A study published in Nature Sustainability on November 16 estimates that 403 US hyperscale data centers used 81.8 terawatt-hours of electricity from May 2024 through April 2025. Gianluca Guidi and colleagues put associated carbon dioxide emissions at about 25.7 million metric tonnes in their central case. They describe that electricity use as about 1.8 percent of US consumption, with Virginia alone accounting for roughly one-quarter of the studied load. The sites were in the contiguous United States. The estimate is for the identified operating sample over that year. Nature Sustainability paper · Publisher’s research summary

The important contribution is the link between the location of a load and the generation mix used to describe it. An annual national number gives a sense of scale. The facility-level approach can show which states account for that electricity and how much the estimated footprint changes when the grid has a different mix. The uncertainty is just as important. Electricity was modeled from facility information and load assumptions. Emissions were assigned using regional averages. Readers can use the result to judge the size and concentration of the observed build-out. A specific site's demand at a particular hour remains a different question.

What the researchers counted#

Hyperscale facilities hold substantial computing and storage capacity for cloud services, data processing and AI. The research starts with facility information and narrows it to a set whose locations and attributes could be validated for the analysis. The authors’ reproducibility materials mention an initial universe of 675 facility records and a 403-facility analytical set. Those counts describe the collection and filtering process. Neither is a census of every American data center. Even among buildings with substantial computing equipment, a project that was announced, a building that was under construction and an active site contribute differently to the energy use of a given year. Reproduction documentation and coverage notes

That perimeter matters because this is a historical estimate. A new facility can be large on a utility filing, irrelevant to the study’s year and important to a grid forecast. A site mentioned in an old directory may have changed its equipment or operating status. Validating locations and dates makes it possible to tie observed infrastructure to an electricity estimate. It also leaves some cases outside the available dataset. A selection restricted to verifiable records can be useful even when it leaves out facilities whose location, size or status cannot yet be corroborated. A reader should keep the number of analyzed sites and the year attached to the resulting energy estimate.

One way to see the distinction is to think of the grid as having an operational list and a planning list. The first might have sites with substations already in use. The second might include some of those and several years of possible additions. A project can occupy both lists while its load grows from an initial hall to a later phase. A planning list needs probabilities and start dates. A historical list needs actual operating windows and a measure of draw for the year in question. Mixing the two gives an apparent time series that rises as developers announce projects and falls as they cancel them, with no equivalent changes yet taking place at the meter.

The authors’ public repository makes this boundary explicit. It includes model code, aggregate state and balancing-authority outputs and synthetic inputs. A data-use agreement prevents publication of identifying inputs for the real facilities. The synthetic rows allow a user to check the machinery for processing data and validating its modeling code. They do not turn into the hidden facility observations. Aggregate totals and public factors can be checked independently, while full replication on the original sample requires access to the restricted inputs. That is a useful distinction when readers evaluate what the paper’s public materials let them verify. Author repository and data availability

It also matters why a record could not be validated. Facilities with a publicly visible industrial footprint, a well-documented operator and a long history may be easier to confirm than recent or commercially quiet sites. Missing data could therefore change the composition of a sample in ways that a simple count does not show. An operator with ten recorded buildings may have a different average electrical capacity from one with a single building carrying its name. An independent survey could help by reporting which characteristics are known for unverified candidates and where coverage is strongest. That would allow a reader to interpret the 403-site estimate without assuming that the remaining facilities have the same distribution of sizes and locations.

Turning electrical capacity into energy#

Facility electrical capacity describes a rate in megawatts. Electricity consumed over a period is energy in megawatt-hours or terawatt-hours. The study chooses several facility-load coefficients to model how much of capacity the facilities draw on average. Its central coefficient is 0.58. At 0.48, the authors’ public results give 67.7 TWh and 21.3 million tonnes of carbon dioxide. At 0.58 they give 81.8 TWh and 25.7 million tonnes. A sensitivity case at 0.663 gives 93.5 TWh and 29.4 million tonnes. The high-load case at 0.70 gives 98.6 TWh and 31.0 million tonnes. The span shows what happens under those assumptions, rather than the probability of any particular total. Final scenario results

Imagine a facility with 100 MW of usable capacity under the simplified model. Holding the load at 58 percent over a 365-day year would give roughly 508,000 MWh. Holding it at 48 percent would give around 420,000 MWh. This is an illustration of the denominator, not an observed facility from the study. It shows why an error in the load assumption can matter as much as an error in the capacity figure. The coefficient describes facility electrical use. It cannot be interpreted as the fraction of time every server is busy, the fraction of workloads that involve AI, or the sector’s share of the power drawn during the worst hour for a utility.

A server rack with two load assumptions and different annual electricity uses.

