Where the cloud touches the earth: the material burden of AI

How AI’s expansion is colliding with the physical limits of power, water, and infrastructure

As artificial intelligence expands in range and use to become the backbone of the digital economy, the servers that sustain it are colliding with the limits of the physical world. Power, water, and minerals, and not data, are now the scarcest inputs in the AI age. The boundless optimism, on the part of the somewhat over-zealous “tech leaders”, of a frictionless digital transformation is giving way to a more grounded understanding: one in which intelligence may be virtual, sure, but its foundations are intensely material.

Every model runs on energy pulled from finite grids; every query depends on coolant, copper, and land. The cloud, it turns out, casts a long shadow, and our bubbled, immersed embrace of technology in the last couple of decades, now crowned by the surge in AI uptake, is feeling some shade.

The scale of the machine

In less than five years, AI has transitioned from the realms of research to a bedrock of daily infrastructure, even if its baseline dependability, or for that matter the uses to which we gleefully put it, are still in question.

The world’s largest cloud providers — Amazon, Microsoft, Google, and Alibaba — have all committed to doubling or tripling their data-center footprints by 2030. Each site, in practice, is an industrial complex in of itself: a veritable steel-and-glass organism exerting its own gravity on resources, drawing hundreds of megawatts of electricity, thousands of tons of concrete, and millions of liters of cooling water. These are not abstract symbols of digital progress or simply quiet commercial activity in suburban industrial parks, but physical systems whose sheer growth is shaping local ecologies and national planning.

According to the International Energy Agency (IEA), global data-center electricity demand could double by 2026 to over 1,000 terawatt-hours, roughly the annual energy consumption of Japan (IEA). Yes, Japan. AI accounts for an expanding share of that load, and here’s a statistic to start with: a single ChatGPT-style query can consume ten times more power than a typical search engine request, and every generative model trained today leaves a carbon footprint measured in thousands of tons of CO₂. As frontier models scale, that’s to say the most advanced systems that are exceeding the capabilities of current models, these resource inputs compound, reinforcing a feedback loop in which capability drives consumption, and consumption drives a commensurate infrastructure expansion.

And yet, the growth continues relatively unchecked, pulled forward by investment, policy, and that “invisible guiding hand” of free-market speculation. For governments, AI equals competitiveness; for companies, it equals valuation, and the calculus of progress has become a material one, where ever more computation capability demands ever more resources. The tension lies in the mismatch between the sleek narratives of “tech will save us” digital innovation and the slower, heavier realities of the systems required to sustain it. What is marketed as “agile” and “transformative” increasingly depends on infrastructure that is neither.

In short: whatever form these systems, this technology, takes, the digital still exists purely within the constraints of the physical world. Let’s consider several examples to clarify how this plays out.

Case 1: Arizona’s water paradox

The desert city of Phoenix has become a global data-center hotspot. Microsoft, Amazon, and Meta have all expanded their operations there in recent years, despite worsening water stress from the overdrawn Colorado River, an issue that was very present well before the advent of the AI sector. Local reporting has traced this conflict through growing public concern and utility-level negotiations (AZCentral), and rightly so, since each hyperscale facility can require up to 4.5 million liters of water a day for cooling, the equivalent of thousands of households. Utilities and operators have turned to recycled or reclaimed water, but shortages are catching up faster than efficiency gains. In effect, digital expansion is competing directly with urban survival.

The irony runs deeper than the Colorado River: the very infrastructure powering climate modeling and drought forecasting depends on a resource it helps deplete, and with more projects planned along the I-10 corridor, Arizona’s “cloud economy” is becoming a literal drain. The region’s delicate balance between tribal water rights, the needs of agriculture, and significant urban growth now includes an industrial sector whose needs are both immense and continuous. Cooling is not optional; it is the condition for computation.

