The cloud feels placeless from the user’s side of the screen, but the data center carbon footprint is shaped by very physical things: buildings, power lines, local grids, and the size of the facilities doing the work.
Key findings
- The analysis matched 1,211 identified U.S. data-center facilities to EPA eGRID2023 electricity subregions.
- By facility count, the median data center sits in a grid region with 286.6 g CO₂/kWh, below the U.S. average of 348.0 g CO₂/kWh.
- Only 29.7% of identified facilities are in grid subregions above the national average.
- When facilities are weighted by physical footprint, average carbon exposure rises to 360 g CO₂/kWh, slightly above the U.S. average.
- RFC West and MRO West stand out as regions where meaningful data-center clusters overlap with higher-carbon electricity grids.
When we talk about “the cloud,” we often use language that makes digital infrastructure feel weightless. A photo moves from a phone to a laptop, a company stores its records somewhere offsite, a streaming service delivers a film, and an AI tool answers a question in seconds.
From the user’s side, the experience feels almost atmospheric: always available, hard to locate, and largely detached from the physical world. That convenient invisibility is part of the appeal.
Every one of those services depends on buildings, land, cooling systems, fiber-optic cables, substations, backup generators, and regional electricity grids. The cloud is a large and growing layer of industrial infrastructure embedded inside the economy, even though most of us encounter it only through screens.
Once we see it this way, the climate question becomes richer than simply asking whether data centers use a lot of electricity. We also have to ask where that electricity is being used, because the same kilowatt-hour can carry very different carbon consequences depending on the grid that supplies it.
This is why the data center carbon footprint cannot be understood only as a national or corporate total. A facility in a coal-heavy grid region is exposed to a different carbon reality from one in a region with more hydroelectricity, nuclear power, wind, solar, or lower-carbon gas generation.
The service delivered to the user may look identical, although the electricity behind it is local. The carbon map behind the cloud is therefore a map of real places.
A recent location-based analysis of U.S. data centers helps make that map visible. The analysis matched 1,211 identified facilities from the IM3 Open Source Data Center Atlas with EPA eGRID data on the carbon intensity of regional electricity grids.
In practical terms, each data center was placed on the U.S. electricity map and assigned the carbon intensity of the grid subregion where it is located. The result is more subtle than the usual argument about digital technology and climate change.
If we count each identified facility once, U.S. data centers appear to be located on relatively clean grids. The median facility in the analysis sits in a grid subregion with a carbon intensity of 286.6 grams of CO₂ per kilowatt-hour, compared with a U.S. average of 348.0 g CO₂/kWh.
Only 29.7% of the identified facilities are located in grid subregions above the national average. On a simple facility count, the cloud looks cleaner than many readers might expect.
When the same analysis gives larger facilities more influence, the picture changes. Weighting facilities by physical footprint raises the average carbon exposure to 360 g CO₂/kWh, slightly above the U.S. average.
A simple count of data-center locations suggests a cleaner-than-average pattern, while a size-aware view points toward higher exposure. For a sector where a small colocation facility and a hyperscale campus can both appear as a single “data center,” this is central to understanding what the map is telling us.
Why the data center carbon footprint depends on local electricity
The climate arithmetic behind electricity use is simple enough to write in one line:
Emissions = electricity demand × carbon intensity of electricity
The first half of the equation asks how much electricity a facility uses. The second asks how carbon-intensive the electricity system is in the place where that power is drawn.
For many climate questions, both parts matter. A large facility on a relatively clean grid may still be significant because it uses so much electricity, while a smaller facility on a more carbon-intensive grid may have a higher emissions rate per kilowatt-hour but contribute less in total.
For data centers, the demand side is difficult to study from public information. A company may disclose electricity use for its global operations, and a proposed campus may come with a capacity figure or utility interconnection request, but annual electricity consumption by individual facility is rarely public.
Physical size helps, although only imperfectly. Square footage does not tell us the density of servers, the efficiency of cooling systems, the utilization of computing equipment, or the design choices that determine how much power a facility actually consumes.
The carbon-intensity side is more visible. EPA’s eGRID detailed data estimates the emissions associated with electricity generation across different parts of the U.S. grid.
By combining data-center locations with these grid subregions, we can ask a narrower and still useful question: are identified data centers mostly sitting on cleaner or dirtier parts of the electricity system?
