Somewhere in northern Virginia, a construction crew is pouring concrete for yet another windowless building that will draw more electricity than a small city. Multiply that scene across dozens of sites in Texas, Ohio, Georgia, and overseas, and you arrive at a number that has jolted grid planners: AI data centers worldwide now consume an estimated 29.6 gigawatts of power, according to the International Energy Agency’s 2025 electricity analysis. That is nearly equal to the peak electrical demand of the entire state of New York, a load that covers 20 million residents, their offices, hospitals, factories, and subway trains.
A single industry segment rivaling the electrical appetite of America’s fourth-largest state is not a curiosity. It is a structural shift in how power grids will need to operate for the foreseeable future, and the consequences are already showing up in utility planning documents, interconnection backlogs, and early signs of rising capacity charges for ordinary ratepayers.
Putting 29.6 gigawatts in context
The clearest yardstick comes from New York’s own grid operator. The New York Independent System Operator’s summer outlook projected statewide peak demand at 31,471 megawatts for the summer of 2025, or roughly 31.5 gigawatts. The 29.6 GW data center figure falls within about 6 percent of that ceiling. NYISO listed total capacity resources at 40,983 MW, leaving a reserve margin of roughly 9,500 MW above projected peak. That cushion exists to handle heat waves, equipment failures, and unexpected demand spikes. It was never designed to absorb an entirely new category of industrial consumer appearing in concentrated bursts.
The IEA, in its broader examination of energy and AI, frames the surge as a supply-chain and grid-planning problem, not merely an efficiency puzzle. Training large language models and running inference at scale requires power around the clock, unlike homes and offices where demand rises and falls with the time of day. A single 500 MW data center campus can stress local transmission lines the way a small city would, but without the natural overnight dip that lets operators schedule maintenance and recharge storage.
Stanford’s AI Index, which tracks the broader trajectory of compute demand, documents how the hardware required for frontier models has been roughly doubling every six to ten months. Each new generation of chips draws more watts per rack, and each new model generation requires more racks. That compounding is the engine behind the 29.6 GW figure, and nothing in current industry roadmaps suggests it is about to slow down.
Where the numbers get fuzzy
No single audited dataset breaks the 29.6 GW figure down by region, operator, or facility type. The IEA’s estimate draws on member-country energy statistics and its own modeling, giving it more rigor than a typical industry white paper, but the underlying methodology does not fully separate hyperscale campuses from smaller colocation facilities, enterprise server rooms, or cryptocurrency operations. Readers should treat the headline number as a credible order-of-magnitude estimate rather than a census-grade measurement.
Grid-level data has its own gaps. New York’s Department of Public Service publishes aggregate capacity and demand forecasts through its public document portal, but those filings do not isolate how much load comes from data centers specifically. NYISO’s interconnection queue, where proposed generation and large-load projects file for grid access, offers clues, but queue filings reflect intent rather than guaranteed construction. Projects are routinely delayed, downsized, or withdrawn.
Competing estimates from investment banks and consulting firms vary by as much as 30 percent depending on assumptions about chip efficiency gains, workload migration, and the pace at which new generation comes online. Some scenarios assume rapid improvements in server utilization and cooling that would moderate growth; others assume new AI applications will more than offset those gains. Policymakers are working with wide confidence intervals at the exact moment they need precision.
The bottleneck is not watts, it is wires
The most immediate pressure point is not total electricity supply but the physical infrastructure that delivers it. Grid operators in the mid-Atlantic and Northeast, particularly PJM Interconnection, which manages the grid across 13 states and the District of Columbia, have watched their interconnection queues swell with data center requests. In northern Virginia alone, the largest data center market on Earth, Dominion Energy has reported load-growth projections that dwarf anything the utility planned for a decade ago.
Traditional planning models were built around steady, predictable load growth of 1 to 2 percent per year. AI-driven demand introduces step changes: hundreds of megawatts arriving in a single project on timelines that often outpace the three-to-seven-year cycle needed to permit and build new transmission lines and substations. That mismatch between demand speed and infrastructure speed is the core reliability risk, more than the absolute size of AI’s load today.
The risk is not evenly distributed. Regions with ample spare capacity, robust transmission, and access to low-cost renewables can absorb new data centers with relatively modest systemwide effects. In constrained areas where aging plants are retiring and new lines face local opposition, even a handful of large AI campuses can consume most of the remaining margin. Northern Virginia, parts of central Ohio, and pockets of the Dallas-Fort Worth metroplex are already testing those limits.
What tech companies are doing about it
The largest hyperscalers are not waiting for utilities to solve the problem. Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart a unit at the Three Mile Island nuclear plant in Pennsylvania. Amazon acquired a data center campus adjacent to Talen Energy’s Susquehanna nuclear station. Google has signed agreements with multiple geothermal and advanced nuclear developers. These deals reflect a recognition that grid-supplied power alone may not arrive fast enough or in large enough quantities to feed planned AI expansion.
On the hardware side, Nvidia’s latest Blackwell GPU architecture is designed to deliver more computation per watt than its predecessor, and competing chipmakers are making similar efficiency claims. But history suggests that efficiency gains in computing tend to unlock new workloads rather than reduce total consumption, a dynamic economists call the Jevons paradox. Cheaper, faster inference could make AI ubiquitous in applications from drug discovery to autonomous vehicles, driving aggregate power demand higher even as each individual query uses less energy.
What it means for your electric bill
For utility customers, the practical consequence is direct. Concentrated data center loads compete for the same transmission capacity and generation reserves that keep household electricity affordable and reliable. When a hyperscaler signs a long-term power purchase agreement for 300 or 500 MW in a single grid zone, that capacity is no longer available to cushion a summer heat wave or an unexpected plant outage.
Ratepayers in regions with heavy data center development may see higher capacity charges on their bills as utilities invest in new lines, substations, and peaking plants to maintain reliability standards. In some jurisdictions, regulators are already debating whether large data center customers should bear a greater share of grid-upgrade costs. Georgia’s Public Service Commission, for example, has scrutinized Georgia Power’s plans to add generation capacity partly driven by data center demand, questioning whether residential customers should subsidize industrial growth.
Policy responses are beginning to coalesce around three levers. First, requiring large loads to fund or accelerate grid upgrades, reducing the cost burden on existing customers even if it slows project timelines. Second, tying new data center approvals to on-site or contracted clean generation, limiting emissions impacts even where local transmission constraints persist. Third, improving transparency through better reporting of sectoral electricity use and clearer interconnection data, which would narrow the uncertainty bands that currently surround AI’s power footprint.
A heavy industry that looks like a tech company
The best-supported conclusion as of mid-2026 is that AI data centers have become a first-order consideration in power-system planning, even if their exact global consumption remains imprecise. The verified grid statistics from New York, the modeling work from the IEA, and the compute trends documented by Stanford all point in the same direction: digital infrastructure now behaves, electrically, like a heavy industry. It demands baseload power, concentrates in specific geographies, and reshapes the economics of every grid it touches.
How quickly regulators, utilities, and technology companies adapt will determine whether the rise of AI strains grids to the breaking point or catalyzes a new round of investment in more resilient, cleaner power systems. The 29.6 gigawatt figure is not the final number. It is the opening bid.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
