When Amazon Web Services faltered in northern Virginia in the small hours of 19–20 October 2025, the disruption felt less like a technical glitch and more like the sudden removal of a piece of critical infrastructure. Payment systems stalled, banking apps timed out, cargo tracking screens froze, and public portals that citizens now treat as basic utilities stopped responding. Monitoring firms logged millions of individual outage reports as failures rippled outward from a single region in the United States to users on several continents. Subsequent analysis traced the episode to a fault in the automated management of Domain Name System records for DynamoDB in the us-east-1 region. An empty record propagated through a system that underpins hundreds of thousands of endpoints, automation failed to self-correct, and engineers had to intervene manually to restore service.
The incident will eventually join a familiar list of cloud outages that specialists recall and most users forget. Yet it should be read less as an isolated failure and more as a visible marker of a deeper structural issue. A single software bug in one region briefly disrupted services across the global economy on a quiet weekend. That raises a more strategic question. If a transient error in DNS automation can produce that level of disturbance, what would a prolonged loss of underlying hardware do in a crisis, and who controls the materials, factories, and grid components that determine how quickly such damage can be repaired.
The Virginia cluster is not an abstract “cloud.” It is an industrial complex tied to a stressed regional grid through a small set of substations and large power transformers. Those substations, in turn, depend on equipment built in a limited number of factories, using metals and specialty materials refined in an equally narrow group of countries. The October outage was a software event. Its broader significance lies in what it reveals about how much of today’s digital power rests on a concentrated and vulnerable physical base.
A planner looking at current investment in artificial intelligence infrastructure would start with straightforward numbers. A single hyperscale facility built to house dense AI training and inference workloads now routinely requests interconnection capacities on the order of 80 to 150 megawatts. Individual campuses can grow beyond that range when multiple halls and backup systems are included. At this scale, data centers are among the largest incremental loads on regional grids, rivaling steel mills or industrial parks.
The physical bill of materials is extensive. High-capacity campuses require thousands of tons of copper for cabling, busbars, and switchgear; large volumes of grain-oriented electrical steel for transformer cores; and tens of thousands of tons of concrete and structural steel for shells, cooling towers, and ancillary buildings. Inside each rack, smaller quantities of gold, silver, tantalum, cobalt, and rare earth elements are embedded in processors, memory modules, capacitors, and power electronics. Optical transceivers and long-haul links depend on high-purity silica, indium phosphide, and other compound semiconductors.
Those inputs do not appear spontaneously. They arrive through supply chains that exhibit their own bottlenecks and geopolitical leverage points. On the materials side, China still refines the bulk of the world’s rare earth elements and dominates magnet production, and it occupies a central position in the processing of gallium, germanium, and other critical technology metals.
Beginning in August 2023, Beijing imposed license requirements on exports of gallium and germanium products. In December 2024 it went further, announcing a halt to exports of gallium, germanium, antimony, and related materials to the United States, framing the step as a response to expanding U.S. controls on advanced semiconductor equipment. In October 2025, China tightened restrictions again, extending controls to a range of rare-earth-related technologies and processes, and constraining the ability of Chinese nationals and firms to support rare-earth projects overseas without approval.
These measures have not resulted in a sustained cutoff. A trade truce announced in November 2025 included commitments by Beijing to ease some export curbs on critical minerals and resume selected semiconductor-related shipments. The pattern is still clear. Access to specific inputs near the bottom of the technology stack has become an instrument of statecraft in a wider contest over digital industrial capacity. Even temporary or partial restrictions can create uncertainty for long-lead investments in data center equipment and power electronics.
If materials provide one axis of vulnerability, power transformation provides another. Large power transformers occupy a peculiar position in modern grids. They are custom-engineered, heavy, difficult to transport, and essential for connecting high-voltage transmission networks to large loads such as data centers and industrial facilities. A transformer failure at the wrong node can take a major region offline for months if there is no spare.
The U.S. Department of Energy’s 2024 Large Power Transformer Resilience report and a parallel assessment by the National Infrastructure Advisory Council describe a tightening supply situation. Lead times for large power transformers have lengthened to around 36 months in many cases, with upper bounds reported at 60 months in extreme cases. Industry surveys and utility trade publications in 2024–2025 report that some manufacturers are quoting delivery times of 80 to over 200 weeks, depending on specifications, and in some cases have stopped accepting new orders because their books are full for several years.
