TECHNOLOGYJuly 24, 2026

Data centers become the new frontlines of global power

Reading time: 8 min

Attacks on Amazon’s cloud facilities in the Gulf underscore how data centers have moved from digital infrastructure to contested strategic assets.

An Amazon Web Services facility in Northern Virginia, U.S., a region renowned as the world’s largest data center market, is commonly referred to as Data Center Alley. © Getty Images

In a nutshell

Geopolitics is carving the world into rival compute empires
Data centers are evolving from real estate into regulated infrastructure
Control of data centers now means economic and military power
For comprehensive insights, tune into our AI-powered podcast here

Data centers have emerged as a crucial element in the realms of geopolitics, energy security and industrial strategy. On the morning of March 1, 2026, three coordinated Iranian drone strikes hit Amazon Web Services facilities in the Gulf, disrupting regional cloud services for days to weeks. The episode crystallized a transition that had been building for years: The quiet warehouses of the digital economy are now contested strategic assets.

The scale of what is being built is difficult to overstate. Meeting projected demand for computational power will require approximately $6.7 trillion of cumulative investment worldwide by 2030, of which roughly $5.2 trillion is tied to artificial intelligence. Hyperscale operators alone plan to spend about $630 billion on capital expenditure in 2026, more than four times the level recorded in the year of the OpenAI GPT-4 launch. The figure exceeds the combined defense budgets of France, Germany and the United Kingdom, and it is being committed by a handful of private companies operating outside the conventional perimeter of strategic planning.

The market itself has expanded accordingly. Estimates converge on a global revenue base of $380 billion to $395 billion in 2025 and a trajectory toward $700 billion to $900 billion by the early 2030s, implying a compound annual growth rate of 11 to 13 percent. Data-center systems spending alone reached $489.5 billion in 2025, a 46.8 percent increase over the previous year. Such growth is no longer an extrapolation of digital transformation; it is the physical correlate of generative AI, with AI-related workloads now accounting for roughly 19 percent of total cloud spending, up from 8 percent two years earlier.

Power becomes the new bottleneck

This expansion has shifted the locus of strategic risk. The bottleneck is not financing, talent or even chips. It is power generation. The International Energy Agency projects that data-center electricity demand will rise from 415 terawatt-hours in 2024 to roughly 945 terawatt-hours in 2030 in its base case, with a high-demand path approaching 1,200 terawatt-hours. Goldman Sachs Research forecasts a 165 percent increase in data-center power consumption by 2030.

Grid interconnection in primary markets (larger data center hubs) in the U.S. and Europe now routinely takes four to seven years; in northern Virginia, the world’s largest cluster, queues are longer still. Vacancy in U.S. primary markets stood at a record 1.4 percent at the end of 2025, while project cancellations rose from six in 2024 to 25 in 2025, almost all driven by power and grid constraints rather than tenant demand.

The response has been a vertical reintegration of compute, energy and finance. Microsoft has signed a 20-year contract to restart the Three Mile Island nuclear plant. Amazon has acquired a nuclear-adjacent campus from Talen Energy in Pennsylvania. Google has partnered with Kairos Power on small modular reactors. More than 22 gigawatts of nuclear capacity have been earmarked for data-center use globally, with the first commercial small modular reactor deployments expected in the early to mid-2030s.

Oct. 16, 2024, Pennsylvania, U.S.: Microsoft has signed a 20-year deal to restart the Three Mile Island nuclear plant to power its data centers. © Getty Images

Behind the meter, gas turbines, fuel cells and grid-scale batteries are being deployed at a pace that recalls the build-out of independent power producers in the 1990s. Hyperscalers – massive cloud providers such as Amazon, Meta, Microsoft and Google, which operate at global scale – are no longer customers of utilities; in many markets, they are becoming utilities.

This change carries political consequences. Communities that once welcomed data centers as low-impact employers now contest them as energy-intensive neighbors. Ireland has imposed a de facto moratorium on new connections in the Dublin region. The Netherlands has restricted new builds in the Randstad. Several American states, including Virginia and Georgia, are reviewing how grid upgrade costs are allocated between data-center tenants and household ratepayers. Water consumption has emerged as a parallel concern: a single hyperscale facility can withdraw between 11 and 19 million liters per day. Sustainability is no longer an investor-relations exercise but a license-to-operate question.

Geopolitical fragmentation and sovereign AI

The drone strikes on AWS facilities in the Gulf accelerated a fragmentation that had been visible for several years. Compute infrastructure is now treated as critical infrastructure subject to sovereignty considerations comparable to those traditionally applied to ports, telecommunications or energy. The European Union has launched seven AI factories across 17 member states. India, Japan, Saudi Arabia and Norway have announced sovereign AI initiatives. The number of distinct compute blocs, anchored by separate chip stacks, regulatory regimes and tenant pools, is transitioning from a binary U.S.-China divide toward five to seven semi-autonomous spheres.

A second structural shift concerns the workload mix itself. Until recently, public attention focused on training: the construction of frontier models in centralized, gigawatt-scale campuses. By 2027 to 2028, inference, the deployment of trained models in real-world applications, will eclipse training as the dominant category of processing power.

