In March 2026, Iranian drones struck Amazon Web Services facilities in the United Arab Emirates and Bahrain. Physical infrastructure was damaged. Cloud services across the region went dark. For the first time in modern conflict, commercial hyperscale data centers became deliberate military targets.
Iranian state media called it a strike against "the enemy's technological infrastructure." Most coverage landed on the obvious: cybersecurity implications, defense posture, Middle East escalation. All of that is real.
None of it is the structural finding.
The Classification Gap
The structural finding is simpler and more uncomfortable. Compute infrastructure that entire economies now depend on—payments processing, logistics routing, healthcare records, government service delivery—is still classified, in most jurisdictions, as private commercial real estate. The drones did not create a new vulnerability. They exposed a gap that was already load-bearing: the distance between the role this infrastructure actually plays and the protection it actually receives.
The U.S. government designates 16 critical infrastructure sectors. Energy. Water. Transportation. Healthcare. Financial services. These are the systems whose failure, in the government's own language, would have a debilitating effect on national security, the economy, or public health. The designation triggers federal oversight, resilience requirements, and coordinated protection planning across government and private sector.
AI compute infrastructure serves all 16 of those sectors. It is not among them.
Technically, the United States does include an Information Technology sector among its 16 critical infrastructure categories. But the designation was written when IT infrastructure meant servers in corporate basements and telecom switching stations, not hyperscale AI compute campuses consuming as much electricity as mid-sized cities. The governance architecture exists in name. Its scope, enforcement mechanisms, and resilience requirements have not been updated for infrastructure operating at this scale or this level of economic dependency.
That is the structural condition this article examines.
What the Numbers Reveal
Scale tells part of the story. U.S. data centers consumed roughly 176 terawatt-hours of electricity in 2023, about 4.4% of total national demand. In 2018, the figure was 76 terawatt-hours, or 1.9%. Researchers at Lawrence Berkeley National Laboratory, preparing an analysis for the Department of Energy, project that number could reach 325 to 580 terawatt-hours by 2028. That would represent between 6.7% and 12% of total U.S. electricity consumption.
When a single category of infrastructure begins shaping grid planning, generation investment, and transmission timelines, its continuity stops being a private operational matter. It becomes a macroeconomic condition.
Capital intensity tells the rest. Developing one megawatt of data center critical load capacity in the United States costs between $9.3 million and $15 million, according to benchmarks from Cushman & Wakefield. That kind of capital density means that physical damage is not merely disruptive. It is slow to reverse. Recovery requires construction lead times, constrained supply chains, and the complex commissioning of power and cooling systems. These are the same frictions that make outages in traditional critical infrastructure sectors economically devastating.
And these are the peacetime numbers. Survey data from the Uptime Institute shows that a majority of data center operators report their most recent significant outage costing more than $100,000, with roughly one in five exceeding $1 million. Before anyone accounts for the additional complexity of damage from kinetic military strikes.
The Thinness Underneath
Cloud architecture is built on a premise that sounds like redundancy but operates as something less. Availability zones within a geographic region provide failover for component failures: a server goes down, traffic reroutes. The architecture was engineered for this. It was not engineered for correlated physical destruction across multiple facilities in the same region.
This is structural thinness—the condition where the margin between normal operation and failure is narrower than the system's own architecture acknowledges. The redundancy existed on paper. In practice, the "backup" sat within strike range of the same threat that could take out the primary. When the test arrived, the margin that everyone assumed was there turned out to be thinner than the blueprint suggested.
Multi-region and multi-cloud architectures exist. The tools to distribute compute across geographies are commercially available. The structural condition is that most organizations have not deployed them. The cost, complexity, and regulatory friction of true geographic redundancy mean that the default configuration for most production workloads remains single-region. The capability exists. The adoption does not.
The consequences compound because of what compute infrastructure has become. The cost of performing inference at GPT 3.5-level capability fell more than 280-fold between late 2022 and late 2024, according to Stanford's Institute for Human-Centered Artificial Intelligence. McKinsey estimates that by 2030, inference will surpass model training as the dominant data center workload, representing more than half of all AI compute. Inference is not a research activity confined to episodic training runs. It is embedded in everyday operations: payment processing, supply chain routing, clinical decision support, fraud detection, government service delivery.
When inference is woven into the operational fabric of daily economic life, a compute outage stops resembling a software glitch. It starts resembling a power outage in its cross-sector reach. And the amplification dynamics follow directly: the more sectors that depend on a single infrastructure layer, the wider the damage propagates when that layer fails. AI does not create new failure modes here. It accelerates the existing ones by deepening the dependency on infrastructure whose structural protection has not kept pace with its structural importance.
The Governance Gap
The United Kingdom designated data centers as Critical National Infrastructure in 2024, placing them alongside energy and water for resilience prioritization. The European Union's NIS2 framework subjects cloud computing and data center service providers to heightened cybersecurity and resilience obligations. These are real moves.
Most jurisdictions have done neither.
The absence is not an oversight in the bureaucratic sense. It is a structural condition: governance absence describes situations where the frameworks that should exist to manage a recognized dependency simply have not been built. The dependency grew faster than the governance. AI compute infrastructure was commercial real estate when the classification frameworks were written. It became a foundational utility while the frameworks stayed the same.
