<\/p>\n
In retaliation for<\/span> the ongoing U.S.\u2013Israeli war, Iran responded with a novel form of counterattack. For the first time in military history, private sector data centers came under deliberate attack.<\/p>\n In an era when companies known for e-commerce, social networks, and search engines have also become close collaborators with militaries, is bombing their servers fair game?<\/p>\n Three days after the U.S. and Israel began their joint bombardment, the Islamic Revolutionary Guard Corps launched kamikaze drone strikes against Amazon-owned data centers in the United Arab Emirates and Bahrain that provide an array of cloud computing services to customers throughout the Middle East. The impacts and subsequent fires \u201ccaused structural damage, disrupted power delivery to our infrastructure, and in some cases required fire suppression activities that resulted in additional water damage,\u201d according to Amazon, resulting in service outages across the region.<\/p>\n The motive behind the attack, according to Iranian state television, was not to block people from ordering groceries or posting to social media, but rather to highlight \u201cthe role of these centers in supporting the enemy\u2019s military and intelligence activities.\u201d Though only Amazon\u2019s centers are known to have come under fire, a March 11 tweet<\/a> from the quasi-official Tasnim News Agency listed dozens of regional facilities, including data centers owned by Microsoft, Google and others, deemed \u201cEnemy Technology Infrastructure\u201d suitable for targeting.<\/p>\n It\u2019s unclear if the Amazon data centers struck by Iranian drone strikes are used for military purposes or civilian purposes, or both. And it\u2019s unknown if the attacks in any way hindered the militaries of the U.S., Israel, or their allies in the Gulf from using AI or other cloud-based services in their war efforts. But with Amazon, Google, and even Facebook parent company Meta are all eager partners of the Pentagon that augment the destructive power of the United States in Iran and elsewhere, server farms may now have the same status as factories building bombs and warplanes.<\/p>\n Scholars of international law and the laws of armed conflict say that when a military runs on the cloud, the cloud becomes a legal military target. But the cloud is an abstraction, not a physical site \u2014 a global network of millions of chips in servers spread across hundreds of massive buildings across the planet, servicing both civilian apps and state tools used to surveil and kill. Separating the former from the latter is an extremely difficult task.<\/p>\n \u201cThe legality turns on whether the specific facility, at the specific moment, is genuinely serving the military operations of a party to the conflict in a way that offers a concrete and definite advantage to the attacker,\u201d explained Le\u00f3n Castellanos-Jankiewicz, a lawyer with the Asser Institute for International and European Law in The Hague.<\/p>\n <\/p>\n Sometimes the split between military and civilian use is straightforward. Microsoft, for example, helps run the Joint Warfighter Cloud Capability, which the Pentagon says provides it with \u201cgreater lethality.\u201d This work involves the processing of classified data, which the government does not want commingling with civilian tech. Cloud computing services are generally offered via geographically distinct \u201cregions,\u201d each made up of many physical data centers. Customers typically select the region that is closest to them to minimize lag time. Microsoft\u2019s US DoD Central and US DoD East regions are \u201creserved for exclusive [Department of Defense] use,\u201d according to the company, and are serviced by data centers in Des Moines, Iowa, and Northern Virginia, respectively.<\/p>\n Amazon offers similar cloud regions exclusive for Pentagon use, though the location of these data centers is not public. Oracle, another JWCC provider, operates Pentagon-specific facilities in Chicago<\/a>, Phoenix<\/a>, and Virginia<\/a>. Companies are understandably tight-lipped about where exactly on the map these facilities stand, in no small part because Iran, or any country at war with the U.S., would have reason to target them.<\/p>\n \u201cA data center that is used solely or primarily for military applications is targetable,\u201d said Ioannis Kalpouzos, an international law scholar and visiting professor at Harvard Law, \u201cand a center that supports the Pentagon\u2019s JWCC falls in that category.\u201d<\/p>\n The march of data center construction<\/a> has become a point of contention<\/a> across the United States and around the world, with communities frequently \u2014 and sometimes successfully \u2014 rallying to block what they view as enormous resource-draining<\/a> eyesores. But for those living in the widening shadow of data centers, planned or built, their status as military targets may be unsettling beyond concerns over water and energy consumption.