Mapping Data Centers: A Fragmented Undertaking

1. The Cloud's Roots

Digital infrastructure is often described through the metaphor of the "cloud," which suggests an immaterial reality with no fixed location. This image conceals a substantial physical footprint: online services, data storage, and, increasingly, the training and operation of artificial intelligence models all rely on data centers: industrial buildings filled with servers, connected to electrical grids and telecommunications networks. Their consumption of electricity, water, and land is considerable, and its effects are felt well beyond the sites themselves.

The recent rise of generative artificial intelligence has changed the scale of these facilities. In Memphis, Tennessee, the company xAI brought online Colossus, a site that by early 2026 housed more than half a million graphics processors. Its power supply, provided in part by gas turbines installed before all required permits had been obtained, has drawn local opposition. In Texas, the Stargate campus in Abilene, backed by OpenAI, Oracle, and Crusoe, targets a capacity of 1.2 gigawatts, comparable to that of a nuclear reactor. In Virginia, Loudoun County, around Ashburn, has the highest concentration of data centers in the world.

These infrastructures transform the territories that host them: land use, grid connections, water consumption, tax revenue, noise pollution. They attract growing interest from public authorities, researchers, and nearby residents. Yet one observation stands out to anyone looking into the question: no state or international organization maintains an exhaustive public registry of these facilities, their location, their capacity, or their resource consumption.

Faced with this gap, a diverse range of actors has set out to map data centers: specialized companies, public laboratories, regulators, research teams, associations, citizen collectives, and communities of volunteer contributors. This article offers an overview of these mapping efforts. The interactive map accompanying the text locates the initiatives surveyed as well as a few representative data centers; the colored links in the text point directly to it, each color corresponding to a type of initiative.

2. The Obstacles to Mapping

The first obstacle lies in the very definition of what constitutes a data center. The category covers heterogeneous realities, from a technical closet of a few dozen square meters housing a telecom operator's equipment to campuses of several hundred hectares dedicated to training AI models. The guide for supporting data center implementation, published in November 2025 by France's Ministry of the Economy, distinguishes five families. Enterprise centers, small and often urban, are operated directly by their owners (banks, telecom operators, government agencies). Colocation and hosting centers let an operator lease shared infrastructure (power, cooling, security) to clients without owning their servers. Cloud service provider centers are run by the American hyperscalers (Amazon, Microsoft, Google) as well as by a handful of French and European players. "Edge" centers, decentralized and small, are located close to users to reduce latency. High-density centers, finally, are dedicated to high-performance computing and AI model training.

Across these categories, power capacity varies by a factor of more than a thousand. Depending on the threshold used (installed capacity, floor area, dedicated grid connection), the same region can be counted as having ten or a hundred data centers, which accounts for much of the discrepancy between different surveys. Public debate benefits from distinguishing these categories: the issues raised by a 500-kilowatt neighborhood site have little in common with those of a gigawatt hyperscale campus. This heterogeneity is not offset by any dedicated regulatory framework: the data center is not a legal category in its own right. Depending on the country and the project, it may be classified as an industrial facility, a telecommunications infrastructure, or a simple commercial building. As a result, this body of work has no stable, shared scope.

The second obstacle is the sector's opacity. Operators often invoke, jointly, trade secrecy and security requirements to limit the publication of information about their facilities. Development projects are frequently carried out through purpose-created project companies operating under code names that conceal the end client's identity. Permit applications and impact studies, when public, are sometimes redacted of power or water consumption figures. Non-disclosure agreements between developers and municipalities further restrict what even local elected officials can disclose. Under these conditions, mapping amounts less to compiling available data than to conducting an investigation.

The final obstacle is the sector's pace of change, which quickly renders any snapshot obsolete. Work from York University (Data Centred) illustrates this for Canada: the capacity of announced or under-construction projects there represents nearly fourteen times that of the operating fleet. A map limited to operational facilities would therefore capture only a fraction of the phenomenon, even though it is precisely at the project stage that the public can weigh in on land-use decisions. Tracking the full life cycle of projects, from announcement to commissioning, including cancellations, becomes a methodological requirement in its own right.

3. Families of Actors

In the face of these obstacles, the landscape of data center mapping has taken shape gradually, driven by the needs of actors with different purposes.

