AI Data Center Energy Demand Drives a Wave of Grid Rule Rewrites

Regulators are now writing the rules for AI power

For most of the AI boom, the debate over electricity was about forecasts: how many terawatt-hours, how fast, and how bad. By late 2026 the debate has moved on. The open questions are about rules: who connects to the grid first, who pays for new substations, who can be switched off when the system is under strain, and how much water a cooling tower may evaporate. The newest marker of that change is a report from the International Energy Agency, Powering Artificial Intelligence, prepared for France's G7 Presidency. It reviews 34 jurisdictions that together hold about 90% of the world's data centre capacity16. The agency counted 134 policies and regulations governing how data centres connect to power systems, and almost 40% of them were announced or put into effect in the previous 12 months16.

That pace is the story. Governments are not waiting for demand estimates to settle. They are writing rules as the load arrives, and the rules they choose will decide where AI capacity gets built and who carries the cost.

The demand numbers agree on direction, not on size

The IEA's April 2026 report, Key Questions on Energy and AI, is the reference most coverage relies on. It found that data centre electricity use rose 17% in 2025, while consumption at AI-focused facilities jumped 50%. Global electricity demand grew 3% over the same period12. The agency projects total data centre consumption roughly doubling from 485 TWh to 950 TWh by 2030, or about 3% of global demand, with AI-focused sites tripling to 465 TWh12. Our World in Data puts the 2025 figure at about 1.5% of world electricity generation, roughly what Germany produces in a year, and estimates AI-focused centres at 155 TWh of that total13.

The consensus breaks down below the headline. Gartner's June 2026 forecast puts 2025 consumption at 447 TWh rather than 485 TWh, and projects 565 TWh for 2026, a 26% jump14. Some widely shared trackers still use the IEA's older 2024 baseline of about 415 TWh, rising to roughly 945 TWh by 20301117. Estimates that include crypto mining are much higher: Energy Institute/S&P Global figures reach about 788 TWh for 202513. Estimates of the AI share of 2030 demand vary widely, from around 285–320 TWh in one compiled dataset to the IEA's 465 TWh1214. These AI-specific numbers are modelled, not measured, because no global system meters AI electricity use separately14.

Uncertainty also points the other way. A May 2026 investment-bank analysis cited by the Congressional Research Service expected only 50–60% of data centre capacity scheduled for the next year or two to come online on time6. One study prepared for an environmental group argued that projections systematically overestimate how many facilities will be built6. In May 2026 the North American Electric Reliability Corporation (NERC) said many grid operators had lowered their summer 2026 demand projections because large loads were connecting more slowly than expected. Even so, it still expected peak demand to be 11 GW higher than in 2025, mostly because of data centres2.

The takeaway is that arguing over the exact 2030 total misses what matters. Demand is local and concentrated, and the strain shows up first in specific places. Data centres use about 21% of Ireland's electricity14. In Northern Virginia, summer peak load is up 23% and winter peak 45% since 201914. In the UK, 140 proposed data centre projects could together need 50 GW, which is 5 GW more than the country's current peak demand15.

Efficiency gains are being outrun

The IEA found that the energy needed for a single AI task has fallen by at least an order of magnitude a year recently. Total consumption keeps rising anyway, because users and the complexity of tasks, especially agentic workloads, are growing faster18. The agency estimates that a standard request to an AI agent uses about 1.1 Wh of GPU electricity, while an agentic request with reasoning uses about 50 Wh13. A simple text query, by contrast, uses less electricity than a television over the same period14.

The hardware is changing just as fast. The IEA reports that AI server power density rose elevenfold between 2020 and 2025 and expects another fourfold increase by 2027. At that point a rack the size of a refrigerator could draw as much power at peak as 65 households14. A typical hyperscale facility already uses as much electricity as about 100,000 households, and the largest campuses under construction are expected to use about twenty times that19. The money behind this is large. Capital spending by the biggest technology and infrastructure companies passed $400 billion in 2025 and is expected to rise another 75% in 202612.

Washington chose orders over a single national rule

In the US, the key moment came on June 18, 2026. The Federal Energy Regulatory Commission (FERC) unanimously issued show-cause orders to PJM, MISO, SPP, CAISO, ISO New England and NYISO, telling each either to justify its large-load interconnection rules or to reform them310. The orders followed a 2025 directive from Energy Secretary Chris Wright1. FERC Chairman Laura Swett called the issue "the biggest priority our country is facing at the moment"8.