The United States imports a large share of these transformers and much of the grain-oriented electrical steel that goes into them, from a small group of suppliers in Asia and Europe. Domestic capacity exists but is limited. Expanding it requires capital investment, skilled labor, and time, and it competes with demand from transmission projects, renewable interconnections, and aging asset replacement.
The transformer backlog is already delaying grid modernization and renewable integration projects. Each additional hyperscale data center that applies for an interconnection agreement enters a queue that also contains semiconductor fabs, electrified transport infrastructure, and basic distribution upgrades. In practice, a proposed cloud region is competing with a hospital expansion and a new residential development for the same constrained set of transformers, steel, copper, and engineering attention.
Placed alongside those facts, the October AWS failure looks less like a one-off glitch and more like an early warning. On that weekend, the trigger was a DNS automation bug. Next time, it could be the loss of a large transformer at a substation feeding a major cloud campus, with a replacement two hundred weeks away. It could be a minor component whose export is suddenly restricted during a diplomatic crisis, slowing the commissioning of a facility intended to serve defense and emergency workloads. From the vantage point of an end user, the effect is the same. Compute capacity disappears, or never comes online as expected, and the chain of material dependencies that produced the shortfall remains largely invisible.
Deterrence when the bottleneck is a factory, not a shipyard
For most of the twentieth century, strategic planners drew maps around oil fields, ore deposits, shipping lanes, and rail junctions. War plans revolved around tanker tonnage, refinery throughput, and the capacity of shipyards to repair hulls under fire. The infrastructure of deterrence was heavy, visible, and straightforward to count.
The infrastructure that underpins contemporary deterrence looks different, but it is no less physical. Intelligence, surveillance, and reconnaissance systems rely on constellations of sensors generating continuous streams of data. Precision targeting depends on fusing, filtering, and delivering those signals to operators in near real time. Logistics and maintenance are increasingly managed through predictive tools and shared dashboards. Financial markets and sanction mechanisms that form part of crisis signaling operate through digital platforms. All of this rests on a computing base that is distributed in software but concentrated in hardware.
Ukraine’s experience since February 2022 illustrates both the promise and the limits of this model. Faced with the risk of missile strikes on government data centers and sustained cyberattacks, Ukrainian officials moved core registries and critical workloads into Western cloud platforms and foreign data centers. By mid-2022, tens of petabytes of government, financial, and educational data had been migrated from local servers to public clouds hosted across Europe.
Those migrations allowed ministries to continue operating after physical facilities were damaged, kept tax and property records available, and preserved key datasets for eventual reconstruction. They also relied on the assumption that the underlying infrastructure in host countries would remain stable, powered, and reachable even as Ukraine’s own grid and networks came under attack. In practice, the continuity of a sovereign state’s digital functions depended on the resilience of data centers and transmission corridors in several NATO and EU members.
The United States has absorbed that lesson in its own planning. The Pentagon’s Joint Warfighting Cloud Capability contracts and its broader effort to implement joint all-domain command and control assume that commercial clouds will handle large portions of unclassified and some secret workloads. These platforms are expected to support distributed training, planning, and logistics, and to provide elastic capacity that can be surged in a crisis. Defense planners now treat major U.S. cloud providers as latent extensions of the national command network, even when they remain privately owned.
This choice links deterrence credibility to the practical question of whether U.S.-based companies can keep data centers functioning under stress. The conventional image of an attack on command and control still tends to focus on satellite jamming, cable cuts, or strikes on fixed headquarters. A more plausible scenario for the next decade involves a combination of mundane and deliberate pressures: cyber intrusion that degrades grid operators, extreme weather that pushes transmission assets beyond their tolerances, and targeted physical sabotage of substations that serve cloud campuses, all occurring against a backdrop of already scarce transformers and long replacement queues.
Imagine a crisis in the Taiwan Strait or the Persian Gulf in which, roughly at the same time, several substations feeding large American cloud regions suffer failures. Some events might be attributable to storms or equipment aging, others to sabotage, others perhaps to cascading operator error. If the lead time for procuring and installing replacement equipment is measured in years rather than months, and if spare inventories are thin, the resulting capacity loss could endure well beyond the news cycle. Even if no classified network is directly affected, the knock-on effects on training pipelines, logistics scheduling, financial clearing, and routine governmental operations could be significant.
Such scenarios are not speculative curiosities for rival planners. Chinese analysts can study open-source accounts of the October AWS outage and of Western transformer shortages, cross-reference them with their own experience building state-directed computing clusters, and explore how long-duration infrastructure disruptions might affect an adversary’s ability to manage two geographically distinct crises simultaneously. The bottleneck in such cases is not a shipyard repairing warships. It is a factory slot producing high-voltage transformers, a mill rolling electrical steel, or a smelter refining a seemingly obscure metal that sits deep inside power electronics.
