A No-Hype Framework for Reading AI Electricity Forecasts

A forecast is not a promise, and an announced data centre project is not a meter reading. Yet claims about AI electricity often place historical estimates, developer requests and modelled futures on the same chart as if they carried equal certainty. The result can exaggerate both precision and inevitability. Use the following audit before repeating a headline number or making a budget, grid or policy decision. It is designed for public reports, utility filings, vendor claims and research papers. The goal is not to choose the smallest estimate. It is to understand what each number represents, which assumptions drive it and what evidence would cause the outlook to change. Step 1: label the evidence before reading the number Put every important value into one of four buckets. An observation comes from a defined measurement, such as a facility meter over a stated interval. An estimate reconstructs an unknown total from partial evidence, models or proxies. A scenario asks what would happen under a specified set of assumptions. A forecast identifies an expected future path, sometimes with probabilities or alternative cases. These labels are not interchangeable. Lawrence Berkeley National Laboratory's 2024 U.S. report estimates historical national data centre electricity use using previous studies and equipment shipment data because there is no single national data centre meter. It then presents a scenario range through 2028. The International Energy Agency similarly describes uncertainty in both current and future consumption and uses a Base Case plus sensitivity cases. Treating the upper edge of a scenario range as an observed trend discards the authors' method. Write a one line evidence label: "estimated U.S. annual data centre energy, reconstructed from equipment data" is far more informative than "AI used X." If the source does not expose a method, boundary and date, the number is not ready for a high confidence decision. Step 2: align units, geography and system boundaries Power and energy answer different questions. A gigawatt is a rate at a moment; a terawatt hour is energy accumulated over time. Nameplate or requested connection capacity cannot be converted to annual energy without an assumed load factor or hourly profile. Peak demand matters for resource adequacy and network capacity, while annual energy matters for generation, fuel and many emissions calculations. Next align geography and scope. A global share can remain limited while one grid region faces a concentrated connection queue. A data centre total includes non AI services unless the study separates workloads. A facility number includes cooling and power conversion; an IT number may not. Upstream generation losses, water use and hardware manufacturing are separate boundaries again. Do not compare two percentages until their denominators match. "Share of total electricity," "share of electricity demand growth" and "growth in data centre electricity" can all be true while describing different quantities. Record the numerator, denominator, place, start year and end year beside every percentage. Step 3: inspect the baseline before the growth rate A dramatic percentage can begin from a small or uncertain baseline. Check whether the starting value is measured or estimated, whether its methodology changed between report editions, and whether cryptocurrency, enterprise server rooms, edge facilities or behind the meter generation are included. Revisions are evidence about uncertainty, not necessarily evidence of error. The IEA's 2025 Base Case estimates global data centre electricity at about 415 TWh in 2024 and projects about 945 TWh in 2030. Its sensitivity cases change the path based on AI adoption, hardware and software efficiency, supply chains and energy bottlenecks. Those assumptions are part of the result. Quoting only the 2030 endpoint strips away the analytical conditions that produced it. Use at least two methods where possible. LBNL begins from chips and other equipment shipments. EPRI's 2026 scenarios use data on operational capacity, construction and announced development. Agreement between differently constructed ranges can increase confidence in broad scale; disagreement can reveal sensitivity to project pipelines, utilisation or equipment assumptions. It does not justify averaging incompatible figures into a supposedly precise answer. Step 4: discount the project pipeline with observable milestones Connection requests can exceed realised demand. Developers may explore multiple sites, delay projects, reduce their scale or never build. Even a completed facility can ramp over years and operate below requested capacity. Forecasts should therefore explain how they screen duplicate or speculative projects and how they translate requested megawatts into likely peak and annual use. FERC's 2025 letter to U.S. grid operators identifies contracts, financial security deposits and physical site control as examples of objective readiness criteria. It also asks how operators prevent double counting and estimate actual use relative to requested service. Those questions form a practical checklist for any pipeline based forecast. ERCOT's 2025 long term load report shows the adjustment process in practice. For its adjusted forecast, the Texas operator applied an average 180 day project delay derived from recent projects. It also reduced non cryptocurrency data centre additions to 49.8% of requested megawatts based on peak consumption at sites entering service from 2022 through 2024. Those values are ERCOT specific empirical adjustments, not universal discount factors. Their importance is methodological: requests were tested against realised behaviour instead of accepted at face value. Step 5: stress the assumptions that can move the result Build an assumption table even if the report does not provide one. Include the pace of AI adoption, amount of computation per task, model and software efficiency, accelerator shipments, server utilisation, facility PUE, construction lead times, grid connection constraints and retirement of older equipment. For pipeline studies, include project completion and ramp rates. For annual electricity, include hourly load shape. Separate efficiency from total demand. More efficient chips or models can reduce electricity per unit of work, while lower cost and wider use can increase the number of units performed. Neither outcome should be asserted without a demand response assumption. Likewise, a supply bottleneck can delay load rather than eliminate eventual demand. The IEA's High Efficiency Case reports materially lower 2035 data centre electricity than its Base Case for the same level of digital service demand, while its Lift Off and Headwinds cases vary adoption and infrastructure constraints. The spread is not noise to be hidden; it identifies decisions and developments worth monitoring. Step 6: turn the range into triggers, not a single verdict For planning, keep a central case and credible lower and upper cases. Then attach observable triggers: signed power contracts, site control, equipment orders, construction progress, energisation, server installation and measured ramp. Update the range as those milestones occur. Record the report version because fast changing development pipelines can cause substantial revisions. Match the decision to the uncertainty. Long lead transmission planning may need to preserve options before every project is certain. Near term procurement should demand stronger readiness evidence. Public communication should state what is measured, what is estimated and what remains conditional. A range should not be presented as though every point is equally likely unless the study says so. Finish with a reproducible forecast card: source and date; measured or estimated baseline; forecast horizon; geography; AI only or all data centres; power or energy; included equipment; project screening rules; utilisation and efficiency assumptions; low, central and high cases; and update triggers. If those fields cannot be filled, communicate the claim as preliminary. The honest conclusion may still be that rapid growth is plausible and infrastructure action is needed. The framework simply prevents plausible growth from becoming false certainty. Good forecasting preserves the difference between evidence already observed, estimates of the present, conditional scenarios and the future path judged most likely. Sources Energy demand from AI https://www.iea.org/reports/energy and ai/energy demand from ai 2024 United States Data Center Energy Usage Report https://eta.lbl.gov/publications/2024 lbnl data center energy usage report Powering Intelligence 2026: Executive Summary https://powering intelligence.epri.com/executive summary.html Chairman Rosner's Letter to the RTOs/ISOs on Large Load Forecasting https://www.ferc.gov/news events/news/chairman rosners letter rtosisos large load forecasting 2025 Long Term Hourly Peak Demand and Energy Forecast https://www.ercot.com/files/docs/2025/04/08/ERCOT 2025 Long Term Load Forecast Report.pdf Cover image credit Cover image by International Energy Agency , made available under CC BY 4.0 . WIKIVISE cropped and converted the source image.