Private actors led the way, for commercial purposes. The oldest reference still active, Data Center Map, was launched in 2007 to make up for the lack of transparency in a market that was then hard to read, even for its own participants. Fed by operators' voluntary submissions, who see it as a channel for visibility, the site now lists several thousand facilities and has become the sector's main documentary reference. Baxtel and Aterio occupy a comparable, more specialized niche: the former offers detailed technical profiles by site, the latter paid datasets aimed at investors and equipment makers. These datasets cover in particular the United States and Canada, including projects under construction. These products provide information on services and capacity, much less on environmental or local impacts.

Public actors have entered the field more recently, following two distinct logics. The first is planning: in the United States, the National Renewable Energy Laboratory (NREL), a federal Department of Energy laboratory, published a map in late 2025 listing more than 4,000 data centers that are operating, under construction, or planned. The map relates them to electricity transmission lines and major fiber-optic backbones, in order to anticipate where power demand will materialize. The second is regulation: in France, the electronic communications regulator (Arcep) collects environmental data (electricity consumption, energy efficiency, water use) from operators every year as part of its "Pour un numérique soutenable" survey, which it publishes in aggregated form. This remains one of the few schemes worldwide in which a public authority legally compels operators to report environmental information, unlike methods that depend on their goodwill alone. On a smaller scale, the Institut Paris Region maintains an Observatory of Data Centers in Île-de-France that publicly and exhaustively documents the region's facilities, capacity, and projects, offering a rare example of complete institutional mapping confined to a specific territory.

Researchers have long taken an interest in the environmental impact of digital technology, and the rise of artificial intelligence has now given new momentum to the question of data centers specifically. The Epoch AI institute, which specializes in the quantitative study of AI trends, launched a data hub in November 2025 on the world's largest AI data centers. Its method combines high-resolution satellite imagery, building permits, and public records, and the entire dataset is published in downloadable form. In Canada, a team at York University produced in 2026 the first overview of the national fleet by tracking each project from announcement to commissioning, thereby documenting the marked gap between announcements and actual builds.

Advocacy and citizen actors form the most diverse family. In France, the collective Le nuage était sous nos pieds ("The cloud was beneath our feet"), formed in Marseille in 2023 amid contested expansion of the city's port-area campuses, tracks existing facilities, announced projects, and the mobilizations opposing them. In the United States, the Piedmont Environmental Council, a long-established environmental organization in Virginia, created in 2023 a map of existing and proposed data centers in Loudoun County. The case of the Ontario Data Centre Tracker deserves mention: this provincial tracker is the work of a single resident. Its methodology, informal and poorly documented, calls for caution regarding the published figures, but the very existence of the initiative testifies to a gap in public information that individuals are taking it upon themselves to fill. In Quebec, where no citizen mapping had existed until now, the NRS collective (Numérique Responsable et Soutenable) announced that a map from its research project À la recherche des centres de données québécois ("In search of Quebec's data centers") would be published sometime in 2026. It would be the first initiative of its kind in the province.

Open-source communities and the media round out the picture. In France, the company Hubblo launched in 2026 an open inventory of the country's data centers, built from public sources structured according to a standardized contribution schema. On OpenStreetMap, volunteer contributors document data centers one by one using the telecom=data_center tag. This open, imperfect but freely correctable data layer feeds other productions, such as the world map produced by the French outlet Next. That map cross-references OpenStreetMap's data centers with indicators of water stress, grid carbon intensity, and historical satellite imagery. This journalistic work shows what open-data aggregation makes possible, while also raising the question of attribution: the value of the final map rests largely on upstream community work that deserves to be explicitly credited.

This overview reveals a landscape that is rich but fragmented: each map answers a particular question, and scopes, definitions, and licenses differ enough to make comparisons difficult. The next section examines the collection methods that give rise to these differences.

4. Mapping Methods

Behind this diversity of actors, four main collection methods can be identified, which initiatives combine in varying proportions.

Declarative collection consists of obtaining information directly from operators. It can be voluntary, as with Data Center Map, where operators fill out their own listings, or mandatory, as in Arcep's annual survey. Reliable for what is declared, this method is by construction silent on what operators choose to withhold, and its coverage depends on the incentive, commercial or regulatory, to report.