In this sense, Arizona illustrates a wider structural tension: the incentives that draw hyperscale server operators to resource-stressed regions rarely align with long-term ecological stability; these are simply linear “cause and effect’ incentives developed to attract immediate economic inflows, with no long-term consideration. Sunlight, flat land, and tax advantages may benefit companies, but the externalities, as always, fall on communities forced to deal with growing scarcity. And as demand rises, the politics of water allocation become even more complicated, while the associated costs of digital growth become harder to ignore.

Case 2: Ireland’s grid strain

Across the Atlantic, Ireland has arguably become Europe’s data-center capital, and a distinct warning sign. Dublin’s cluster already consumes 21 percent of the nation’s entire electricity, and this is projected to reach 30 percent by 2030. In response, the national grid operator has imposed connection limits in some regions and ordered new renewable-integration requirements. The consequences are already visible in planning delays, constrained grid capacity, and rising debate over whose demand should be prioritized (Central Statistics Office).

Ireland’s dilemma represents an energy hierarchy in miniature: what once symbolized modernization, the idea of attracting the infrastructure of global technology firms, now strains public services and stated national climate targets. When power is scarce, choices have to be made, and what once sounded like abstract “innovation” becomes a real, practical question: who gets electricity, and who doesn’t? This isn’t just an engineering problem, something to magic away with clever tech solutions; it’s a governance issue. Data centers are built in a few years, while power grids take decades to establish. That gap creates a certain tension that no efficiency tweak can fix.

What’s important here is how Ireland highlights that concentrated digital infrastructure can definitely distort national planning. A country with limited flexibility in its ability to generate power is finding its energy future increasingly shaped by the needs of a single industry, one tempted in, again, for the promise of its immediate economic inflows. And what’s more, even with additional renewable energy sources coming online, data centers still need a steady, always-on power supply. As digital demand grows, that means more pressure to either build new power, import it at greater cost, or simply divert it from homes and other existing industries.

Case 3: Singapore’s controlled reboot

Singapore suspended data-center approvals for three years after recognizing that its limited land and energy supply could not sustain unchecked growth. In 2022, it lifted the moratorium with strict caps: namely that new projects must meet stringent efficiency thresholds and integrate renewable sources. The policy evolution marked a shift from the straightforward pursuit of economic expansion to one of regulation, showing that sustainability doesn’t come automatically with technology, but has to be imposed deliberately. (EDB Singapore).

The city-state’s approach demonstrates something interesting, that restraint can be a strategy. Yet the inexorable global push to build AI capacity undercuts this approach, since limits in one country, such as in Singapore, simply shift investment elsewhere, sending infrastructure toward the places with the fewest restrictions. And Singapore’s rules have already reshaped regional competition, seeing neighboring countries market themselves as alternative hubs, even when their power grids or water supplies are already under strain.

Singapore’s strategy underscores an uncomfortable truth: as long as AI development prioritizes scale above all else, sustainability becomes a moving target. Efficiency improvements matter, although they rarely outpace the steep upward curve of demand, and it must be noted that this kind of responsible slowdown currently remains the exception, not the rule.

Material limits

Behind every server hall lies a deeper extraction chain. AI hardware requires vast quantities of copper, lithium, and rare earth elements for chips, wiring, and cooling systems, and semiconductor fabrication is among the most resource-intensive industrial processes in the world, relying on enormous electricity loads and complex water treatment systems. These realities create a second-order vulnerability: even if a data center operates efficiently, the upstream infrastructure enabling it may not.

Policy responses are emerging. In China, data-center growth is increasingly shaped by energy constraints and the geography of power supply, with renewable resources concentrated in some regions and demand in others, requiring long-distance transmission and policy planning to manage rising electricity use. (Carbon Brief). Likewise, Kenya is pushing for stricter international accountability around the environmental impact of data centers, calling for clearer emissions reporting, greater use of renewable power, and global standards that reflect the real energy and water costs of AI infrastructure. (Daily Nation).

Physics, however, still sets the floor. Every watt consumed has to be generated somewhere, and every unit of waste heat has to go somewhere else. No matter how smart each frontier AI advance is, it must still conform to the laws of thermodynamics: cooling systems can be made more efficient, for sure, but they cannot be eliminated.