Inference is fundamentally different from training. It is latency-sensitive, distributed and tied to where users actually are. It scales with the diffusion of AI through enterprises, vehicles, industrial systems and consumer devices, rather than with the production of new models.

By 2035, it is anticipated that inference will make up over 70 percent of AI computing, leading to a major realignment in where capacity needs to be located. The traditional model for deploying new sites – a 50- to 100-megawatt campus situated on agricultural land near metropolitan areas – is increasingly out of sync with the workloads it aims to support. Latency-sensitive applications, including real-time copilots, autonomous vehicles and industrial robotics, impose strict round-trip time limits (the maximum duration allowed for data to travel to the AI server and back) that campuses several hundred kilometers away simply cannot meet. As a result, there is a growing need to deploy capacity in locations facing the greatest challenges with land availability, power sources and regulatory approvals.

More by Henrique Schneider

Labor, water, lead times and war

Four secondary risks compound this picture. Labor markets are under acute strain, with industry estimates indicating roughly 50,000 unfilled construction and operations roles globally. Shortages of high-voltage electricians, mechanical fitters and certified operators are most severe in Europe and North America.

Water consumption is politically contested in Arizona, the Iberian Peninsula and parts of Asia, where cooling-related withdrawals compete with municipal and agricultural demand.

The supply chain for transformers, switchgear and high-voltage cables has become a structural bottleneck: lead times that stood at roughly 12 months in 2022 now extend to 30 to 60 months in 2026, with prices up by 30 to 60 percent.

And the March 2026 Gulf strikes have made physical hardening, geographic dispersion and bloc-level redundancy a board-level concern rather than a facilities matter.

The next decade will be defined less by how much computational power is built than by who controls it, where it is located and how it is powered. Capital is abundant, but megawatts, transformers and political consent are not. The asset class is transitioning from real estate, with its standardized yields and tenant covenants, into regulated infrastructure, with longer durations, greater integration and closer proximity to the state.

Scenarios

The trajectory of the next decade is sufficiently contested that planning around a single base case is imprudent. Three scenarios bracket the operating environment that investors, operators and governments must prepare for. They are not mutually exclusive; elements of each will appear across geographies and segments.

Likely: Distributed intelligence network

Inference dominance, combined with grid scarcity and the maturation of factory-built modular data centers, is pushing capacity outward. Hyperscale campuses continue to train frontier models, but the marginal new megawatt is being deployed in metro areas, embedded in commercial real estate and industrial sites, in nodes of one to five megawatts that can be installed in eight to 12 weeks.

The total market is projected to reach roughly $1.1 trillion by 2035, while the modular segment is expected to expand from approximately $42 billion to over $170 billion.

Real-estate-style returns compress as value shifts toward equipment manufacturers, energy infrastructure and supply-chain integrators. Power utilities, modular fabricators, liquid-cooling specialists and component vendors capture a disproportionate share of the upside.

The strategic implication is that geographic diversification at the megawatt scale becomes a defensive necessity. In contrast, exposure to specialized supply-chain players becomes the primary source of returns above the cost of capital.

Moderately likely: Sovereign compute fortress

Geopolitical fragmentation accelerates following further infrastructure attacks and tighter export controls. Computers are treated as national infrastructure across all major economies. Vertically integrated operators, often with state participation, control power, chips, cooling and applications inside each bloc, a pattern already visible in the European AI Factories initiative and in Saudi Arabia’s HUMAIN venture.

Sovereignty-compliant capacity in premium jurisdictions trades at sustained premiums of 60 to 100 percent over commodity capacity. Long-duration leases of 10 to 15 years with regulated tenants become standard. The total market is somewhat smaller in nominal terms, around $1 trillion by 2035, but more profitable on a risk-adjusted basis. Investment increasingly resembles regulated infrastructure rather than real estate.

The idea here is that jurisdiction itself becomes a valuable asset, generating yield on its own. Operators that combine reliable regulatory standards with stable energy sources and tenant bases stand to gain significant sovereignty premiums.

Less likely: Efficiency correction

Continued gains in chip performance per watt, on the order of eight to 10 times by 2030, combined with more efficient model architectures, will outrun demand growth. The speculative build-out of the 2024-2028 period results in measurable overcapacity. Vacancy in U.S. primary markets rises from 1.4 percent to 8 to 15 percent by 2030, then settles around 8 to 10 percent.

Capital expenditure pivots from greenfield construction to retrofitting air-cooled facilities for liquid cooling. Asset values in secondary markets decline by 30 to 50 percent below replacement cost. Distressed mergers and acquisitions define the 2030-2033 cycle, with operators that have secured power, investment-grade tenants and low leverage emerging as consolidators.

The risk highlighted by the Institute for Energy Economics and Financial Analysis is that the cost of misallocation falls disproportionately on ratepayers and on lenders that financed speculative capacity against weak-demand covenants.

Contact us today for tailored geopolitical insights and industry-specific advisory services.