The Gulf strikes made this absence operationally concrete. When an affected UAE insurance platform needed to shift workloads outside the region after the attack, the migration required regulatory approval. Local data residency rules mandated that insurance-related data remain hosted domestically. The governance designed to protect data integrity actively prevented the rapid failover that operational resilience required.
This is where two structural conditions collide. Data sovereignty rules—which represent one form of governance—sit in direct tension with failover architectures, which represent another form of governance. Neither has structural priority over the other in current frameworks. During the Gulf strikes, both failed simultaneously. The sovereignty rules could not flex to accommodate the emergency, and the failover architecture stalled behind a regulatory approval process that was never designed for wartime speed.
No jurisdiction currently has a mechanism for "emergency data mobility" during defined infrastructure crises. That mechanism does not exist because the infrastructure it would govern is not yet classified as the kind of infrastructure that warrants emergency protocols. The classification gap creates the governance gap, which creates the operational gap, which the drones revealed.
The Concentration Underneath the Concentration
Gulf states are not building data centers as a defensive measure. They are building them as an economic strategy. Saudi Arabia, the UAE, and their neighbors are positioning themselves as global hubs for AI compute and inference at scale. AWS is investing more than $5.3 billion to build a new cloud region in Saudi Arabia. Google Cloud and the Saudi Public Investment Fund announced a $10 billion AI hub partnership. Microsoft committed $15.2 billion to UAE data center capacity through 2029.
The investment logic follows market incentives. But it creates a structural condition worth naming: geographic concentration of compute capacity in a region experiencing active military conflict. The same forces driving the investment (capital availability, energy infrastructure, strategic ambition) are producing a concentration pattern that mirrors the very vulnerability the March strikes exposed.
This is the pattern that critical infrastructure designation was designed to address. When essential systems concentrate in ways that create single points of failure, designation triggers the planning, protection, and resilience requirements that diffuse the risk. Without designation, the concentration deepens on market logic alone, and the structural margin between operating normally and regional failure keeps getting thinner.
The drones were the Gulf's version of the test. Other geographies carry different versions. A sustained grid failure in northern Virginia, where a significant share of global internet traffic passes through data centers concentrated in a single county, would produce a similar structural outcome through a different mechanism. Extreme weather events, increasing in frequency and severity, threaten physical infrastructure with the same correlated regional failure pattern the drone strikes demonstrated. The threat vector varies by geography. The structural condition does not: concentrated infrastructure with thinner margins than the architecture acknowledges.
What This Means for Organizations
The structural argument from the World Economic Forum is that governments should treat AI infrastructure as critical infrastructure. The structural observation from this analysis is that AI infrastructure already functions as critical infrastructure, in every way that matters operationally. The only thing missing is the governance that acknowledges it.
For organizations, the implication is not abstract. Every organization running production workloads on a single cloud provider in a single geographic region is carrying a version of the same structural condition the Gulf strikes revealed. The redundancy they are paying for may be thinner than the architecture suggests. The failover they assume will activate may face regulatory constraints they have not mapped. The governance gap is not just a government problem. It is a structural condition inside every organization that depends on compute infrastructure for daily operations.
The scale of the dependency is still growing. IMF modeling suggests AI-driven productivity gains could raise global GDP by several percentage points over the next decade, concentrated in economies with stronger digital infrastructure. That is the upside of the dependency. The Gulf strikes revealed the other side: when the infrastructure that enables the productivity gains also concentrates the failure risk, the upside and the downside live in the same place.
The question for any organization relying on this infrastructure is not whether their cloud provider has redundancy. It is whether that redundancy has been tested against the actual threat landscape, or only against the threat model the architecture was designed for.
The Monday Morning Audit
Ask your infrastructure team three questions this week:
Where is our compute actually running? Not the provider's name. The physical locations. How many availability zones, in how many regions, in how many countries? If the answer is "multiple zones in one region," you are carrying the same structural condition the Gulf strikes exposed. Zones in the same geography can fail together.
What would prevent us from moving workloads across borders in an emergency? Data residency rules, regulatory requirements, contractual obligations, technical dependencies. Map them. The UAE insurance platform discovered these constraints during the crisis, not before it. The time to identify regulatory barriers to emergency failover is before you need the failover.
When was the last time we tested failover against a scenario that was not a component failure? Cloud providers test for server failures, network partitions, software bugs. What hit the Gulf was none of those. They were correlated physical destruction across multiple facilities. If your disaster recovery plan assumes the backup region is still standing, you are testing against the comfortable scenario, not the structural one.
Source: Jabbour, O. (2026). “When data centres become targets: It’s time to treat AI infrastructure as critical infrastructure.” World Economic Forum, April 2, 2026. weforum.org
The structural dynamics in this post connect directly to The 16 Systems Your Life Depends On, which maps the critical infrastructure sectors this analysis argues AI compute should join. The AI Amplification page examines how AI deepens existing structural dependencies rather than creating new ones. The Four Frequencies framework is described at The Four Frequencies. The diagnostic that measures these conditions for organizations is at Organizations.
This analysis publishes monthly. The Frequency Report goes deeper: with a structural tracker across twelve sectors, reader observations from the field, and a full four-frequency diagnostic each month.