Artificial intelligence no longer competes only in the realm of algorithms and capital. It competes for rivers, power grids, and the right to draw watts from a national grid. The nations that understand this are rewriting the rules of industrial policy. The ones that don’t are already losing ground.

In the summer of 2023, Montevideo ran out of safe drinking water. The culprit was drought—but the accelerant, officials later acknowledged, was a planned data centre that would have drawn heavily on the Río de la Plata basin during peak demand. The facility never opened; the city’s taps turned saline anyway. It was a preview. The geopolitics of AI—long framed as a contest over algorithms, capital, and export-controlled chips—has acquired a harder, more physical character. It is now a fight over water, electricity, and the land beneath both.

That shift matters for everyone from Pentagon planners to municipal water boards in Phoenix. The compute infrastructure powering the AI boom is not weightless. It is anchored to specific places, draws on finite natural resources, and strains grids that were never designed for it. The countries and regions that control those resources—or can build grid capacity fastest—are accumulating a structural advantage that no number of AI researchers can offset.

The scale of what’s being built is still poorly understood outside a narrow circle of energy analysts and infrastructure investors. Data centres supporting AI operations are projected to consume 1,580 terawatt-hours per year of electricity by 2034—a figure comparable to India’s entire national power consumption today. That projection comes from FP Analytics, drawing on IEA modelling, and it was published before DeepSeek’s January 2025 breakthrough suggested that inference costs might fall sharply, potentially accelerating adoption and driving even more aggregate demand.

The water dimension is less discussed and arguably more alarming. Global data centres consumed an estimated 560 billion litres of water in 2023 for cooling alone, according to the International Energy Agency. A peer-reviewed analysis published in late 2025 put the AI sector’s water footprint at between 312.5 and 764.6 billion litres by year-end 2025—and that range reflects genuine uncertainty about how fast inference workloads are scaling, not a methodological flaw. The honest answer is that nobody knows exactly how thirsty AI is, because tech companies’ environmental disclosures remain inconsistently audited.

1,580 TWh Projected annual electricity demand from AI data centres by 2034 — roughly equivalent to India’s current national consumption. Source: FP Analytics / IEA modelling, 2025.

The most vivid case study is also the most embarrassing for a tech industry that prides itself on rational planning. Northern Virginia—”Datacenter Alley”—handles approximately 70 percent of global internet traffic. Dominion Energy, the regional utility, projects that summer peak load will increase by 70 percent between 2022 and 2045, driven almost entirely by data centre demand. The grid was not built for this. It cannot be upgraded fast enough without significant capital commitments that ratepayers—not shareholders—will largely absorb.

Ireland tells a similar story from a different angle. Data centres accounted for 21 percent of Ireland’s total metered electricity in 2023, exceeding all urban households combined. Dublin’s grid operator paused new approvals until 2028. What followed was effectively a forced regulatory evolution: new facilities must now generate their own power on-site, export excess capacity back to the grid, and commit to 80 percent renewable procurement within a set period. In practice, this means technology companies are becoming utility operators—a structural shift with no clear precedent in industrial history.

Mexico’s Querétaro state and Uruguay’s capital offer cases where water stress and data centre expansion collided directly. In both instances, the draw on aquifers during drought conditions forced local authorities into uncomfortable trade-offs between digital infrastructure investment and basic residential water security. Accelerated AI adoption could result in an additional 4.2 to 6.6 billion cubic metres of water withdrawal by 2027, including both on-site cooling and electricity generation upstream. That figure, from WestWater Research, covers the US alone.

What makes these cases geopolitically significant is not their local drama but their systemic implication: the placement of compute infrastructure is no longer a purely commercial decision. It is an act of resource allocation with consequences for communities, national grids, and bilateral relationships.

Western policy has focused obsessively on semiconductor export controls as the primary lever for managing AI competition with China. That focus is rational but incomplete. The control of compute power—where it is built, who can access it, and on what terms—has a physical layer that chip export rules do not fully address.

Can export controls actually stop China’s AI advance?

Export controls can delay but not decisively stop China’s AI development. They restrict access to leading-edge chips, keeping Chinese labs dependent on lower-performance hardware. Yet China has closed much of the capability gap through model efficiency gains, achieving near-parity on benchmarks despite compute constraints—suggesting that raw chip access is a limiting but not determining factor.

Since October 2022, the US has imposed successive waves of export controls on advanced semiconductors. The January 2025 AI Diffusion Rule divided the world into three tiers, imposing hard caps on GPU imports and AI model weights. The Trump administration then rescinded the most stringent provisions in May 2025, re-restricted H20 sales to China in April, reversed course again in July, and by December had announced a scheme allowing Nvidia to sell H200-class chips to China in exchange for a 25 percent revenue stake. The incoherence has been, as Chatham House observed in April 2026, the “worst of both worlds”—damaging US commercial interests without achieving clear strategic goals.

