Sovereign AI in Global South could consume 30M liters water per cluster, study warns

A new paper from researchers including Muntaser Syed (arXiv:2607.13443) examines the environmental cost of sovereign AI infrastructure in the Global South, where over $200 billion in state-led AI commitments have been announced between 2024 and 2026. Using public water stress data, grid carbon intensity factors, and GPU power specs, the study models hypothetical deployments across four regions: UAE, Bangladesh, India, and Kenya. In the UAE, a single 1,024-GPU cluster using evaporative cooling would consume over 30 million liters of water annually in a country classified as 'extremely high' water stress. Bangladesh, meanwhile, has sovereign AI policy documents calling for centralized GPU procurement but offers no guidance on where to site data centers in a nation where more than a fifth of land floods annually and the power grid struggles with reliability.

The authors identify a 'sovereignty-sustainability trilemma': no country can simultaneously maximize AI sovereignty, minimize environmental impact, and maintain affordable resource access for citizens. They propose design principles for environmentally responsible sovereign AI, including mandatory water usage effectiveness (WUE) reporting, climate-vulnerability siting assessments, and a preference for frugal small language models over frontier pre-training in resource-constrained settings. The study underscores a growing tension between digital sovereignty ambitions and the physical infrastructure costs that fall hardest on countries least equipped to absorb them.