China’s national compute strategy and material leverage
China’s Eastern Data, Western Computing initiative offers a useful contrast with the more fragmented U.S. landscape. Launched in 2022, the program links eight national computing hubs to high-capacity grid corridors that move power from energy-rich western regions to demand centers in the east. By mid-2024, Chinese officials reported cumulative investment of around 6.1 billion dollars under the initiative, with roughly 1.95 million server racks installed and about two-thirds of that capacity already in use. Private capital has followed, as investors shift from crowded AI chip plays into power, grid infrastructure, and metals associated with data center growth.
Official statements describe Eastern Data, Western Computing as a way to balance regional development and reduce strain on coastal grids. At the same time, technical publications and procurement documents make clear that the hubs are expected to serve national objectives in security, industrial policy, and AI development. Beijing is experimenting with marketplaces that allow workloads to be shifted between clusters while keeping control in state-aligned hands. Cloud and AI infrastructure is being treated as a strategic resource that requires centralized coordination across fiber, power, and compute.
China’s posture on critical minerals and materials fits the same pattern. The sequence of export controls on gallium, germanium, rare-earth-related technologies, and other inputs since 2023 signals a willingness to use deep-supply-chain leverage in response to U.S. and allied semiconductor restrictions. The recent partial relaxation under a trade truce does not remove the underlying structural dependence or the possibility of future tightening.
None of this makes China immune to its own vulnerabilities. Eastern Data, Western Computing must cope with regional water constraints, local resistance to new transmission lines, and the technical challenge of coordinating large distributed clusters. The domestic power grid faces its own stresses from electrification and climate-driven extremes. Nevertheless, the initiative reflects a view that compute should be integrated with national energy and industrial planning, and that control over transformers, lines, and materials is part of that effort.
In the United States, by contrast, the AWS complex that failed in October sits at the intersection of local zoning decisions, state-level utility regulation, federal export controls, and private capital flows. The Department of Defense and other federal agencies ride on top of that matrix but do not govern it. No single actor is responsible for treating the combination of grid infrastructure, cloud campuses, and supply chains as a coherent strategic system.
Allies and the new dependency chain
The dependence on U.S.-centered cloud infrastructure is not purely domestic. American allies, particularly in Europe and East Asia, are now deeply embedded in the same ecosystems. Ministries in Tallinn, Warsaw, Tokyo, and Canberra use many of the same platforms, and sometimes the same geographic regions, as U.S. companies. NATO experimentation with federated clouds for command and control relies heavily on services delivered from U.S. or U.S.-linked data centers. For smaller states, participation in these ecosystems has been part of the implicit bargain of alliance, alongside troop presence and defense cooperation.
Outages like the one in northern Virginia underscore the asymmetry in control. From the perspective of a European or Asian capital, an incident that begins in a U.S. region can slow or degrade services that support basic functions far away, from payroll processing to emergency communication. Measurements of the October event show that traffic rerouting and congestion affected users as far as Singapore as dependencies on us-east-1 propagated through global architectures.
In political systems where public debate already questions overreliance on U.S. digital platforms, visible episodes of fragility can feed momentum for “sovereign cloud” initiatives. European policymakers, for example, have pushed projects under the general banner of digital sovereignty, ranging from regulatory constraints on data flows to efforts to build European-controlled cloud offerings. Large providers have responded with region-specific arrangements, including commitments to keep European customer data within the continent under local legal regimes and to create sovereign public and private cloud configurations operated by European personnel.
These developments have an ambivalent security effect. In principle, diversified infrastructure could reduce exposure to single points of failure in U.S. regions. In practice, smaller national clouds may lack the scale, redundancy, and security investment of major hyperscalers. Fragmentation can complicate joint operations, as forces struggle to federate workloads and data across heterogeneous platforms in a crisis.
From Washington’s viewpoint, this is a strategic issue, not merely a matter of technology policy. If U.S. data centers and supply chains are perceived as robust and well managed, allied dependence on them reinforces American influence and underpins collective deterrence. If they are seen as brittle or under-resourced, the same dependence becomes a liability that adversaries can highlight, both in their planning and in their political messaging to third countries.
A more explicit conversation with allies about shared compute dependencies, contingency planning, and joint investment could help manage this tension. At present, many of these questions are handled in separate tracks: defense procurement, digital regulation, industrial policy. The AWS incident suggests that they belong in a single discussion about alliance resilience.
