The physical stack: electricity, water, land and heat
Compute is continuous, concentrated and quality-sensitive. Modern AI facilities require not simply large quantities of electricity but dependable electricity, transmission capacity, voltage stability, redundancy, cooling and emergency backup. Grid hardware can be as constraining as generation: large transformers, switchgear and related components can carry procurement lead times measured in years.
Water analysis requires equal discipline. Withdrawal, consumption and discharge are different quantities; on-site cooling water should be distinguished from water consumed indirectly in electricity generation. The paper therefore cautions against universal “litres per query” claims and evaluates water impact in the context of local climate, technology and power mix.
The physical stack extends further upstream and downstream: semiconductor fabrication, metals, embodied carbon, fibre routes, subsea cables, hardware lifecycle, waste heat and network concentration all shape the true infrastructure requirement.
Different systems, converging policy questions
United States. Compute demand is increasingly moving upstream into generation, including nuclear uprates and dedicated supply. The policy challenge is to avoid socialising grid and water costs while preserving a credible pathway for strategically important capacity.
China. “East Data, West Computing” demonstrates the advantage of coordinating compute geography with energy geography. The model does not eliminate constraints, but it treats computing power as an infrastructure-planning problem rather than a stand-alone property development.
European Union and Ireland. Europe is moving from voluntary sustainability claims towards mandatory measurement. Ireland has gone further by conditioning new grid access on local generation, storage and additional renewable supply.
Australia. The debate is increasingly about grid stress, water, emissions and social licence. A technically viable facility can still become politically unviable when local impacts are poorly allocated or explained.
South Africa and wider Africa. The issue is not “population first, data centres later”. It is whether scarce infrastructure is expanded rather than captured, and whether strategic digital investment improves access, reliability, affordability and local benefit.
Taiwan. Compute sovereignty intersects directly with island-grid resilience and the semiconductor supply chain, making energy-utilisation review and industrial-benefit tests strategically significant.
Orbital compute. Space-based computing may eventually reduce some terrestrial land and water pressures, but it shifts externalities towards launch demand, radiation, thermal rejection, spectrum, debris and orbital governance.
The Infrastructure Sovereignty Test
The paper proposes a recurring policy test for large compute projects: scale and speed of growth; electricity-system capacity and reliability; water exposure; grid-hardware and network dependencies; climate and operational resilience; regulatory transparency; additionality and cost allocation; and local, distributional and strategic benefit.
The strongest common principle is additionality. Renewable certificates or annual accounting alone are not enough where a physical grid is constrained. Regulators should ask whether new load adds dependable capacity at the relevant location and time, including 24/7 Carbon-Free Energy and hourly matching where appropriate.
Infrastructure sovereignty is not autarky. It is the capacity to host strategically important compute without degrading essential services, imposing uncompensated costs on households or becoming dependent on infrastructure the state cannot effectively govern.
Policy recommendations
The recommendations are grouped into four policy families: Transparency & Reporting; Grid Integrity, Additionality & Flexibility; Planning, Resilience & Community Rights; and Industrial & Geopolitical Strategy.
They include facility-level energy and water reporting, dual water accounting, cost-causation rules, additionality requirements in constrained systems, dynamic tariffs and curtailment arrangements for flexible workloads, cumulative infrastructure assessment, climate-resilience tests, early disclosure of local impacts, hardware lifecycle plans, and public-value conditions on incentives.
Conclusion
The relevant policy choice is not between embracing AI and protecting electricity or water. The real task is to allow strategic compute to grow while forcing its physical costs into the investment decision rather than externalising them to households, existing industry or future public budgets.
Where infrastructure is abundant, governments should enable growth. Where it is constrained, hyperscale compute should arrive with additional infrastructure, transparent resource accounting and a credible local-value proposition. AI may remain digital in output, but its strategic competition is increasingly physical.