Annualizing the central estimate gives about 9.34 gigawatts of average demand across the selected sites. That is simply 81.8 TWh spread evenly across the hours in twelve months. It helps relate the energy number to a rate of use. The same average could accompany a relatively steady load or a more variable profile. A utility evaluating a local connection needs a different set of facts, including the expected power during system peaks, how quickly the load ramps up and what happens if a large load trips. A national annual average can anchor a scale comparison. In a connection study, the grid’s location and time of stress determine which power number matters.

The scenarios also address a common problem with older comparisons. Changes in the use of server hardware can push facility draw in different directions. The workload mix and installation schedule matter. A center can operate while still adding servers. A newer data hall can have less overhead per unit of computing but a higher total draw. A figure for facility electricity does not reveal the useful computation completed or the overhead spent on cooling. Where those distinctions matter, the denominator should be explicit. Capacity, load, power usage effectiveness and electricity per completed job serve different readers. Applying a server-utilization assumption as if it were the observed utilization of an entire building can hide important differences.

In a large facilities study it makes sense to use a transparent set of demand scenarios when direct readings are unavailable. It remains possible that the error is uneven across regions. A fast-growing cluster could be running below established capacity because new halls have opened. Another cluster might have nearly full older buildings. Applying the same coefficient to both yields a consistent model but cannot show these differences. Metered totals for a representative group could test whether the central assumption is reasonable overall and whether specific regions need separate treatment. That kind of validation would also show whether the scenario spread reflects the main uncertainties or misses an important source of variation.

The denominator that moved the emissions number#

The final study uses the US Environmental Protection Agency’s eGRID 2023 Revision 2 for annual generation and emissions attributed through balancing authorities. Under the total-output method, carbon dioxide reported for generation is divided by all the reported net electricity generated in that balancing authority. That includes generation with no direct carbon dioxide emissions at the power plant, such as nuclear and many renewable sources. Multiplying the resulting regional factor by modeled facility use provides the study’s attribution. This explains why the same load has a different estimated footprint in regions with different generation mixes. EPA eGRID · Authors’ emissions method and denominator

One earlier version of the work reported a larger range, approximately 37 to 54 million tonnes, for a similar electricity range. The final repository documents a change in the emissions denominator during revision. Its combustion-output diagnostic divides by electricity generated from combustion. Applied as a rate to all consumed electricity, that denominator implies a higher result than a factor that also counts the nuclear and renewable generation in the region. The published headline uses the total-output factor. The change is not an observed decline in physical emissions between two years. It is a correction to how the same modeled electricity use is turned into a carbon estimate. June preprint · Authors’ final alignment notes

In the final figures the load-weighted average factor for the sample is about 314 grams of carbon dioxide per kilowatt-hour. The authors give 348 grams for the US eGRID total-output reference. Dividing 25.7 million tonnes by 81.8 TWh yields about 314 grams using the rounded central values. The sample’s location-weighted rate is thus below that national factor on the same basis. That result can coexist with high absolute emissions because the sites have a large total volume of use. A lower average factor may be consequential when deciding where a new facility could operate. It does not determine whether any particular new load would have the same factor. Final emissions comparison

The authors give a mix of approximately 54 percent fossil, 21 percent nuclear and 25 percent renewable electricity when they associate facilities with regional generation. These are modeled shares on an accounting boundary. The researchers do not trace an individual wire out of a plant to a particular facility. A region can import and export power, and generation mix changes through the day. A large buyer may have a contract for an energy attribute. Company disclosures may apply those contracts under a separate market-based method. A physical question about the generation needed to serve a change in demand calls for data on dispatch, power flows and timing. The study’s average factors are suited to the narrower question it asks about attribution.

The use of a 2023 grid dataset also deserves to stay with the values. The estimated loads are from a 2024–2025 window and their emissions are modeled with the specified eGRID vintage. A new power plant, grid interconnection or retirement can change the grid mix. Comparing the study to a later estimate works best when the vintages, regions and methods are published. Sometimes a difference demonstrates a real change. Sometimes it arises because an updated dataset gives a new view of a similar set of physical activities. Keeping the grid data version and the window of facility activity visible gives the reader a way to distinguish the two. Final analytical basis

An electricity emissions estimate also has a boundary around what it counts. Constructing a building and manufacturing the servers create other emissions. Operating a generator at a campus would require knowing how the fuel and electricity are accounted for in that case. Cooling has implications for electricity and can also affect water use, another resource with very local conditions. These issues are important to a complete environmental assessment. The 25.7 million tonnes describe attributed carbon dioxide from the modeled electricity use and cannot stand in for a lifecycle footprint or a local exposure measurement. A more comprehensive comparison can build on this work by reporting additional quantities separately with their methods and owners.

