This year has demonstrated that data centers are becoming prime targets in modern warfare. The most prominent examples have emerged from the Russia-Ukraine conflict. Since late September, both sides have repeatedly struck each other’s data centers, severely disrupting civilian internet connectivity, including digital services critical to civilian populations. Nor is the phenomenon limited to that theater. Data centers have also been targeted in the ongoing war in the Middle East, as we examined in a previous Just Security piece.
This new frontier of warfare raises a question that has received far less attention than whether States may lawfully attack such facilities: what must States do to protect civilians and civilian objects under their control from the effects of enemy attacks on data centers? The answer lies in the law of armed conflict (LOAC) rules on “passive precautions.” Those rules dictate whether and under what conditions a State must ensure that its own military facilities are not co-located with civilian objects and civilian population centers.
The legal issues are pressing, as States are deciding now where to locate data centers for military or dual civilian-and-military use. France, for instance, is building data centers for exclusive military use. On Sept. 29, it unveiled “Arcadia,” a command-and-control and intelligence-fusion platform housed in a “vast underground data center” designed to provide “operational efficiency for the French armed forces.” The center is located in the historic Mont-Valérien fortress, on the outskirts of Paris.
The French initiative has obvious strategic logic. Dedicated military data centers reduce European militaries’ dependence on U.S. cloud providers such as Microsoft Azure or Amazon Web Services. Concerns over that reliance, and over so-called “data sovereignty” more generally, have spread across Europe. As another example, the Dutch Ministry of Defense is also investing in “a sovereign, military-controlled cloud” for classified information, hosted in a dedicated data center.
In this article, we briefly explain how data centers’ role in warfare implicates those rules on passive precautions. We then examine how those rules apply to kinetic attacks before turning to distinct questions raised by cyber operations.
Data Centers in Warfare
Data centers house networked servers and related equipment for storing and processing data, often with remote access. Because servers consume vast amounts of electricity, they generate a great deal of heat. As a result, a data center’s operation relies on power, cooling, and connectivity. Each presents a vulnerability during an armed conflict that an attacker can exploit to neutralize the center’s operations and the services that rely on it. This is significant because many actors do not build their own centers but instead rent servers or computing capacity from commercial data center (or “cloud”) providers. Thus, a single data center may simultaneously host servers supporting, for example, financial transactions, hospital patient records, and military capabilities.
The military significance of data centers is growing as armed forces increasingly integrate AI into intelligence analysis, targeting, command and control, and logistics. These applications require substantial computing power, particularly from hyperscale data centers, both to train AI models and to operate AI-enabled systems. Because building and maintaining these facilities requires considerable resources, many militaries rely on commercial providers to meet their computing needs.
This growing military reliance on data centers has made them increasingly prominent targets in warfare, despite civilian dependence on the same centers. Russia struck Ukraine’s governmental data center with a cruise missile early in the 2022 invasion, but by 2026 data centers are being attacked on a more widespread basis. In late September, Russian attacks damaged at least six data center facilities in Ukraine operated by multiple commercial entities, including ones located in Kyiv. Ukraine responded in kind in early October, attacking three data centers operated by Yandex (think Russia’s Google) and a data center that Sberbank, Russia’s largest bank, was building. And, while preparing this article, another Ukrainian drone hit a data center near St. Petersburg belonging to VK, the Russian technology group behind VKontakte, one of Russia’s largest social networks.
The civilian consequences have been significant. Russian attacks on Sept. 23 cut internet service to about 100,000 households and took Ukraine’s online court system offline for more than two hours. The damage also disrupted civilian and military television broadcasts. Ukrainian attacks on Yandex data centers in Russia also had a major impact, disrupting access to the internet, maps, ride-hailing, food delivery, and other online services, including some reporting that it affected healthcare. The attacks have also resulted in immediate deaths and injuries. For example, one of the Russian attacks on data centers in Kyiv caused four deaths, while the Ukrainian attack on the data center in Kaluga injured nine people.
The conflict in the Middle East provides other examples of the consequences for civilians when data centers are attacked. Iranian attacks targeting Amazon Web Services data centers in Bahrain and the United Arab Emirates (UAE) caused “devastating fire, power outages, and further damage as firefighters fought the blazes,” while users in the UAE lost access to banking and food-delivery applications that relied on Amazon’s cloud.
