Hyperscale data centers used 2% of US electricity in 2024-25
A study estimates that 403 US hyperscale data centers used 68–99 TWh in May 2024–April 2025, or 1.5%–2.2% of national electricity, but could not isolate AI workloads.

By James Calloway
An estimated 403 U.S. hyperscale data centers consumed 68 to 99 terawatt-hours (TWh), a unit of electrical energy, between May 2024 and April 2025, according to a new study. The researchers put that consumption at 1.5% to 2.2% of total US electricity use, roughly the amount used by Wisconsin or Indiana. That scale matters because the data-center buildout has left the industry without a settled plan for clean electricity supply and its effects on electricity customers, commonly called ratepayers.
What did the study actually measure?
The research team tallied energy use and estimated emissions for hyperscalers operating in the United States between May 2024 and April 2025. Hyperscalers operate very large data centers, while the study also included facilities that could not be confirmed as dedicated AI sites. The summary does not classify either measure by AI-only facility or workload.
The central limitation is explicit. The study says researchers “weren’t able to differentiate between AI compute and more traditional data services.” As a result, the 68 to 99 TWh estimate covers the selected data centers as a whole rather than isolating electricity consumed by AI.
The study also produced emissions estimates, but its summary provides no emissions total or emissions-per-facility figures. It therefore cannot support a separate claim about the emissions associated with AI computing. Readers can use the electricity range as a hyperscale baseline, but not as an AI emissions estimate.
Practically, this is a national hyperscale baseline, not an AI-specific electricity or emissions total.
How did researchers identify 403 facilities?
The team started with 675 facilities listed in a commercial database assembled from news reports, company statements and satellite imagery. Researchers then applied two filters intended to remove entries that did not meet the study’s definition of a large data center. This method produced a verifiable count where previous reporting had struggled to establish what was being built.
They excluded facilities without a reported electrical capacity. They also removed sites they could not verify as large buildings with the expected cooling and electrical connection equipment. Those exclusions covered 272 facilities believed to be large data centers.
The surviving inventory contained 403 data centers with an average electrical capacity of about 40 megawatts (MW), a measure of power capacity rather than annual energy consumption. The average does not show how capacity was distributed across the sites or how intensively each facility operated during the study period.
The research summary does not name the commercial database or provide the full facility list. It also gives no minimum, maximum or median capacity, limiting independent checks on the 40 MW average and the representativeness of the sample.
For readers, the inventory is broad but incomplete, and its average capacity does not describe what each site actually consumed.
Why does the study call its estimate conservative?
The research team calls the result “a very conservative estimate” because it excluded 272 facilities believed to be large data centers. Some lacked a reported electrical capacity, while others could not be verified through their physical infrastructure. The omission supports the team’s characterization of the inventory as conservative.
That label describes the boundary of the facility inventory, not perfect certainty about every calculation. The summary reports no separate power-use estimate for the excluded sites. Their contribution therefore remains outside the published 68 to 99 TWh range.
The AI classification creates a different uncertainty. Some of the 403 included facilities may not operate AI data centers specifically, according to the study. Treating the entire aggregate as AI consumption would therefore attribute electricity use from conventional services to AI.
Two caveats point in opposite directions. Missing large facilities can understate total hyperscale consumption, while including conventional sites can inflate an AI-specific figure if the total is mislabeled. The aggregate hyperscale estimate does not need either correction, but any attempt to assign it to AI does.
The estimate probably undercounts hyperscale sites, but that does not make the separate AI share any more precise.
How uncertain is the power-use range?
Reported electrical capacity is potential output, not a direct measure of electricity consumed. Actual demand depends on how heavily a facility operates. Researchers can account for that difference through a load factor, defined here as the share of reported capacity in use.
The study cites other research for two operating assumptions. Conventional data centers have run at an average load of about 48 percent of capacity. AI training loads have run higher, at around 70 percent of capacity.