Open-source investigation, or OSINT (open source intelligence), reconstructs information from public documents: building permits, grid-connection requests, planning records, corporate registries, local press, job postings. This method, central to the work of the Le nuage était sous nos pieds collective, demands considerable time and a detailed knowledge of local administrative procedures. It remains, however, the only method that can document projects before their official announcement, at the moment when the information holds the most value for public debate.

Remote sensing exploits a physical characteristic of data centers: their cooling equipment, sized in proportion to installed computing power, is visible in aerial and satellite imagery. Epoch AI's analysts have thus built models linking the diameter and number of a site's fans to its cooling capacity, and hence to its electrical load, with a precision they consider adequate for tracking large sites. Imagery also makes it possible to follow construction progress image by image, from earthworks to roofing, as at Memphis or Abilene. The method, however, is poorly suited to small urban facilities, which are often indistinguishable from ordinary buildings.

Crowdsourcing aggregates thousands of individual contributions. On OpenStreetMap, each contributor documents the facilities they know about; the sum of these contributions gradually builds a world map whose quality varies by region but which has two unique properties: it is freely reusable, and anyone can correct it. Errors there are public, discussed, and traceable, something no proprietary database can claim.

None of these methods suffices on its own. The most robust surveys combine them, as does the Observatory of Data Centers in Île-de-France, which maintains satellite tracking of previously declared centers. Next does the same by cross-referencing OpenStreetMap data with environmental indicators and historical imagery. These initiatives would benefit from greater interoperability, which is currently hindered by differing licenses and formats.

5. Mapping as Contestation

The siting of data centers has triggered local opposition of a new magnitude, particularly in the United States. The Data Center Opposition Report, which tracks organized online opposition groups, counted more than 500 local groups and 525,000 members as of June 2026, a sevenfold increase since December 2025. These groups are present in more than 40 states, forming at a rate of roughly two new groups per day. This opposition has itself become a mapping subject. The Coalition for Responsible Data Center Development tracks the citizen organizations formed around siting projects, producing a map of the mobilization that lets local groups identify one another and share experiences across state lines. In a different register, Data Center Watch, a project of the company 10a Labs, tracks opposition on behalf of a professional audience: its reports counted around 75 projects blocked or delayed by local mobilization in the first quarter of 2026, representing some $130 billion in investment. The fact that tracking opposition has become a business-intelligence product speaks to the scale the phenomenon has reached.

The relationship between contestation and mapping is not, however, reducible to this relation between object and observer. Neighboring residents' groups are often the first indicators that a project is arriving in a territory: attentive to rezoning requests, filed permits, unusual land transactions, and weak local signals, they pick up on clues that frequently precede official announcements. Their reports feed directly into citizen maps, such as those of the Piedmont Environmental Council or the Le nuage était sous nos pieds collective, which add projects as soon as they are detected.

The relationship also works in the other direction: mapping equips contestation, giving residents the means to situate a project within its local context and compare it to precedents, so as to raise precise questions during public consultations. The result is a mutual reinforcement: opponents produce part of the information the maps use, and the maps in turn strengthen their capacity to act. Local groups are therefore not an anomaly in the mapping landscape: they form a network of territorial observation, on which a good number of the initiatives described in this article depend.

This review points to a few lessons. A map is never neutral: one built by a broker, a regulator, or a residents' collective shows neither the same objects nor the same attributes, and transparency about definitions, sources, and licenses matters as much as the data itself. Sustainability is a second challenge: many citizen maps rest on a handful of volunteers, when they do not depend on a single person, and their long-term upkeep remains fragile in a fast-moving sector. Finally, linking up with existing open data, OpenStreetMap foremost among them, both pools the collection effort and ensures results remain reusable.

Despite these caveats, a clear trend emerges. Data center mapping, long the domain of sector specialists, is gradually becoming a tool for territorial information and democratic participation. Knowing where these facilities are, who operates them, what they consume, and what is planned is a precondition for any informed public deliberation about their siting. The initiatives surveyed here, for all their limitations and blind spots, lay the first stones: what remains is to connect them and make them last.