The mounting frequency of such attacks should not be taken to mean that data centers are necessarily lawful targets. Whether a particular facility may be targeted depends on whether it qualifies as a “military objective” pursuant to Article 52(2) of Additional Protocol I (for Parties) and its customary law equivalent. Article 52 defines military objectives in universally accepted terms as “objects which by their nature, location, purpose or use make an effective contribution to military action and whose total or partial destruction, capture or neutralization, in the circumstances ruling at the time, offers a definite military advantage.” Importantly, if a facility serves both military and civilian purposes, it qualifies as a military objective subject to attack, provided its use makes an effective contribution to military action and its destruction or neutralization offers a definite military advantage in the circumstances (DoD Law of War Manual, § 5.6.1.2). We caution readers that even then, the attack remains subject to the proportionality rule, precautions in attack, and any applicable special protections.
Precautions Against the Effects of Attacks
As data centers become indispensable to AI-enabled warfare, nearby civilians and those relying on the same infrastructure face increasing risks, from immediate physical injury to loss of access to data and disruption of essential services. Civilian infrastructure may also become collateral damage during an attack on a data center that qualifies as a military objective. Since cyber operations pose unique risks to data centers and the functions that rely upon them, we take on this issue in the next section.
The Rule. In light of such risks, attackers must take so-called “active” precautions in attack under Article 57 of Additional Protocol I and customary law to avoid harming civilians and civilian objects (ICRC Customary IHL study, rules 15-21; DoD Law of War Manual, § 5.11). They include choosing among feasible targets, weapons, and tactics to avoid such harm while still achieving the military advantage sought, as well as abiding by the rule of proportionality.
But the obligation to provide such protection does not rest on attackers alone; defenders must also act to protect civilians and civilian objects. Article 58 of Additional Protocol I provides,
The Parties to the conflict shall, to the maximum extent feasible:
(a) without prejudice to Article 49 of the Fourth Convention, endeavour to remove the civilian population, individual civilians and civilian objects under their control from the vicinity of military objectives;
(b) avoid locating military objectives within or near densely populated areas;
(c) take the other necessary precautions to protect the civilian population, individual civilians and civilian objects under their control against the dangers resulting from military operations.
This obligation reflects customary law in international armed conflict (ICRC Customary IHL study, rules 22-24; DoD Law of War Manual, § 5.14; Kupreškić, ¶ 524). Its customary status in non-international armed conflict is less well-settled. The ICRC considers the general obligation to be applicable in such conflicts (commentary to rule 22), while characterizing the more specific obligations to avoid locating military objectives near densely populated areas and to remove civilians and civilian objects from their vicinity as arguably applicable (rules 23–24). The experts who adopted Tallinn Manual 2.0 split over whether the rules enjoyed that status in non-international armed conflict (rule 121, ¶ 2).
Four features of the obligation shape its application to data centers. First, it is limited to what is “feasible,” a term the drafters used “to illustrate the fact that no one can be required to do the impossible” (AP I Commentary, ¶ 2245). Feasibility is assessed contextually. A precaution’s expected protective effect matters, but so too does its practical availability, the military consequences of adopting it, and the risks it may itself create. Accordingly, the DoD Law of War Manual defines feasibility as “practicable or practically possible, taking into account all circumstances ruling at the time, including humanitarian and military considerations” (§ 5.2.3.2). The fact that a measure cannot eliminate danger does not, by itself, render it infeasible.
As technology gives defenders new means of protecting civilians, “the understanding of the ‘feasibility’ test needs to evolve accordingly” (Jensen, p. 174). For instance, in the context of data centers, cybersecurity techniques and backup protocols to protect data from loss are evolving, providing new ways to protect services from disruption.
The second feature of the passive precautions obligation is that a State need only take precautions against reasonably foreseeable enemy operations and effects. After this year’s attacks, strikes on data centers supporting military operations as a category of attack are clearly foreseeable in future conflicts involving technologically advanced States; this reality alone triggers the obligation to take measures to protect civilians and civilian objects against them. Knowing the enemy intends to strike a specific data center makes the risk more concrete and may require additional feasible passive precautions. These could include, for instance, warning, sheltering, or evacuating civilians who work or live nearby.
Third, the obligation applies in both peacetime and armed conflict. As the ICRC commentary to Article 58 explains, the obligation “must imperatively be translated … first of all into measures to be taken already in peacetime” (¶ 2244). Although data and access to computing power can often be relocated relatively quickly, the infrastructure needed to support them cannot. Data centers take years to plan and build, and relocation or transfer options depend on the availability and geographic distribution of existing facilities. Consequently, peacetime decisions about constructing military data centers, and whether armed forces use commercial centers, can have lasting implications for civilian protection during armed conflict.