The team combined those external benchmarks with the capacity data for 403 facilities. That calculation produced the annual range of 68 to 99 TWh. The summary does not publish the step-by-step conversion, facility-level assumptions or exact weighting assigned to the two load benchmarks.
The range therefore carries more uncertainty than a single percentage would suggest. Some included sites could support conventional services, some could support AI training, and the summary does not allocate electricity between those categories. It also does not show how much of the difference between the range’s endpoints came from operating loads rather than facility selection.
The range is a modeled estimate whose uncertainty starts with the mix of workloads inside the inventory.
Why does the AI attribution gap matter?
One flashpoint for the technology industry’s push toward Large Language Models has been the rush to convert investment into hyperscale data centers. According to the source, that rush has not allowed the industry to plan how to supply the required electricity cleanly or with minimal impact on ratepayers. Even basic information about what was being built has been difficult to collect.
The study offers a point of reference that can be tracked as more facilities come online. Its 403-site inventory establishes a dated baseline for both capacity and estimated electricity use. A later inventory using comparable methods could show whether hyperscale consumption is rising faster than the wider electricity market.
That comparison would still require careful interpretation. A higher hyperscale total could reflect more facilities, greater operating loads, a different facility mix or a combination of those changes. Without workload-level data, it would not establish how much of the increase belonged to AI rather than traditional data services.
The inventory also offers limited help with emissions planning because the summary gives no emissions result. It therefore cannot show whether cleaner electricity supply, new generation or other interventions would meet the sector’s projected demand.
For anyone assessing the buildout, the study supplies a scale reference, but not the AI allocation needed for a complete energy and emissions assessment.
What does the research summary leave unanswered?
The summary provides enough information to establish the headline range and the basic filtering method. It leaves several material questions open:
- Workload allocation: How much of the 68 to 99 TWh came from AI compute, and how much came from more traditional data services? The research team says it could not distinguish between them.
- Emissions: What emissions total did the study estimate, and what method did it use? Neither figure appears in the research summary.
- Inventory access: Which 403 facilities were included, and which 272 candidates were excluded? The summary does not provide the facility list or name the commercial database.
- Calculation detail: How did the 48 percent and 70 percent load assumptions translate into the final lower and upper estimates for each facility? The step-by-step method is not disclosed.
- Study identity: The research summary does not identify the authors, their institutions or the publication carrying the study.
These omissions do not erase the useful baseline, but they constrain independent verification and prevent precise attribution of the environmental burden to AI. The most important unresolved issue is not the overall size of hyperscale demand. It is the inability to say how much of that demand belongs specifically to the AI boom driving construction.
The study can anchor follow-up work without yet supporting a complete AI accounting.
Can future inventories build a better baseline?\n
A repeat inventory could turn this one-year estimate into a trend line rather than a single data point. The first analytical test would be whether the 403 verified facilities consumed more electricity during a comparable later period, while new sites entered the inventory.
A stronger follow-up would need to recover the 272 excluded facilities and establish their electrical capacity. It would also need to classify workloads more precisely and disclose the emissions estimate, inventory sources and calculation method. Without those steps, another aggregate would reproduce much of the present uncertainty.
The source makes no commitment to such a follow-up, and the research team has not said whether it will update the inventory. It also provides no schedule for adding the excluded sites or separating AI from conventional services. Those details would determine whether the baseline becomes a durable tracking system or remains a one-time estimate.
The next meaningful test is whether later inventories close those gaps; until then, 2% is a baseline for hyperscale sites, not a measured AI total.
Source: arstechnica.com
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
722 articles
Related articles
- Fossil Fuels Supplied 54% of Power to US Hyperscale Data Centers
- 403 U.S. Data Centers Used as Much Electricity as Michigan
- 403 largest U.S. data centers rivaled Denmark’s emissions
- Google's Emissions Jumped 50% in 2024 to Feed Generative AI
- Voters Approved 40 of 41 Ballot Measures Restricting AI Data Centers