Geography explains the national number#

Virginia contained 142 of the selected facilities. The publisher says the state accounted for about 21 TWh of the central load, roughly a quarter of the 81.8 TWh total. Oregon had 56 selected sites and Ohio 38. Counts and energy volumes need to be read together. With roughly 35 percent of the facilities and about 26 percent of modeled electricity, Virginia’s share of the site count is greater than its share of the estimated load in this sample. This is a descriptive comparison, from one period and its selection. It shows why the number of facilities in an announcement is a poor substitute for their capacity or annual draw. Publisher’s state breakdown

States and power areas supply different boundaries. A state accounts for jobs, land use and some permitting decisions. A balancing authority manages its assigned power system, which can cover parts of several states or smaller regions. A nearby community can see the effects of a particular interconnection or generator that get blurred in larger aggregates. The study helps because it puts facilities and power factors in specified regions. A reader evaluating local decisions still needs more detail about when electricity is consumed, what generation or transmission has been added and where the costs of connecting a large load will fall. The national fraction is a reason to look for those figures.

Some coverage put the study alongside comparisons with the electricity used by an entire state or country. The comparison helps readers visualize a large energy unit when both figures measure annual use. It can obscure the industrial geography. A concentration of loads near one part of a network is a different planning problem from an equal annual total dispersed across towns. A facility using 100 MW for a long interval is also different from an intermittent peak in a system that can shift load. When readers see the country comparison, the next question is which point about the data center industry the writer wants it to establish. Energy scale, grid pressure and emissions intensity should each retain their evidence. Associated Press coverage

The timing of the study's publication brought it into a broader discussion of data center power and climate commitments. The International Telecommunication Union and COP31 presidency presented a voluntary AI pledge the same week, including measures related to infrastructure energy and reporting. The study gives one empirical reason to pay attention to reporting boundaries. A statement about clean energy can be tested with the region, period, generation source and claimed basis made explicit. An electricity figure can be used to track progress only if a later reader can tell whether it covers the same facilities and uses the same measuring unit. The published modeling and the separate pledge have different roles in that discussion. ITU description of the Antalya AI Pledge

How the estimate fits with larger surveys#

Lawrence Berkeley National Laboratory has modeled the electricity use of the broader US data center sector. Its June 2026 update modeled reference-case electricity consumption of 649 TWh in 2030, around 11.8 percent of projected US electricity. That is a forward scenario about a wider universe that responds to assumptions about equipment shipments, demand and design. It cannot be read as an observed successor to 81.8 TWh, which is an estimate of the 403 identified sites for a historical year. A ratio between those figures would compare multiple boundaries at once. An appropriate growth comparison would require a consistent group, baseline date and model or a careful bridge between changes. Berkeley Lab’s national energy-use update

The Department of Energy’s Energy Information Administration launched a data center energy-use survey pilot in March 2026. A survey can ask operators about measurements, equipment and sources in ways that a public-directory study cannot. The first challenge is to get comparable responses across differently defined facilities, especially when some report a part of a building and others report a campus or a portfolio. Respondents also need to know whether the question asks for actual annual use, peak contracted capacity or projected load. Independent estimates remain useful during that process. They can flag regions where local measurement should be checked and give readers a way to ask how new results fit the public operating picture. EIA pilot survey announcement

A useful future comparison would show what happened to the already verified sites first. Did they use more electricity because existing halls became busier or because another phase entered service? A separate count would handle sites added to the sample. A change in emissions could then be split into the change in electricity demand and the change in the factors applied to it. Additional information on on-site generation and the difference between annual and hour-level matching would answer questions that cannot be settled with annual regional factors. This kind of bridge would turn a pair of headline numbers into a test of which part of the system changed.

Better data can also make uncertainties more specific. Publishing capacity and annual metered use on a consistent perimeter would let readers compute a load coefficient instead of borrowing a scenario. Publishing several hourly demand profiles for representative facility types would show when pressure falls on the network. Disclosing the rules for what remains confidential can help separate missing facts from details that simply should not be public. Aggregate reporting is often sufficient to evaluate a region’s energy trend. Maintaining a way for qualified reviewers to check the underlying information helps make the aggregate credible without creating a public inventory of sensitive operating details.

A reporting format could follow a simple sequence. State which sites are operating, identify the period and units, and report the method and grid factor. Label whether the power number represents a commitment, a peak or an average. Provide an annual energy total and, where available, the range of actual loads during times when the local grid is most constrained. Then explain how revisions to the site set or factor changed earlier comparisons. That creates an account a reader can follow over time. It also gives regulators and operators a way to reconcile their views when one speaks about future infrastructure and another speaks about consumption that has already occurred.

The study’s conclusion is strongest when stated with all three parts attached. It models about 82 TWh of consumption by 403 working facilities in one year, assigns 25.7 million tonnes of carbon dioxide using regional total-output factors, and shows how sensitive the estimates are to facility load. That is large enough to demand a more consistent accounting of the industry. It is also precise enough to show which facts are still missing. The next convincing study will make it easier to distinguish a change in the number of facilities, a change in their actual use and a change in the energy mix that supplies them.