That approach gives us a carbon-exposure map rather than a full emissions inventory. It tells us the carbon intensity of the grid surrounding each facility, which is a necessary part of understanding the climate implications of where digital infrastructure is built.
This matters because data centers are no longer obscure back-office infrastructure. They have become part of public debates over artificial intelligence, local electricity demand, water use, permitting, grid reliability, land use, and corporate climate claims.
If we want to understand the data center carbon footprint with any seriousness, national averages are too blunt. The geography of the grid has to be part of the story.
By count, data centers cluster in relatively cleaner regions
The first pass through the data is somewhat reassuring. Among the 1,211 identified U.S. data-center facilities in the analysis, most are located in grid regions with below-average carbon intensity.
The study began with the IM3 Open Source Data Center Atlas, an OpenStreetMap-derived dataset that includes data-center points, buildings, and campuses. Because buildings can sit inside campus boundaries, the analysis used a campus-aware facility definition to avoid double-counting obvious overlaps.
The final base unit consisted of campuses, buildings outside campuses, and point locations that were not already contained within a building or campus. Each of those 1,211 facilities was then spatially joined to one of 27 EPA eGRID2023 subregions.
The match rate was 100%, meaning every identified facility could be assigned a grid carbon-intensity value. That gave the analysis a consistent geography for comparing locations.
The resulting distribution leans toward moderate and lower-carbon regions. The mean facility-location intensity is 317.8 g CO₂/kWh, below the U.S. average of 348.0 g CO₂/kWh.
The median is lower still, at 286.6 g CO₂/kWh. Roughly seven in ten identified facilities are located in subregions below the national average.
The distribution chart makes the pattern clearer. Of the 1,211 identified facilities, 807 fall in the 200–400 g CO₂/kWh band, while only seven are in regions above 600 g CO₂/kWh.
Several large clusters explain why the count-based result looks cleaner than many readers might expect. Northern Virginia, the Pacific Northwest, and California all host substantial data-center activity, and each sits in grid regions that are below the national average in this analysis.
Loudoun County, Virginia, is the most famous example. The county has become so closely associated with cloud infrastructure that it is often called “Data Center Alley.”
In this dataset, Loudoun alone contains 144 identified facilities, the largest county concentration in the country. Its eGRID subregion, SERC Virginia/Carolina, has a carbon intensity of 269.2 g CO₂/kWh, well below the U.S. average.
Santa Clara County, California, another major cluster, contains 74 identified facilities and sits in the CAMX subregion at 194.3 g CO₂/kWh. The Pacific Northwest, supported historically by large hydroelectric resources, also hosts many identified facilities in a below-average grid region.
Seen this way, the map complicates the assumption that a larger data-center sector automatically means more infrastructure on the dirtiest grids. Many of the country’s most visible data-center clusters are not located in the highest-carbon parts of the U.S. electricity system.
That finding deserves to be taken seriously. It shows that the geography of digital infrastructure already reflects a mix of forces: access to reliable power, large grid infrastructure, hydropower in some regions, corporate siting preferences, and proximity to network routes and customers.
At the same time, a facility count is a blunt measure of an electricity-intensive industry. It tells us where facilities appear, not how much electricity flows through them.
Why accounting for size changes the picture
Data centers vary enormously. Some are small colocation facilities serving many customers, while others are hyperscale campuses operated by or for the largest technology companies in the world.
A dataset that counts each facility once can make the sector look more evenly distributed than its electricity demand really is. This is why the study also looked at physical footprint where those data were available.
Footprint information is present for 92.3% of facilities in the dataset. It is an imperfect proxy for demand, and the analysis treats it that way.
A larger footprint may indicate a larger site, although electricity use also depends on server density, cooling design, power utilization, redundancy requirements, and whether the mapped polygon represents a building footprint or a broader campus area.
Even with those caveats, the size-aware view is revealing. When each facility is counted equally, the mean carbon exposure is 317.8 g CO₂/kWh.
When facilities are weighted by physical footprint, the mean rises to 360 g CO₂/kWh. That is slightly above the U.S. average of 348.0 g CO₂/kWh.
Counting facilities makes the cloud look cleaner than average. Giving more weight to larger facilities pushes the picture above the national grid average.
This is the central result of the analysis. The identified data-center fleet looks relatively clean when each facility has the same weight, and it looks less clean when larger facilities carry more influence.
