The White House expanded its Ratepayer Protection Pledge on July 23, 2026, saying that more than 200 additional utilities, data-center developers, cooperatives, states, and other participants had joined the effort. The administration says the pledge now covers 80 percent of electricity delivered to US homes and businesses and 263 million people. Those are administration claims, not an independent audit of customer protection.1
The distinction matters because the pledge is voluntary. It does not displace state public utility commissions, Federal Energy Regulatory Commission jurisdiction over interstate transmission and wholesale rates, utility tariffs, or negotiated service agreements. Its significance is political and directional: the federal government is pressing large computing customers to finance the generation, network upgrades, and dedicated facilities their projects cause, and to keep paying for contracted capacity even when actual use falls short.2
That formulation moves the AI-energy debate away from one of its most visible but least useful symbols—the electricity assigned to a single prompt. The immediate infrastructure question is whether a 100-megawatt, 500-megawatt, or gigawatt-scale campus can be supplied on schedule without shifting generation, transmission, equipment, water, or cancellation costs to other customers.
Utilities do not reserve transformers, reinforce transmission, procure capacity, or construct substations in response to individual prompts. They invest against forecasts, contracted service levels, ramp schedules, expected utilization, and the probability that a proposed project will actually arrive. Per-query energy remains relevant to efficiency and emissions accounting. It is not the unit on which an electric system plans capital.
The load is becoming system-significant
The International Energy Agency estimates that data centers used about 415 terawatt-hours of electricity worldwide in 2024, roughly 1.5 percent of global consumption. Its base case reaches about 945 terawatt-hours in 2030, with AI-accelerated servers accounting for nearly half of the net increase. The global share remains modest, but the demand is concentrated in particular regions and can arrive faster than generation, transmission, and permitting.3
For the United States, Lawrence Berkeley National Laboratory’s June 2026 update places reference-case data-center electricity use at 649 terawatt-hours in 2030. Its compounded uncertainty range is 521 to 843 terawatt-hours, or 9.5 to 15.3 percent of total US electricity consumption. The report does not predict that every announced project will be built. It models equipment shipments, utilization, operating life, cooling performance, and other assumptions that remain uncertain.4
EPRI reaches a similarly broad range using project-development scenarios, estimating that data centers could account for 9 to 17 percent of US electricity consumption in 2030 and that nominal information-technology capacity could reach 56 to 132 gigawatts. The breadth of these estimates is itself a planning fact: utilities must make physical commitments before the industry’s final load trajectory is known.5
Why prompt arithmetic cannot carry the policy debate
A 2024 ACM study of 88 models across 10 tasks found more than a 1,450-fold spread in energy use per 1,000 inferences. Under the study’s test conditions, text classification averaged about 0.002 kilowatt-hours per 1,000 inferences, text generation about 0.047 kilowatt-hours, and image generation about 2.9 kilowatt-hours. Task, model, hardware, output length, batching, and utilization all changed the result.6
A separate production-oriented analysis estimated a median 0.34 watt-hours for a typical frontier-model query under its assumptions, rising to 4.32 watt-hours when test-time computation increased token use fifteenfold. The authors also estimated that combined improvements in models, serving systems, and hardware could reduce energy per query by roughly eight- to twentyfold.7
These studies do not establish a universal number. They show why one does not exist. Per-query estimates can inform product design, fleet efficiency, and emissions inventories. They cannot determine whether a particular campus requires a new transmission line, whether its load arrives in 2028 or 2031, or who pays if the project is canceled after long-lead equipment has been ordered.
Local systems carry the national load
Northern Virginia shows how a manageable national percentage can become a dominant regional planning issue. EPRI estimates that Virginia is currently the only state in which data centers consume more than 20 percent of electricity and projects a 39 to 57 percent share by 2030 across its scenarios.5
The US Energy Information Administration reports that summer peak load in PJM’s Dominion zone reached 23,905 megawatts in 2025, 23 percent above 2019, while the 2025–26 winter peak was 45 percent above the 2019–20 level. PJM expects summer peak load in the zone to grow an average of 5.4 percent annually over the next decade, largely because of data centers.8
Wholesale and retail effects depend on supply, network conditions, market rules, retirements, weather, and the timing of load additions. Data centers are not the sole cause of rising prices in constrained regions. They are, however, a major new source of concentrated demand, and their effect cannot be inferred from a national average.