The fourth feature of the obligation is that it extends to “dangers” (Note: This will also become important in the more detailed cyber discussion below.) The meaning of that term determines whether the obligation is triggered by the mere loss of services. A majority of the Tallinn Manual 2.0 experts limited “dangers” to death, injury, and damage, while a minority would include “a major disruption of day-to-day life”; all agreed that mere “inconvenience or irritation” does not qualify as the danger the obligation is meant to preclude (rule 121, ¶ 9). As a result of this interpretive difference, potential physical destruction caused by a kinetic strike on a data center would qualify as a danger triggering the obligation under either view, but large-scale loss of internet access by a civilian population as a result of a cyber operation would amount to damage only under the minority view, unless that loss were likely to result in some form of physical harm to objects or individuals (e.g., by interrupting medical care).
A wide variety of measures can be taken to satisfy the passive precautions obligation as it applies to data center attacks. We discuss several of the most prominent below.
Segregating Military and Civilian Infrastructure: Segregating military from civilian data and related computing is an especially effective passive precaution, one contemplated in Article 58 (a) and (b). Dedicated military data centers, such as those of France and the Netherlands, can reduce the risk of data center attacks affecting civilians and civilian services. Whether such separation is legally required turns on feasibility in the circumstances. Where a dedicated military facility is practicable, its protective benefits may make reliance on a shared civilian facility difficult to justify; where it is not, States must consider other feasible measures.
In that regard, States could consider acquiring mobile tactical data center units for use during an armed conflict to provide edge computing capabilities and reduce their force’s reliance on commercial data centers. Demand for such units has risen in Russia amid the Ukrainian strikes, with one Russian manufacturer now marketing “a containerized data center.” However, because the enemy will likely target them as military objectives, mobile centers must be kept as far from populated areas as feasible (see below).
When military use of commercial or other “dual-use” centers (e.g., those serving both military and non-military governmental functions) is unavoidable, physically separating military servers, for instance in dedicated halls with their own power and cooling, may limit the civilian effects of a precision strike targeting the portion of the data center involved in military activities. Views differ on whether such segregation would mean the areas not used for military purposes qualify as civilian objects (see our discussion). But regardless of whether they do, the fact remains that, if feasible, such measures may still be required as a passive precaution to mitigate the negative effects of attacks or other operations (e.g., cyber operations, see below) on civilian services.
States could also achieve segregation by imposing regulatory or legislative restrictions on shared use by the military, on the one hand, and by civilian entities such as hospitals, emergency services, and financial institutions, on the other. States might also reduce civilian exposure to domestic kinetic attacks by relocating or replicating civilian data in facilities abroad. Shortly before Russia’s 2022 invasion, Ukraine passed legislation allowing the government to move data to the cloud. Amazon Web Services soon moved petabytes of data from government bodies, universities, and private firms such as PrivatBank into its data centers, and much of Ukraine’s digital infrastructure became “hosted in data centers across Europe.” The practice continues. After Russian strikes forced Kyiv’s Parkovyi data center to cease operations this fall, its operator reported that one of “the most important thing[s] is that … clients’ data had been deployed at backup sites in the EU.”
Although a different matter altogether, hosting military data and conducting military activities from abroad also implicates the law of neutrality. Belligerents may not use neutral territory for military purposes, and a neutral State must not allow such use (Hague Convention V, arts. 1-5; DoD Law of War Manual, § 15.2). Thus, for example, a neutral that allows a data center in its territory to be used to store belligerent intelligence data or operate its AI-enabled systems would almost certainly breach its duties under the law of neutrality.
Locating Military Data Centers: In addition to segregation, States must also consider the location of military objectives; Article 58(b) (and its customary law counterpart) requires States to “avoid locating [them] within or near densely populated areas.” The United States is not Party to Additional Protocol I. Still, the DoD Law of War Manual acknowledges that “[i]t may be appropriate to avoid placing military objectives … in urban or other densely populated areas,” while cautioning that “it often may not be feasible” (§ 5.14.1). The Manual also states more emphatically, “military commanders and other officials responsible for the safety of the civilian population must take reasonable steps to separate the civilian population from military objectives and to protect the civilian population from the effects of combat” (§ 5.14.1). To this end, and if feasible, States must locate military-use data centers away from concentrations of civilians.
In this regard, commercial considerations can push commercial, military, or dual-use data centers toward or away from concentrations of civilians. On the one hand, large facilities need affordable power and often cooling water, which draws some to sparsely populated areas. For example, Norway and Finland have selected data center sites based on the availability of hydropower. But commercial pressures can pull facilities into cities because of proximity to fiber networks, internet exchange points, customers, skilled workers, and the low latency that time-sensitive military applications sometimes demand. Kyiv’s Parkovyi data center, located in the heart of the capital, illustrates this pull. Where such factors make location away from civilian centers infeasible, other precautionary measures take on greater importance.