Coverage differs on how far the orders go. One account says they require data centres to pay the full cost of grid upgrades tied to their own connections3. Others describe cost allocation as one of five reform categories that operators must address or defend18. The original Department of Energy proposal included a 20 MW national threshold and 100% cost responsibility for assigned upgrades, but FERC issued region-by-region orders rather than a final rule, and litigation is possible7. The narrower reading seems more accurate: FERC set the direction and left the details to each region. Its stated reason is that one national approach would not fit every region10. Thresholds already range from 10 MW to 75 MW across operators7.

The jurisdiction fight has not gone away. The National Association of Regulatory Utility Commissioners (NARUC), which represents state utility regulators, warned FERC that federal control over load interconnection would weaken states' ability to protect retail customers on affordability1. Two Senate bills would explicitly give FERC that authority6. FERC also told NERC on July 16 to develop reliability standards for "computational loads"6. NERC plans to file proposed requirements by the end of 2026, partly in response to incidents in Virginia and Texas where large loads dropped offline unexpectedly6.

Texas, which falls outside FERC's orders, has moved more forcefully. On August 3 the governor paused data centre development pending an audit of projects in ERCOT's interconnection queue2. ERCOT has also stopped approving energization for AI facilities of 75 MW or more until it files a verification report, due by December 107.

Flexibility is becoming the price of a connection

Across these systems, one idea keeps coming up: data centres that agree to be curtailed get connected faster. SPP's conditional service, accepted by FERC from July 1, offers the quickest hookup in exchange for possible temporary curtailment7. Texas law requires remote curtailment during firm load-shed events7. NERC found that summer 2026 peak demand in ERCOT was forecast to be 3.7 GW lower than in 2025, partly because more data centres can now be curtailed2. The Department of Energy went further in May with an emergency order allowing PJM to run backup generation at data centres as a last resort to prevent blackouts6.

The IEA gives this approach a global scale. It estimates that non-firm connections, which can be restricted when the network is stressed, could add 750–900 GW of grid hosting capacity worldwide16. In practice, a data centre's willingness to accept interruptions is turning into a form of currency. Developers who accept them will move up the queue, and those who insist on firm, always-on power will pay more or wait longer. Waits are already long: the average time through the US interconnection queue has risen to nearly five years, and more than nine in California9. Self-supplied power has been slower to arrive than the hype suggests. Only about 2 GW of announced behind-the-meter power was operating by mid-20266.

Cooling is now a permitting question

As rack densities rise, cooling choices are coming under regulation too. More than 60 bills on data centre water use were introduced in US states in 2025 and 202626. Early bills mostly required disclosure. Newer ones regulate cooling technology directly23. A South Carolina bill would require zero net water withdrawal and ban the use of groundwater or municipal water for cooling. A Kansas bill would require closed-loop systems and let local officials sue violators24. Illinois's POWER Act proposal treats closed-loop cooling as the efficiency baseline and requires quarterly public reporting30. Virginia has taken a softer route, tying grant money to the use of treated wastewater24.

The engineering trade-off is real. Evaporative cooling uses a lot of water but relatively little energy. Air cooling uses almost no water but the most energy. Direct liquid and immersion cooling suit dense GPU racks, but some coolants contain PFAS, the persistent "forever chemicals," and need special handling27. An Illinois industry representative put it simply: air cooling tends to use more energy, and liquid and evaporative methods tend to use more water30. That means a water mandate written in isolation can add to the grid problem. This is why the IEA calls for a holistic approach that covers energy and water efficiency together16. Nationally, data centres' direct water use is small, about 2% of US consumption by CRS's estimate, but the effects on individual communities can be large22. The Information Technology and Innovation Foundation (ITIF) argues for watershed-specific performance standards rather than technology mandates, and against federal preemption21.

Where this leaves things

The IEA's six principles are transparency, predictable connections, joint energy and water efficiency, fair cost allocation, grid-supporting flexibility, and dialogue. They read like a summary of where regulators are already going16. The agency's most important point may be that the effect on household bills depends largely on how regulators assign infrastructure costs16. The demand estimates will keep being revised. The bigger question now is whether flexible, cost-reflective rules can be agreed faster than new campuses can be built. FERC's regional approach, Texas's pause, and the patchwork of state water laws all suggest that, for now, the rules are lagging behind.