The hidden footprint

The least visible cost of artificial intelligence is also one of the most consequential: water. Large AI systems don’t just consume electricity; they depend on vast quantities of freshwater to keep servers and power plants cool. Analysis found that training a single GPT-class model can require hundreds of thousands of liters of water, used both directly in data-center cooling and indirectly through electricity generation. (MIT). Renewable energy helps to improve the math by reducing emissions, but it does not remove this pressure for such volumes of water. Wind and solar power are intermittent, which means more infrastructure, more storage, and more redundancy, all of which extend the material and water footprint of AI systems. These costs are not abstract, either, since, based on resource availability, they shape where data centers can be built, which regions absorb the strain, and who ultimately pays for digital growth.

For lower-income countries courting data-center investment, the trade-off is often stark. Jobs, connectivity, and tax revenue come bundled with long-term demands on land, water, and already-stressed grids. So, in practice we can see that AI shifts environmental pressure to the places hosting its hardware, even when the benefits flow elsewhere.

The return of thermodynamics

How about them laws of thermodynamics? You just can’t beat them for consistency. The digital economy’s most inconvenient truth, as we have seen, is that it cannot escape physics. Information generates heat; computation requires energy; and every gain in machine intelligence rests on material systems that still obey thermodynamic limits. The more intelligence is outsourced to machines, the more dependent it becomes on extractive infrastructures that efficiency alone cannot overcome.

See: AI and the capacity to govern

Technical improvements do, of course, matter, and advances in liquid cooling, modular data-center design, and smarter workload scheduling all reduce waste at the margins. However, growth, undoubtedly fueled by rampant AI market speculation, continues to outpace those gains. Even AI applications specifically designed for sustainability, from climate modeling to optimized building design, all rely on compute-intensive processes that expand the system’s overall footprint. Inescapably, at scale, the tools designed to improve environmental outcomes risk amplifying the pressures they are meant to solve.

Designing restraint

Some governments are beginning to accept that efficiency is not the same as stewardship. Singapore’s conditional data-center moratorium, Europe’s emerging green-data-center standards, and growing U.S. debates around “compute accountability” all point toward a common conclusion: technological ambition has to be constrained by resource limits, and should not be assumed to transcend them.

GYST has explored this pattern before, particularly in the context of “Industry 5.0: geopolitics in the age of intelligent machines” where technological systems are no longer just tools but strategic infrastructure whose scale and consequences demand political oversight rather than market autopilot. 

The open question is whether those constraints arrive in time, and whether governance can move faster than infrastructure investment, which, if precedent is cited, would likely not be the case. If current trajectories hold, the energy systems of the 2030s may simply resemble digital extraction zones that exhibit the same characteristics of our industrial expansion in the last couple of centuries: vast, humming complexes where silicon replaces steel, and the cloud imposes real, local costs on every region that hosts it.

The warning is simple. Shortcuts built on speed, opacity, or overconfidence offer control in the moment, but fragility over time. This, once more, is the same conclusion we find in our analysis of whichever geopolitical, economic, or technology issue we research. Sustainability does not emerge automatically from better technology, rather, it has to be designed, enforced, and, above all, slowed down a little to better fit the world it depends on. Everything else burns too hot to last.

Read this. Notice that. Do something.

Read this: International Energy Agency — Electricity 2024, a sober overview of how data centers and AI are reshaping global electricity demand, grid planning, and energy security.

Notice that: Read MIT’s clear, accessible explanation of how large AI models consume freshwater through cooling and power generation, and why this cost is largely invisible in public debate.

Do something: World Resources Institute (WRI) — Aqueduct Water Risk Atlas, use this to see where data centers and grid infrastructure intersect with existing water stress and scarcity.

Previously on GYST: Supply-chain shock #4.0: tariffs, trade, and the new geography of risk

Next up: Water shock in Asia: climate, power & the cascading infrastructure crisis