The Environmental Cost of Digital Sovereignty: Water, Energy, and Emissions Impacts of Sovereign AI Infrastructure in the Global South
Muntaser Syed, Marius C. Silaghi, Sheikh Abujar, Sharun Akter Khushbu, Amal El Ahmad
cs.CY, cs.DC
2026-07-15
A comparative stress analysis prices sovereign AI in the UAE, Bangladesh, India, and Kenya: the same 1,024-GPU cluster with evaporative cooling drinks 34.3 million liters of water a year in the UAE (water stress 4.82/5), and emits 7.3x more CO2 on Bangladesh's grid than Kenya's geothermal one.
Between 2024 and 2026, Global South governments announced over $200 billion in sovereign AI commitments: the UAE's Stargate campus plans 5 GW, Saudi Arabia's HUMAIN bought 18,000 Blackwell GPUs, India's IndiaAI Mission has 34,381 deployed. The discourse is geopolitical; the physical bill, water, power, carbon, is nearly absent from national AI strategies. Bangladesh's draft AI policy for 2026–2030 contains zero mention of environmental siting or water.
This paper, submitted to GHTC 2026, does the accounting. The method is simple by design: WRI Aqueduct water-stress data, Ember grid carbon intensities, and published NVIDIA DGX H100 power specs (10.2 kW per 8-GPU node), modeled for a 1,024-GPU cluster (1.31 MW IT load) across PUE 1.20–1.80 and three cooling regimes. The authors are explicit that these are order-of-magnitude estimates, not engineering-grade predictions.
Four cases span fiscal capacity × environmental vulnerability: UAE (rich, extremely water-stressed), India (resourced, highly vulnerable), Bangladesh (constrained, highly vulnerable), Kenya as the African case (constrained, less vulnerable). Five stress vectors are assessed per case: water, energy, carbon, land and siting, and resource competition with human needs. The arithmetic is multiplicative: IT load × PUE gives facility power; × WUE (1.8–3.0 L/kWh for evaporative cooling) gives annual water; × grid carbon intensity gives annual CO2.
The paper also scores 15 actual or planned data center sites (Dhaka, Chennai, Mumbai, Stargate Abu Dhabi, Nairobi Konza, and others) on five climate-vulnerability dimensions, and compares the environmental footprint of four AI development paths: frontier pre-training, a sovereign LLM run, LoRA fine-tuning, and edge inference.
| Metric (1,024 GPUs, PUE 1.80) | UAE | Bangladesh | India | Kenya |
| Annual water, evaporative (M liters) | 34.3 | 32.0 | 22.9 | 13.7 |
| Annual CO2, evaporative (tons) | 9,624 | 14,330 | 13,795 | 1,964 |
| Grid carbon intensity (gCO2/kWh) | 467.5 | 696.1 | 670.1 | 95.4 |
| WRI water stress (of 5) | 4.82 | 1.38 | 3.05 | 0.74 |
Key readings:
The value is not the method, which is a multiplication table, but the assembly of scattered public numbers onto one page where the slogan "sovereign AI" can be priced. Three conclusions carry weight for practitioners:
Siting and cooling technology determine footprint more than cluster size, a 60x spread in water and 7x in carbon, depending on where and how you build. Outsourcing compute is genuinely lower-carbon; countries either accept that premium or hedge with clean-grid siting, where Kenya is the positive template. And for resource-constrained states the defensible path is frugal AI, small-model fine-tuning and edge inference, rather than stretching for frontier pre-training.
Seven design principles follow: mandatory WUE disclosure, climate-vulnerability siting assessments, renewable-first siting, frugal-model defaults for constrained countries, regional shared facilities, environmental impact assessments written into national AI strategies, and desalination-loop accounting for arid deployments.
The authors list theirs plainly: capex data is incomplete and promotional; water and carbon figures use national averages, not site measurements, so results support cross-country comparison but not single-facility design; of the three driving parameters, grid carbon intensity dominates the spread (7.3x across the four countries), though the country ranking is robust across the full parameter space.
Points that stand out on a close read: GPU power from TDP is a lower bound; diesel backup generation (2–3x grid carbon intensity) is mentioned but never enters the quantitative model; the scoring of 15 sites stitches together heterogeneous sources with unexplained weighting; and the "trilemma" framing, sovereignty, sustainability, and affordable citizen access cannot all be maximized, is stronger than four cases can strictly prove; the cases show tension, not impossibility.
Worth noting the authors' position: they cite their own prior work on transistor-level ML energy efficiency, and the paper leans toward a humanitarian-engineering lens. That does not affect the arithmetic, but it shapes emphasis, the emissions advantage of outsourcing gets more ink than its sovereignty costs.