The strain of emerging sensing and the risk of expensive blindness
Looking ahead, the tension between front-end sensing and back-end compute is likely to sharpen. Emerging systems in quantum sensing, persistent space-based monitoring, and high-resolution environmental observation promise richer information about adversary movements and the operating environment. Quantum magnetometers for submarine detection, gravity gradiometers mapping mass distributions, hyperspectral imagers detecting camouflage and subtle changes in vegetation, and dense multistatic radar networks all produce continuous streams of high-volume, noisy data.
The value of these systems depends less on the sensitivity of individual instruments than on the capacity to transport, store, and process their output at speed. If the underlying compute and network substrate cannot keep up, the result is not improved awareness but an overwhelmed analytic apparatus. Data piles up faster than it can be processed, operators are buried in false positives, and decision-makers receive fewer actionable insights rather than more.
There is a real risk that states will overinvest in visible, politically attractive sensors while underinvesting in the less visible infrastructure of data centers, fiber routes, and power systems that make those sensors useful. In such a scenario, pressure in a crisis to exploit these capabilities would lead governments to scramble for any available cloud capacity, including commercial resources that were never designed for wartime resilience. Because those same clouds depend on the transformer fleets, materials, and factories already described, the scramble would take place in a structurally constrained environment.
From a strategic standpoint, this dynamic converts industrial capacity in photonics, power electronics, and transformer manufacturing into a lever over the future shape of sensing and targeting. States that can manage the chain from mineral extraction to deployed compute clusters will be better positioned to field and sustain advanced ISR architectures. The October AWS outage did not involve these systems directly, but it hinted at what happens when optimistic assumptions about elastic compute meet the hard limits of physical infrastructure.
A different frame for policy
Taken together, these trends point toward a conclusion that remains politically uncomfortable in many capitals. The cloud cannot be treated as a neutral utility that can be expanded or reconfigured at will through software. It is an enduring strategic asset anchored in metals, machines, and grids that need to be planned, financed, and protected in ways that resemble, more than they differ from, earlier efforts to secure oil supplies, shipping lanes, and heavy industry.
For the United States and its allies, this implies three broad shifts in mentality and practice.
First, policymakers need to stop treating the grid and the cloud as separate policy domains. Transformer backlogs, electrical steel sourcing, and substation permitting are now directly relevant to the resilience of command and control, intelligence processing, and defense logistics. Energy regulators, industrial planners, and defense officials rarely sit in the same room to think about a specific data center interconnection or a transformer plant expansion; yet the strategic stakes of those decisions now overlap. Incorporating large data center loads into transmission planning, and reflecting defense priorities in decisions about grid hardening and equipment stockpiles, requires new institutional arrangements that cut across existing bureaucratic lines.
Second, resilience has to be understood as more than software-level redundancy. Cloud providers already design for failover across availability zones and regions. That logic needs to be extended down into the hardware and material layers. For some clusters, maintaining reserves of large transformers and critical components, pre-positioned within reach and regularly rotated to avoid degradation, may be justified in the same way that fuel and munitions stockpiles have long been treated as non-negotiable overhead. Diversifying suppliers, investing in domestic production of key equipment, and supporting workforce development in specialized manufacturing are slower and costlier than writing new failover code, but they address the constraints that matter when a region’s physical capacity is impaired.
Third, allied infrastructure should be brought into this picture explicitly. If, in practice, a small number of American regions and associated supply chains function as the compute backbone of an alliance, decisions about where to build, where to harden, and where to hold spare capacity should be shared priorities. That could mean co-financing additional transformer manufacturing capacity, jointly identifying data center clusters that merit enhanced physical protection, and coordinating policies toward critical materials and rare earths that recognize their role in digital infrastructure, not just in electric vehicles or missiles.
The October 2025 outage will fade from public memory. Services resumed, transaction queues cleared, and most users filed it mentally alongside previous disruptions. What should not fade is the underlying lesson. The cloud that underwrites so much economic activity, statecraft, and military planning is not weightless. It flows through a narrow canyon of metals, factories, transmission corridors, and regulatory decisions that are exposed to both accident and coercion.
Strategic competition in this decade will revolve around code, narratives, and alliances, but it will also be shaped in transformer plants, steel mills, optic-fiber draw towers, and compound semiconductor fabs. The failure in Virginia briefly illuminated those foundations. The task for policymakers is to treat them not as the background of digital life, but as elements of national power that require deliberate, sustained management.