Concerns about placing data centers that perform military functions in or near populated areas are becoming concrete. For instance, the University of Michigan is building a major data center in Ypsilanti Township, Michigan, to support Los Alamos National Laboratory’s nuclear weapons research. Citing the Iranian attacks on Amazon data centers, the township’s attorney has warned that the data center will put “a big bull’s-eye target on this entire township.” The French decision to locate its military data center at a military base on the outskirts of Paris raises similar concerns. Paris is, after all, France’s largest and most populous city. Locating the data center on the base makes both the base and the center more attractive targets, thereby exposing nearby civilians to greater risk in the event of an attack.
The decision to place the French data center underground illustrates how a facility’s physical characteristics can affect both the likelihood and consequences of enemy attacks, and thus the passive precautions required. An attacker armed with precision penetrating munitions might strike it directly, keeping most of the weapon’s blast effects inside the facility and reducing risk to civilians. But to neutralize the data center, an attacker without such weapons might instead have to target the facility’s above-ground components, such as power sources, the cooling plant, and fiber connections, which may lie in the surrounding populated area. The foreseeability of these alternative targets bears on France’s obligations to take passive precautions. France would presumably claim the above ground option is not militarily feasible because it does not provide sufficient protection from attack. Regardless, the passive precautions obligations should inform both the decision to locate the facility at the base and the protective measures adopted there.
Other Measures. Article 58(c) requires “other necessary precautions,” such as shelters, “well-trained civil defence services,” and “effective fire-fighting” (ICRC Commentary, ¶¶ 2257-58). The data center analogs would include warning and evacuation plans for those who work or live near facilities with military functions and firefighting capacity scaled to the risk posed by attacks on those facilities.
Moreover, attacks on data centers in Russia and Ukraine illustrate how civilian services’ overreliance on a centralized network or a small number of data centers increases their vulnerability to attack. As Ars Technica has observed, Ukraine’s subsequent “decentralized digital ecosystem consisting of smaller data centers” and its “steps to build out backup infrastructure, including moving some equipment and networks underground or outside the country” contribute to the resilience of civilian services, such as internet access, financial services, and emergency response. Replicating critical civilian data and making advance arrangements for facility repairs can also serve as effective passive precautions in the event of data center attacks (Tallinn Manual 2.0, rule 121, ¶ 3).
Finally, point defenses can be an important passive precaution, as they may prevent attacks on data centers from endangering nearby civilians and civilian objects. Indeed, Russia instructed data center operators to prepare defenses against Ukrainian drone attacks. If circumstances allow, the obligation to take passive precautions requires such measures. Their feasibility depends on factors such as the likelihood of an attack, the defenses’ expected effectiveness, competing demands for scarce assets, and the risks posed by interception weapons or falling debris. A State might reasonably prioritize defending deployed forces over a data center. Conversely, protecting the facility may be required when effective defenses are available and foreseeable civilian risks are substantial. The decisive consideration is feasibility in the circumstances, not whether a defensive measure offers some benefit in isolation.
Passive Precautions and Cyber Operations on Data Centers
For the United States and other States beyond the reach of their adversaries’ drones and missiles, kinetic attacks on data centers are a remote, albeit not unimaginable, prospect. Cyber operations are more likely. Such operations can delete or encrypt data, disable services, or manipulate the industrial control systems that manage a data center’s power and cooling. Neither distance nor kinetic defenses offer protection against them. Whether Article 58 of Additional Protocol I and its customary law counterpart require the State to protect civilians from the effects of cyber operations on data centers depends on how two unsettled questions are resolved, along with the meaning of “dangers,” an issue discussed above.
The first question is whether the obligation applies only to “attacks” or to all “military operations.” The title of Additional Protocol I’s Article 58 includes “attacks,” while subparagraph (c) refers to “the dangers resulting from military operations.” A majority of the Tallinn Manual 2.0 experts limited the customary obligation to attacks, relying on the title of Article 58 and the absence of supporting State practice with respect to operations. A minority countered that the drafters’ choice of “military operations” in a treaty that uses “attack” throughout must have been deliberate (rule 121, ¶ 5).