The reason is intuitive. If bigger sites are more common in somewhat higher-carbon regions, a simple count will understate the carbon exposure associated with the sector’s physical scale.
Footprint should not be read as electricity consumption. A facility twice the size of another does not necessarily use twice as much power, and a large, efficient facility may have a lower emissions profile than a smaller, less efficient one.
A size-weighted exposure estimate is therefore not a direct emissions estimate. Its value lies in showing how sensitive the story is to scale.
A map of dots can be visually persuasive, especially when the dots are numerous. Electricity demand, however, is not distributed one dot at a time.
If larger facilities sit in different places from smaller ones, a count-based view will miss part of the climate geography. For readers, this is a useful way to think about the next wave of AI and cloud infrastructure.
The question is not only how many data centers are built. The size of those facilities, and the carbon intensity of the grids they depend on, will shape the sector’s real-world climate consequences.
Where data-center clusters meet higher-carbon grids
The study also identifies regions where data-center concentration and higher grid carbon intensity overlap. This is where the map becomes especially useful for public discussion.
The largest subregion by facility count is SERC Virginia/Carolina, with 258 identified facilities and a carbon intensity of 269.2 g CO₂/kWh. This region includes the enormous Northern Virginia cluster, and its grid carbon intensity is 78.8 g CO₂/kWh below the U.S. average.
The WECC Northwest region also has a large count, with 179 identified facilities and a carbon intensity of 286.6 g CO₂/kWh. ERCOT, the Texas grid region, contains 113 identified facilities at 332.9 g CO₂/kWh, modestly below the national average in this analysis.
The higher-carbon hotspots appear elsewhere. RFC West, covering parts of Ohio, Illinois, and Wisconsin, has 134 identified facilities and a carbon intensity of 413.4 g CO₂/kWh.
MRO West, covering parts of Iowa, Nebraska, and Minnesota, has 73 identified facilities and a carbon intensity of 417.4 g CO₂/kWh. Both regions sit roughly 65–70 g CO₂/kWh above the national average.
These regions are not necessarily the places most associated with data centers in the public imagination. Northern Virginia and Silicon Valley are more familiar symbols of the cloud’s physical footprint.
The carbon-exposure map shows why visibility and climate relevance are not always the same thing. A famous cluster on a relatively lower-carbon grid may be less carbon-exposed per kilowatt-hour than a less famous cluster in a more carbon-intensive region.
For local governments and utilities, this kind of analysis has practical implications. Data centers are large, concentrated electricity loads, and their arrival can affect transmission planning, generation needs, reliability discussions, and the pace at which local grids need to add clean power.
In a higher-carbon region, new large loads create a sharper question. Will the grid serving them decarbonize quickly enough to keep digital growth from adding to fossil generation?
Why state averages can mislead
One of the quieter findings in the report may be among the most important for future analysis. A state-level map is often too crude for understanding the carbon exposure of data centers.
Electricity systems do not align neatly with state borders. A single state can contain more than one grid region, and a grid region can cross several states.
A facility’s location-based carbon exposure depends less on the political boundary around it than on the power system serving it. The report tested this directly by comparing the subregion-based results with a simpler state-average approach.
For 47% of facilities, the assigned carbon intensity changed by more than 10% when state-average rates were used instead of eGRID subregion rates.
For a lay reader, this finding is easy to grasp once the map is in view. Electricity geography is more complicated than a state-by-state ranking.
A data center announced in “Virginia,” “Ohio,” or “Texas” may sit in a particular grid subregion whose carbon intensity differs from the state average. If we want to understand location-based climate exposure, finer geography matters.
This is especially relevant as data centers become part of local economic-development strategies. Communities often evaluate proposed projects in terms of jobs, tax revenue, land use, water demand, and electricity availability.
Carbon intensity belongs in that local discussion. It should be measured using the best available grid geography rather than the easiest political boundary.
What this analysis can and cannot tell us
The study is most useful when read with its limits in mind. It measures exposure to local grid carbon intensity, not verified facility-level emissions.
Actual emissions would require facility-specific electricity use, and that information is generally unavailable in public datasets. It also uses annual average grid carbon intensity from eGRID.
Annual averages are appropriate for broad comparisons, although electricity grids change hour by hour. The carbon impact of an additional megawatt-hour can depend on which generator is operating at that moment.
