The hidden costs of AI’s data-centre boom’

Artificial intelligence (AI) is transforming how organisations work but the “cloud” that supports it is not a cloud at all. It is a global network of physical data centres: concrete facilities packed with high-density servers, drawing on power grids, water systems and land. As generative AI moves from research labs into everyday consumer products, the demand on that infrastructure is growing in ways the public conversation has not yet caught up with. In a study presented at the 2025 Americas Conference on Information Systems in Montreal, Canada, my co-authors Laura Watkowski (University of Bayreuth, Germany), Jenny Elo (University of Jyväskylä, Finland) and I set out to map what AI’s data-centre boom is doing to the societies hosting it. Drawing on interviews with industry experts and a structured review of media reporting, we identified five systemic tensions: the energy paradox, water strain, hyperscaler dominance, sovereignty erosion and urban displacement. They are interlocking and intensifying each other. The cost of AI The figures the paper documents are striking. Microsoft’s own sustainability reporting acknowledged that its greenhouse gas emissions rose roughly 30% from a 2020 baseline, driven largely by AI infrastructure, a notable departure from the climate pledges the major hyperscalers had set themselves before the generative AI cycle. By 2023, the major hyperscalers (Amazon, Google, Microsoft and Meta) operated close to 992 data centres globally, with capacity having doubled in just four years. A single new hyperscale facility can draw as much electricity in a year as the demand from 350 000 to 400 000 electric cars and projections cited in our analysis suggest global data-centre electricity consumption could roughly double to 1 065 terawatt-hours by 2030. Water is the second front. As compute densities rise, more facilities rely on liquid cooling drawn from already stressed watersheds. Projections in our review suggest global AI-driven demand could push data-centre freshwater consumption above one trillion gallons by 2027. Those numbers translate into local effects. Dublin, where data centres now consume close to a fifth of Ireland’s electricity, faces a de facto moratorium driven by public pushback. Amsterdam and Singapore have paused new builds. Outside Washington, DC, residential land is being repurposed for industrial use to make room for the next campus. Ireland briefly halted new grid-connected construction because of energy and sustainability concerns, only to reverse the decision under industry pressure. Japan is reviving nuclear power in part to keep the AI-relevant industry competitive. These are not anti-technology gestures; they are governments and communities trying to reassert authority over infrastructure that has, until now, been treated as a purely commercial decision. The good news There is real value in naming these tensions clearly and that is what our paper attempts to do. The argument is not that AI computing should stop expanding. It is that the dominant story we tell about it — for example, that more computing is straightforwardly good, that hyperscalers will internalise their own costs and that climate targets will hold — is incomplete. We borrow from organisation studies (Smith and Lewis, 2011) the idea of treating these contradictions as paradoxes: dualities to be navigated rather than problems to be solved by picking a side. Treating sustainability and scalability as a binary trade-off invites bad choices in both directions. Treating them as a paradox invites investment in solutions such as efficient hardware, additional renewable supply, transparent reporting and real planning frameworks that hold both sides in view. Recognising tensions is, in this sense, a precondition for acting on them. The limits of the framing It is worth being honest about what a study like ours can and cannot do. Our work is qualitative and exploratory. Six interviews and 14 industry sources can map a phenomenon; they cannot quantify its effect in any specific country, model the cost-benefit trade-offs of a specific facility or predict how policy interventions will play out. We offer what is sometimes called “a theory of the problem”: a way of seeing, not a complete answer. It is also important to notice what the framing itself chooses. By calling hyperscaler dominance a “tension”, we accept that its benefits — such as scale, efficiency and lower per-unit cost — are real. Others might reasonably argue that what we describe as a tension is in fact a market failure that should be regulated as such. By calling community displacement a “tension”, we treat it as something to be balanced rather than prevented. These are conceptual choices, and they are worth contesting in the open. An African angle The empirical work was conducted in South Africa, which is fitting: we are increasingly part of the s