The disagreement matters. By the majority view, the question is whether a cyber operation against a data center qualifies as an attack. Tallinn Manual 2.0 defines an attack as an operation “reasonably expected to cause injury or death to persons or damage or destruction to objects” (rule 92). In this regard, the ICRC has opined that limiting the notion of attack to operations causing injury, death, or damage “would be difficult to reconcile with the object and purpose of the IHL rules on the conduct of hostilities” (page 8). Along the same lines, a majority of the Tallinn Manual 2.0 experts agreed that interference with functionality qualifies as damage when restoring it requires repairing or replacing physical components. Still, they divided over operations that require only reinstalling the operating system or the bespoke data upon which the system relies. They also acknowledged a view that even temporary interference might qualify as damage (rule 92, ¶¶ 10-11). States have adopted positions across this continuum.
The second question is whether data qualifies as an object. Some States maintain that “only tangible things can constitute objects,” a view shared by a majority of the Tallinn Manual 2.0 experts (rule 100, ¶¶ 6-7). Others disagree, with the ICRC warning that excluding “essential civilian data,” such as medical, social security data, tax, and bank records, from the protection afforded to civilian objects “would result in an important protection gap” (p. 8; see also State positions, CCD COE “Toolkit”). If data is not an object, its loss or alteration does not, on that basis alone, engage the defender’s obligation to safeguard civilian objects. Feasible measures to do so may nevertheless be required pursuant to the broader interpretation of “dangers,” which, as explained, includes major disruption of civilian life.
Consider the risk of a cyber operation against a cloud provider’s data center that hosts data essential to the functioning of both military-related systems and university research systems. If the enemy launches such an operation, it is likely to erase or damage some university research data and probably cause the university’s research-related computer systems to go down. Whether the passive precautions obligation requires the State to take feasible measures to protect the university’s data and services turns on two separate determinations.
First, it is necessary to determine whether the cyber operation against a data center triggers the obligation, either because it extends to all military operations or because the operation qualifies as an attack (which, in turn, depends on the view taken on loss of functionality when there is no physical damage at the data center). If so, it must then be determined whether the effects on the university’s data and systems fall within the obligation’s scope. This depends on whether the affected data qualifies as an object and whether a major disruption of the university’s research activities constitutes a “danger.” Thus, whether feasible passive precautions are required in anticipation of such a cyber operation depends on the legal positions adopted and the consequences reasonably expected from that operation.
Foreseeability also takes on a distinctive character in the cyber context. AI-enabled tools allow an attacker to discover vulnerabilities in the defending State’s military systems that the State is unaware of and to exploit them in ways that spill over into civilian systems. A defender cannot guard against a specific vulnerability or the effects of its exploitation if it does not know the vulnerability exists. That said, if the defender knows the enemy uses such tools, the general risk is foreseeable. This being so, the defender must take feasible measures to mitigate that risk to civilians, such as strengthening the security of shared systems or employing its own AI-enabled defenses.
Although the measures required by the passive precautions obligation largely reflect sound cybersecurity practices, LOAC makes their implementation legally obligatory when feasible. In the context of data centers, such measures may include segregating military from civilian networks and isolating systems supporting essential civilian infrastructure from the internet (ICRC, p. 6). They may also include backing up important civilian data and making advance arrangements for the timely repair of computer systems, measures that should be taken in peacetime (Tallinn Manual 2.0, rule 121, ¶ 3). Other measures to safeguard against enemy cyber operations might include hosting military data and activities in dedicated classified environments; using security protocols for military servers that are separate from those providing civilian services so that an operation targeting the former does not spill over into the latter; isolating a facility’s power and cooling control systems from the internet; and maintaining offline, geographically separated backups of essential civilian data due to their exposure from military use of data centers.
The same logic applies to the AI models hosted by data centers. Operations against models, such as data poisoning, often delete or alter the data that comprise the model. A military that relies on the same model as civilian agencies, such as a facial recognition model shared with law enforcement, exposes those agencies to operations aimed at the armed forces. Maintaining a separately deployed military model, with segregated data, access controls, and supporting infrastructure, can reduce that shared exposure.
Concluding Thoughts
As military operations become increasingly dependent on data centers, the risk to civilians who rely on them will inevitably rise. LOAC does not prohibit the military use of commercial data centers, nor does it require States to eliminate the risks that such use creates. It does, however, require States to take feasible precautions against the foreseeable consequences of enemy operations. States cannot defer those obligations until hostilities begin. Peacetime decisions about the location, segregation, and protection of data centers may determine the risks civilians face when armed conflict erupts. War may be moving to the cloud, but the duty to protect civilians is not lost in it.
– Klaudia Klonowska and Michael Schmitt, Published courtesy of Just Security.