Corporate renewable procurement adds another layer. Many large technology companies buy renewable energy through power purchase agreements, renewable energy certificates, or other market-based instruments.
Those purchases can matter, especially if they help finance new clean generation. They also sit alongside the physical reality that data centers draw power from local grids in real time.
A location-based analysis is therefore not a replacement for corporate emissions accounting. It is a way to keep the physical grid in the conversation.
The data-center location dataset is also incomplete. The IM3 atlas is derived from OpenStreetMap, which makes it transparent and useful, although not a complete census.
Coverage varies by region, operator names are missing for some facilities, and footprint information is imperfect. The study’s cleaning steps reduce double-counting, but they cannot turn an open map into a comprehensive commercial inventory.
These limitations define the claim we can responsibly make. The analysis shows where identified data centers sit on the U.S. grid carbon map, how those locations compare with the national average, and how the interpretation changes when physical size is considered.
That is still a valuable contribution. Public debates often leap directly from “data centers use a lot of electricity” to broad conclusions about climate impact.
This analysis gives us a more grounded intermediate step. Before we estimate emissions, we should understand the carbon intensity of the grids where the infrastructure is located.
What it means for the next phase of the cloud
The growth of artificial intelligence has made data-center energy demand a mainstream concern, although the issue is broader than AI alone. Streaming, cloud storage, business software, scientific computing, financial systems, online commerce, and everyday internet services all depend on the same expanding infrastructure.
As more of the economy runs through data centers, their relationship to the power grid will become harder to ignore. The data center carbon footprint will be shaped by more than server efficiency or corporate clean-energy claims.
The findings here suggest that the climate story will be written at several scales. Nationally, data centers are a growing category of electricity demand.
Regionally, they are unevenly distributed across grids with very different carbon intensities. Locally, a single large campus can become significant enough to influence utility planning.
For companies, the analysis points toward a more physical understanding of climate responsibility. Efficiency still matters, renewable procurement still matters, and the location and size of facilities matter as well.
A company choosing where to build new capacity is also choosing a grid context. That choice carries carbon consequences that may not be visible in a national average.
For policymakers and utilities, the findings suggest that data-center growth should be paired with serious grid planning. In lower-carbon regions, new demand still has to be managed carefully so that it does not strain infrastructure or crowd out other electrification needs.
In higher-carbon regions, the case for faster grid decarbonization becomes more urgent as large new loads arrive. The question is not whether digital infrastructure is good or bad in the abstract, but whether the grids serving it are getting cleaner quickly enough.
For readers, the main lesson is simple enough to carry away. The cloud is not uniform; it has a map, and that map matters.
By facility count, identified U.S. data centers are more often located on cleaner-than-average grids. That finding should temper the assumption that the sector is automatically concentrated in the dirtiest parts of the electricity system.
Once facility size enters the picture, the average carbon exposure rises above the national grid average. That second finding should temper any easy reassurance drawn from the count-based result.
The two findings belong together. They show that data-center carbon exposure depends on both geography and scale.
If we count only facilities, we miss the influence of the largest sites. If we look only at national electricity demand, we miss the regional differences in grid carbon intensity.
As the cloud grows, its climate impact will depend not just on better chips, cleaner corporate contracts, or more efficient cooling. It will also depend on where the next generation of digital infrastructure is built, how large those facilities become, and how quickly the surrounding grids can decarbonize.
The cloud may feel invisible from the user’s side of the screen, but on the carbon map it has become increasingly visible. The task now is to make that visibility useful.
Source note
This article is based on a location-based carbon-exposure analysis combining identified U.S. data-center locations from the IM3 Open Source Data Center Atlas with EPA eGRID2023 subregion carbon-intensity data. The analysis measures exposure to local grid carbon intensity, not verified facility-level emissions.
Figure inventory for WordPress editing
| Figure | Placement | Purpose |
|---|---|---|
| Figure 1 | After the introduction | Map showing data centers on the eGRID carbon map. |
| Figure 2 | After the count-based distribution discussion | Shows most identified facilities are in the 200–400 g CO₂/kWh band. |
| Figure 3 | After the size-weighted result | Shows equal-weighted versus footprint-weighted exposure. |
| Figure 4 | In the hotspot section | Shows data-center concentration versus grid carbon intensity. |
| Figure 5 | In the state-versus-subregion section | Shows why state averages can mislead. |