Infrastructure means equipment, commitments, and cancellation risk
Long lead times make the allocation problem immediate. Reuters reported in July 2026 that generator step-up transformer lead times exceeded 160 weeks and that high-voltage circuit-breaker lead times had reached 125 weeks. Some utilities were buying equipment years before expected need. Such commitments create exposure before a data center consumes its first kilowatt-hour.9
A utility may need to complete studies, secure land, reserve crews, order transformers, procure supply, or reinforce transmission before the customer reaches a final investment decision. The customer may resist binding commitments until the utility can provide a credible price and service date. Each party seeks certainty before the other makes an irreversible investment.
Federal regulators are now treating large loads as a distinct planning and tariff problem. On June 18, 2026, FERC issued tailored show-cause orders under section 206 of the Federal Power Act to PJM, MISO, SPP, CAISO, ISO New England, and NYISO. The grid operators and their transmission owners received 60 days to justify existing tariffs or propose reforms addressing study processes, cost shifting, co-location, flexible service, and generation near large loads; they also received 30 days to report how adequate generation would be secured.10
The orders begin proceedings; they are not final uniform rules. FERC expressly recognized that the six regions differ and that a single national tariff design may not be appropriate.10
Who pays is a tariff and contract question
The Ratepayer Protection Pledge states a cost-causation principle: large data-center customers should finance the incremental supply and infrastructure their projects require. Whether that principle protects other customers depends on the governing instruments.
Study deposits can reimburse engineering work. Upfront contributions can cover dedicated facilities. Letters of credit, guarantees, or other security can match equipment and cancellation exposure. Minimum-demand charges can reduce underutilization risk. Milestone payments can rise as investments become less reversible. Phased energization can link capacity releases to financing, permits, equipment delivery, and verified construction progress. Separate tariffs can distinguish firm service from interruptible or non-firm service.
These mechanisms are not substitutes for one another. A construction contribution addresses capital cost but may not cover future supply commitments. A minimum bill addresses low utilization after service begins but may not compensate for cancellation beforehand. Flexible service can reduce peak-driven investment only when curtailment rights, telemetry, notice, duration, testing, and consequences for nonperformance are enforceable.
Water, emissions, and flexibility follow the project
A 2025 Nature Sustainability analysis estimated that US AI servers could produce an annual water footprint of 731 million to 1.125 billion cubic meters and 24 million to 44 million metric tons of additional carbon-dioxide-equivalent emissions by 2030 under its modeled scenarios. The results vary with siting, grid carbon intensity, cooling design, and efficiency; they are scenarios, not observed outcomes.11
A 2025 life-cycle assessment led by Microsoft and WSP researchers found that advanced cooling approaches reduced greenhouse-gas emissions by 15 to 21 percent, energy demand by 15 to 20 percent, and blue-water consumption by 31 to 52 percent against the study’s air-cooled baseline. Those are design-specific results, not universal guarantees.12
Operational flexibility can also have system value, but evidence must be described accurately. A 2026 study modeled firm, pause, and workload-shifting configurations on a synthetic 2,000-bus Texas system. In the model, flexibility expanded the feasible siting frontier by 9 to 17 percent for one gigawatt of load and 19 to 21 percent for two gigawatts. The authors estimated that pre-certified sites could reach initial energization in 12 to 18 months. Those findings are modeled and depend on standardized flexibility envelopes; they are not evidence of an observed nationwide interconnection program.13
Flexibility has infrastructure value only when a facility can specify how much load it can reduce, for how long, with what notice, through what telemetry, and under what penalties. A voluntary assertion that computing is flexible is not equivalent to a dispatchable contractual obligation.
