Electricity grids worldwide are straining under a new kind of load. Global data centers are on track to consume roughly 1,000 terawatt-hours of electricity this year, a figure driven largely by the rapid expansion of artificial intelligence training and inference workloads layered on top of existing cloud computing demand. That volume of power is comparable to the total annual electricity consumption of Japan, and it forces grid operators, policymakers, and facility designers to confront a basic question: can efficiency gains keep pace with AI-driven growth before power shortages and missed emissions targets become unavoidable?
AI workloads are pulling data-center power demand past historical ceilings
The 1,000 TWh threshold did not arrive overnight. For more than a decade, efficiency improvements in servers, storage, and cooling kept global data-center electricity use relatively flat even as computing output surged. That dynamic has shifted. The International Energy Agency’s analysis of electricity demand ties the acceleration directly to generative AI models, whose training runs and inference queries demand far more power per operation than conventional cloud tasks. Each new generation of large language models requires denser GPU clusters, longer training cycles, and round-the-clock inference serving, all of which translate into sustained electrical load that older efficiency trends cannot absorb.
In that outlook, data centers emerge as one of the fastest-growing sources of electricity demand globally, alongside electric vehicles and heat pumps. But unlike those technologies, data-center loads concentrate in a handful of regions, primarily the United States, parts of northern Europe, and east Asia, creating acute pressure on local grids. In the U.S., Lawrence Berkeley National Laboratory’s 2024 reporting on data-center energy use supplies much of the domestic baseline that feeds into the IEA’s global totals, documenting how American facilities account for a disproportionate share of worldwide consumption. The European Commission, drawing on similar figures, has described data centers as an “energy-hungry challenge,” warning that unchecked growth could complicate national decarbonization plans.
The concentration of demand matters for practical reasons. When a single hyperscale campus draws hundreds of megawatts in a region that was already near capacity, utilities face choices between accelerating new generation, curtailing other customers, or delaying interconnection. Northern Virginia, central Texas, and parts of Ireland have all experienced versions of this bottleneck, where transmission constraints and permitting timelines collide with AI buildout schedules. The 1,000 TWh figure, while global, masks these localized crises where the gap between supply and demand is already binding and where new data-center proposals are increasingly contingent on grid upgrades or on-site generation.
Cooling design holds the biggest untapped efficiency lever
One of the clearest paths to slowing the electricity growth curve runs through cooling, which can account for 30 to 40 percent of a facility’s total energy use depending on local climate and system design. A peer-reviewed study in Resources, Conservation and Recycling developed a hybrid statistical and thermodynamics-based method for estimating efficiency metrics across different U.S. climate zones and data-center archetypes. The research showed that Power Usage Effectiveness (PUE), the ratio of total facility energy to IT equipment energy, varies significantly by geography and cooling technology. Facilities in cool, dry climates using free-air economizers can achieve PUE values well below 1.2, while those in hot, humid regions relying on traditional chilled-water systems often exceed 1.5.
That spread matters enormously at scale. If operators building new capacity in warm climates adopted designs validated for their specific climate zone, such as direct liquid cooling for high-density AI racks or hybrid evaporative systems tuned to local humidity, the aggregate effect could meaningfully slow the electricity growth curve even as AI workloads expand. The idea that climate-zone-specific cooling designs deployed at scale could flatten data-center electricity growth by 15 to 20 percent within a few years is plausible on paper, but it has not yet been tested against post-2024 operational data. Most published PUE and water-use benchmarks rely on modeling and pre-AI-boom facility profiles, and the shift to GPU-dense racks changes heat-density assumptions that older studies used.
Cooling choices also interact with water and land constraints. Evaporative systems can deliver excellent energy efficiency but may be politically untenable in water-stressed regions. Direct liquid cooling reduces fan energy and allows tighter rack packing, but it can require retrofits to existing buildings and new maintenance practices. These trade-offs mean that the “best” design is highly context dependent, reinforcing the importance of climate-aware engineering over one-size-fits-all templates.
Lawrence Berkeley National Laboratory’s work on recalibrating estimates of global data-center energy use provides much of the methodological backbone for these projections. The bottom-up approach used by LBNL researchers, which aggregates facility-level characteristics rather than relying solely on top-down utility data, has been adopted by the IEA and other institutions as a baseline for global accounting. But the recalibration also reveals how sensitive total estimates are to assumptions about facility mix, utilization rates, and cooling overhead. Small changes in average PUE across the global fleet produce swings of tens of terawatt-hours in the final number, underscoring why cooling design is such a powerful lever.
Gaps in metering and regional data cloud the outlook
No primary global real-time metering dataset exists for data-center electricity consumption. The 1,000 TWh figure and its variants are modeled extrapolations built from bottom-up surveys, utility filings, and equipment shipment data. The IEA and LBNL have refined these methods over multiple cycles, but the estimates still carry wide confidence intervals, especially outside North America and western Europe where facility-level transparency is limited. Rapid AI buildouts in regions with sparse reporting, such as parts of southeast Asia, further complicate efforts to track the true pace of change.
Within individual facilities, operators often monitor power at a granular level, but those data are rarely shared publicly and are sometimes treated as commercially sensitive. Where disclosure does occur-through sustainability reports or voluntary benchmarking programs-it may lag by a year or more and may not distinguish between AI-specific clusters and conventional cloud workloads. As a result, policymakers trying to plan grid investments or set realistic emissions trajectories are working from a patchwork of partial signals rather than a coherent global picture.
These data gaps have practical consequences. Utilities weighing whether to approve new interconnections must estimate not only peak load but also the likelihood that AI clusters will run near full capacity for extended periods. Overestimating risks can lead to overbuilding and stranded assets; underestimating can produce reliability problems and emergency curtailments. Without better visibility into how AI workloads evolve over time-how quickly models are retrained, how inference demand grows, and how aggressively operators pursue efficiency-those decisions remain fraught.
Closing the information gap will require a mix of policy and industry action. Governments can mandate more detailed reporting on large energy users, including standardized PUE disclosures and clearer breakdowns of IT versus non-IT loads. Industry consortia can expand anonymized benchmarking programs that allow operators to compare performance without exposing proprietary details. Researchers, in turn, can refine bottom-up models with updated assumptions about GPU densities, cooling technologies, and regional build patterns.
The stakes extend beyond accounting accuracy. If AI-driven data-center growth continues on its current trajectory without corresponding efficiency gains, it will complicate efforts to decarbonize power systems and could crowd out electrification in other sectors. Yet the same modeling work that raises alarms also points to a path forward: targeted improvements in cooling, smarter siting in favorable climates, and more transparent reporting could keep the electricity curve from bending too sharply upward, buying time for grids to adapt while the AI era takes shape.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
