectories.
The efficiency question
Not everyone in the industry sees a straight line from 29.6 GW to grid catastrophe. Chipmakers like Nvidia and AMD are shipping processors that deliver more computation per watt with each generation. Liquid cooling systems, which are replacing traditional air cooling in many new AI facilities, can cut a data center’s total energy overhead significantly. And some operators are co-locating facilities near dedicated power sources, including solar farms, natural gas plants, and even nuclear reactors, to avoid burdening the public grid at all.
These efficiency gains are real, but they face a familiar problem: the Jevons paradox. When each unit of AI computation becomes cheaper and less energy-intensive, demand for computation tends to surge in response, often erasing the savings. The history of computing is littered with examples of efficiency improvements that were swallowed whole by exponential growth in usage.
Where the next 29.6 gigawatts will come from
That is the question grid operators, utility executives, and policymakers are now wrestling with. If AI data center capacity doubles in the next few years, as several industry forecasts suggest, the world will need to find another 30 GW of generation and the transmission infrastructure to deliver it. Building that out takes time. A new natural gas plant can take three to five years from permitting to operation. A large solar or wind farm faces similar timelines, plus the added complexity of energy storage. Nuclear projects, which some tech companies have publicly embraced, operate on even longer horizons.
Grid operators in fast-growing regions are already warning that interconnection queues are lengthening and that reliability margins are thinning. Utility customers are hearing more about potential rate increases, new transmission corridors, and the possibility that large industrial loads could crowd out smaller projects waiting for grid access.
Future editions of the Stanford AI Index may break out training and inference loads separately, or distinguish between retrofitted legacy data centers and purpose-built AI campuses. If grid operators begin publishing explicit breakdowns of AI-related demand in their planning documents, it will become easier to test whether AI is overwhelming the grid or playing a manageable role in broader electrification trends.
For now, the picture is clear enough to act on, even if the fine details remain blurry. AI data centers already draw a power load comparable to New York’s worst summer peak. Many regional grids are strained under far smaller increments of new demand. The economic and scientific promise of AI is enormous, but so is the physical infrastructure required to keep it running. The wires, transformers, and power plants that underpin the system were not built for this, and upgrading them will be neither fast nor cheap.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