The emerging standard is project evidence
The durable policy frame is not whether AI is inherently beneficial or environmentally harmful. A data-center proposal is an infrastructure project whose value and risk depend on location, ramp rate, triggered facilities, supply resources, operating behavior during stress, and liability if expectations fail.
A credible review should therefore disclose five things before major commitments become irreversible: a verified load profile; a cost-allocation map; an enforceable flexibility envelope; a location-specific environmental account; and a public-benefit and exit plan.
The expanded federal pledge gives the first principle—cost responsibility—national political force. FERC’s proceedings may translate parts of large-load integration into enforceable regional tariffs. Scientific research is making the same shift by replacing generic comparisons with location- and system-specific analysis.
The environmental significance of a prompt is real but variable. The public consequences of a data center are physical, contractual, and local. They are determined long before the prompt is typed.
Notes and Sources
1. The White House, “President Trump’s Ratepayer Protection Pledge Secures American AI Dominance, Protects Consumers,” July 23, 2026. https://www.whitehouse.gov/releases/2026/07/president-trumps-ratepayer-protection-pledge-secures-american-ai-dominance-protects-consumers/
2. The White House, “Fact Sheet: President Donald J. Trump Advances Energy Affordability with the Ratepayer Protection Pledge,” March 4, 2026; and “Ratepayer Protection Pledge,” March 4, 2026. https://www.whitehouse.gov/fact-sheets/2026/03/fact-sheet-president-donald-j-trump-advances-energy-affordability-with-the-ratepayer-protection-pledge/
3. International Energy Agency, Energy and AI (Paris: IEA, 2025), “Energy Demand from AI.” https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
4. Sarah Josephine Smith et al., United States Data Center Energy Usage Report: 2025 Update (Berkeley, CA: Lawrence Berkeley National Laboratory, June 2026). https://eta.lbl.gov/publications/united-states-data-center-energy-2025
5. Electric Power Research Institute, Powering Intelligence: Updated U.S. Data Center Scenarios (Palo Alto, CA: EPRI, 2026). https://powering-intelligence.epri.com/executive-summary.html
6. Alexandra Sasha Luccioni, Yacine Jernite, and Emma Strubell, “Power Hungry Processing: Watts Driving the Cost of AI Deployment?,” Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 85–99. https://doi.org/10.1145/3630106.3658542
7. Felipe Oviedo et al., “Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute,” Joule (2026): 102430. https://doi.org/10.1016/j.joule.2026.102430
8. US Energy Information Administration, “Commercial Electricity Sales Have Soared in Virginia, Driven by Data Centers,” May 5, 2026. https://www.eia.gov/todayinenergy/detail.php?id=67664
9. Kavya Balaraman, “US Power Companies Scramble to Secure Equipment as Surging Data Center Demand Strains Supplies,” Reuters, July 9, 2026. https://www.reuters.com/business/energy/us-power-companies-scramble-secure-equipment-surging-data-center-demand-strains-supplies-2026-07-09/
10. Federal Energy Regulatory Commission, “FERC Launches Aggressive Targeted Action to Speed Large Load Integration,” June 18, 2026; see dockets EL26-67-000 through EL26-72-000. https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targeted-action-speed-large-load-integration
11. Tianqi Xiao et al., “Environmental Impact and Net-Zero Pathways for Sustainable Artificial Intelligence Servers in the USA,” Nature Sustainability 8 (2025): 1541–53. https://doi.org/10.1038/s41893-025-01681-y
12. Husam Alissa et al., “Using Life Cycle Assessment to Drive Innovation for Sustainable Cool Clouds,” Nature 641 (2025): 331–38. https://doi.org/10.1038/s41586-025-08832-3
13. Dongjoo Kim, Lin Dong, and Le Xie, “Flexibility-Aware Framework for Efficient Planner-Initiated Siting of Data Center,” Nature Communications 17 (2026): article 6512. https://doi.org/10.1038/s41467-026-72324-9