Still, the controls have had measurable effect. Huawei produced only around 200,000 AI chips in 2025, according to US Commerce Secretary Howard Lutnick’s congressional testimony. Meanwhile, Nvidia‘s Blackwell-generation systems are being deployed in clusters of hundreds of thousands in US hyperscaler data centres. That aggregate compute gap—not individual chip performance—is where the strategic advantage increasingly lives.

Yet China has a structural advantage that chip controls cannot touch: it can build power generation capacity faster than any Western democracy. In 2025 alone, China added over 540 gigawatts of new power capacity, roughly 80 percent of which was solar and wind. The US, by contrast, faces permitting timelines measured in years and grid interconnection queues stretching into the 2030s. Brookings’ April 2026 analysis flagged energy as the “first gap” in America’s AI ecosystem—more acute than the talent or capital shortfalls.

The resource intensity of AI is creating a new class of geopolitical winners and losers that cuts across the traditional developed-developing world divide. Countries with abundant, cheap, low-carbon electricity—Norway, Iceland, Paraguay, Canada’s Quebec province—are seeing data centre investment that would have been unthinkable a decade ago. Countries with stressed water tables and aging grids are discovering that AI ambitions have a hard physical ceiling.

For capital markets, the implications are already visible. Utilities with exposure to data centre demand are trading at premiums not seen since the industrial buildout of the 1990s. In 2025, the largest US technology companies committed more than $300 billion to AI development, hardware, and new data centre construction—a figure that, if sustained, implies total US power demand for data centres roughly doubling by 2030 to 426 terawatt-hours. The investment in nuclear energy—Microsoft‘s revival of Three Mile Island with Constellation Energy being the most prominent example—reflects a sector that has concluded it cannot wait for the grid.

“These companies have effectively decided to become utility operators. The question is whether regulators—or voters—are ready for that.”

— Paraphrased from policy discussions at FP Analytics / World Governments Summit simulation, Dubai, February 2025

For policymakers, the governance vacuum is the central problem. The Paris AI Action Summit in February 2025 produced a framework on inclusive and sustainable AI, but the United States and the United Kingdom declined to sign. Without the two countries that host the most powerful AI infrastructure, any global standard on water disclosure, energy sourcing, or compute access is effectively voluntary. The World Economic Forum noted in mid-2025 that international relations are now defined as much by geotechnology disputes as by traditional territorial ones—but the institutions designed to manage traditional disputes have no clear mandate over data centre siting or GPU allocation.

For smaller economies, the second-order effect is a structural dependency that isn’t yet named as such. When a country’s AI ambitions depend on compute capacity hosted in a foreign jurisdiction—subject to that jurisdiction’s export licensing, its grid reliability, its political stability—it has outsourced a dimension of national sovereignty without a formal treaty to govern it.

The alarm registered in most coverage of AI’s resource intensity is real, but it’s worth engaging seriously with the counter-argument. Several credible analysts argue that the energy trajectory of AI will not follow the straight-line projections. The IEA itself expects that advances in edge computing, quantum computing, photonic microchips, and neuromorphic architectures could each significantly reduce AI’s energy footprint—and if leading AI models accelerate research in those areas, the effect could compound in either direction.

DeepSeek’s emergence is the strongest empirical case for optimism. Its models matched frontier US performance at a fraction of the compute cost, suggesting that the efficiency frontier is not fixed. If Chinese AI labs—constrained by chip access—systematically out-innovate on efficiency, they may inadvertently solve a problem that threatens everyone. Sam Altman acknowledged as much in February 2025, noting that the pressure on compute efficiency was “the most interesting forcing function the industry has faced.”

The water argument also has its limits. Liquid cooling systems are improving, water recycling is becoming standard in newer facilities, and siting decisions are increasingly shifting toward regions with surplus water. The picture is more complicated than “AI drinks rivers.” That said, the governance mechanisms required to ensure responsible siting do not yet exist at the scale or speed the investment cycle demands.

The race for AI dominance has always been described in terms of models, talent, and capital. Those things matter enormously. Yet the contest is now also being fought over kilowatt-hours, aquifer recharge rates, grid interconnection queues, and export licensing regimes that change with each administration’s trade priorities. That is not a metaphor. It is a literal description of where the binding constraints are moving.

Countries that treat AI infrastructure as a purely commercial matter—to be sited by the market and regulated after the fact—are ceding a strategic choice that will be very difficult to revisit. Countries that understand compute capacity as a form of industrial sovereignty, equivalent in long-run importance to port access or electricity generation in earlier eras, are planning differently.

The deepest irony of the AI era may be this: the technology most celebrated for its disembodied intelligence is reshaping geopolitics through the most material of means—water drawn from an aquifer, watts pulled from a line, and the political will to build the infrastructure faster than your rivals.