Every time a tech company trains a new large language model, it draws enough electricity to rival the annual consumption of a small neighborhood. One peer-reviewed lifecycle analysis found that building a single model released 493 metric tons of carbon dioxide equivalent, matching the yearly power use of about 98 American homes, while consuming nearly 2.8 million liters of water. Those numbers keep climbing as models grow larger, and the power industry is now warning that aggregate demand from AI training is approaching gigawatt-scale levels.
Grid strain from AI training runs is accelerating
The tension behind these figures is straightforward: models are getting bigger, but the energy cost of each additional parameter is not shrinking fast enough to offset the growth. A research team that tracked the full model-creation lifecycle reported that the process emitted 493 metric tons of CO2 equivalent and consumed 2.769 million liters of water. That carbon output equals what roughly 98 U.S. households produce in a year through normal electricity use. The finding captures not just the final training run but also the upstream experiments, data processing, and hardware manufacturing that feed into a finished model.
If efficiency gains per additional parameter have plateaued since 2022, as hardware benchmarks and published training logs suggest, total grid load will scale in rough proportion to parameter count. Doubling a model’s size will roughly double the electricity bill unless chip designers or training schedulers deliver a step-change improvement. No widely deployed breakthrough of that kind has arrived yet, which means every generation of frontier models places a heavier load on the same power infrastructure that serves homes, hospitals, and factories.
Primary research quantifying AI energy costs
Three bodies of primary research anchor the current understanding of how much electricity large AI models consume. The earliest systematic treatment came from a 2019 paper on energy and policy considerations for deep learning in NLP, which laid out a framework for measuring compute costs in dollars, kilowatt-hours, and CO2 equivalents. That work showed how hyperparameter search, the trial-and-error process of tuning a model before its final training run, can dominate total energy use, sometimes exceeding the cost of the training run itself by a wide margin.
A 2021 study on carbon emissions and large neural network training extended the analysis to GPT-3 and similar models. Its authors warned that back-calculating energy use without detailed disclosures on datacenter efficiency, hardware specifications, and geographic location can produce unreliable estimates. That caution still applies: most frontier labs disclose little about the power sources, cooling systems, or chip configurations behind their largest training runs, making independent verification difficult.
The most recent lifecycle study, published on arXiv in early 2025, went further by accounting for the entire creation process rather than the final training pass alone. Its 493-metric-ton CO2e figure and 2.769-million-liter water-consumption total represent the broadest scope yet applied to a single model’s environmental footprint. Together, these three papers trace a clear trajectory: as models have grown from hundreds of millions of parameters to hundreds of billions, the energy cost per model has risen by orders of magnitude.
The Electric Power Research Institute and the research group Epoch AI reinforced that trajectory with a joint report on surging power demand from AI model training. The report describes demand reaching gigawatt-scale levels, a threshold that would place AI training alongside heavy industrial loads like aluminum smelting or steel production in terms of grid impact.
Gaps in disclosure and what to watch next
Several questions remain open. The EPRI and Epoch AI report flags gigawatt-scale demand, but the full underlying datasets and modeling assumptions have not been released beyond the press summary. Without those details, independent analysts cannot verify the specific growth curve or the assumptions about future hardware efficiency baked into the projection.
No public primary records detail exact energy figures for models trained after mid-2024. Frontier labs including OpenAI, Google DeepMind, and Anthropic have not published the kind of granular datacenter-level data that the 2021 carbon-emissions paper identified as necessary for reliable estimates. The water-consumption methodology from the 2025 lifecycle study also lacks third-party verification, meaning the 2.769-million-liter figure, while carefully documented by its authors, has not been independently replicated.
For utilities and grid operators, the practical consequence is immediate. New data center construction is outpacing generation capacity in several U.S. regions, and each new frontier model training run can draw tens of megawatts for weeks or months at a stretch. Readers who pay electricity bills in data-center-heavy regions, or who invest in utility stocks, should track two developments closely: whether chip makers deliver meaningful efficiency gains in the next generation of AI accelerators, and whether regulators begin requiring energy and water disclosure from large-scale model trainers. Both factors will determine whether the 98-homes-per-model benchmark stays roughly stable or climbs toward the hundreds.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