<\/p>\n And as Defense Secretary Pete Hegseth aggressively shoehorns AI tools into the military wherever possible, the rapid expansion of data centers means the potential proliferation of legitimate military targets across the United States.<\/p>\n With comparisons between the destructive power<\/a> of AI-augmented warfare and nuclear<\/a> weaponry becoming<\/a> more common<\/a>, the ever-expanding network of American data centers may recreate Cold War anxieties around intercontinental ballistic missile, or ICBM, silo placement. The country\u2019s nuclear launch capabilities were famously clustered in the relatively sparsely populated Upper Midwest, forming a so-called \u201cnuclear sponge<\/a>\u201d that would draw Soviet nukes away from population centers and toward rural areas and farmland.<\/p>\n But the legal calculus around most data centers will be less clear. Google, for example, says the Pentagon uses both its general purpose public cloud<\/a> and smaller specialized air-gapped networks<\/a> that don\u2019t touch the public internet, depending on the sensitivity of the data involved. Even cloud work involving Top Secret<\/a> military data \u201ccan operate within Google\u2019s trusted, secure, and managed data centers.\u201d The company also sells modular mini-data centers<\/a> for use closer to battlefields or bases.<\/p>\n These arrangements, shrouded in both military and trade secrecy, make it hard to assess whether a server is hosting a student\u2019s homework or Air Force R&D, blurring the legality of attacking data centers that may host both. Google may have little control over how governments use its cloud tools; The Intercept has previously reported<\/a> that Google executives worried internally they wouldn\u2019t be able to tell how the Israeli military was deploying its cloud services.<\/p>\n \u201cThe practical challenge is that cloud infrastructure is often technically opaque, even to providers themselves,\u201d Castellanos-Jankiewicz said. \u201cThe services a given data center supports may not be readily ascertainable from the outside or even inside, which complicates the attacker\u2019s legal obligations considerably.\u201d<\/p>\n Amazon and Google\u2019s Project Nimbus similarly provides cloud computing services across the Israeli government, including both civilian agencies and the Ministry of Defense, along with state-owned weapons companies.<\/p>\n \u201cThe picture becomes more legally complex when a data center functions as a so-called \u2018dual-use\u2019 object,\u201d simultaneously hosting military data or capabilities alongside civilian services,\u201d Castellanos-Jankiewicz told The Intercept. \u201cOnce a facility is found to make an effective contribution to military action, the entire physical object can, under the dominant legal view, qualify as a military objective.\u201d<\/p>\n The embrace of commercial cloud computing by the U.S. and others has muddled an already murky legal picture, Castellanos-Jankiewicz explained. \u201cA military\u2019s decision to store classified data or run AI-enabled military systems on commercial cloud infrastructure shared with civilian services could itself raise legal concerns \u2014 particularly if the commingling of military and civilian uses makes a strike more likely or increases the foreseeable harm to civilians when one occurs.\u201d<\/p>\n Determining whether a given data center can be legally attacked under international humanitarian law \u2014 itself comprised of various treaties that not every country adheres to \u2014 relies on a complex series of balancing tests that rarely produce concrete answers. To begin with, every object and person is generally presumed civilian and exempt from attack under this framework. Before launching a strike, a country is supposed to have a verifiable reason to believe a data center contributes to the enemy war effort, and reason to believe an attack will appreciably harm that effort. What \u201ceffectively contributes to military action\u201d will, of course, be a source of disagreement<\/a>.<\/p>\n Anthropic\u2019s Claude large language model was reportedly used to accelerate American airstrikes against Iran; Claude, in turn, was built in part using 500,000 chips<\/a> housed in an $11 billion<\/a> Amazon data center in Indiana. If Claude is now arguably a weapon, is this Indiana site the data equivalent of a bomb factory? Kalpouzos, the Harvard Law visiting professor, told The Intercept it depends on the facts at the moment the bomb hits, not past usage. \u201cIf the facility is currently used in the training of the LLM that is used in the conduct of military operations \u2014 for example, by fine-tuning object classification or user-interaction features \u2014 then this could render it targetable,\u201d he said.<\/p>\n