Insights · Policy brief
The Coupled Constraint
Why AI’s real bottleneck is the energy–water–heat system — and why that makes it a governance problem, not a technical one
Key message
We plan, finance, regulate and defend energy, water and compute as three separate systems. Physically they are one. Every one of them is, at bottom, a heat problem: compute turns electricity into heat that must be rejected with water; desalination spends energy to move water against osmotic pressure; and both increasingly compete for the same power and the same watersheds. Treating this coupled system as three silos is why AI’s water footprint surprised its own builders, why data centres are being sited into water-stressed basins, and why “energy security” and “water security” keep being solved in isolation and failing together. The binding constraint on the AI build-out is not chips, and not power alone — it is the energy–water–heat coupling. That coupling is primarily a governance and risk question, and it can only be managed as one system.
1. Three systems, one physics
Infrastructure policy is organised by sector: an energy ministry, a water authority, a digital or industrial portfolio. Each optimises its own domain. But the three are joined by thermodynamics. Electricity does not disappear when a data centre uses it — the overwhelming majority of it becomes waste heat that has to be moved somewhere, and today that somewhere is largely water. Water, in turn, is increasingly manufactured with energy: in arid regions, drinking water is desalinated seawater, which is one of the most energy-intensive things a utility does. So compute needs water to shed heat, water needs energy to be made, and both draw on the same grid and the same scarce basins. A decision taken in one silo — build here, cool this way, price water that way — silently sets the constraints for the other two.
The failure mode is not dramatic; it is quiet mis-optimisation. Each system looks efficient on its own dashboard while the coupled system accumulates fragility. That is precisely the kind of risk that sectoral governance is built not to see.
2. Water for AI: the compute side of the coupling
The scale is no longer marginal. A single 100-megawatt data centre can consume on the order of two million litres of water per day for cooling — comparable to about 6,500 households (IEA, Energy and AI, 2025). In the United States, direct data-centre water consumption was about 17.4 billion gallons in 2023 and is projected to double or quadruple by 2028 (Lawrence Berkeley National Laboratory, 2024 U.S. Data Center Energy Usage Report). Globally, AI-driven water withdrawal is projected to reach 4.2–6.6 billion cubic metres a year by 2027 — on the order of half the United Kingdom’s annual withdrawal (Li & Ren, Making AI Less Thirsty, 2023; Communications of the ACM, 2025). Most of the water withdrawn for evaporative cooling is not returned to the local supply; disclosed figures from major operators imply roughly 80% is consumed (evaporated) rather than returned.
Two features make this a coupled problem rather than a water problem. First, the indirect footprint from electricity generation dwarfs the on-site figure: LBNL estimated about 211 billion gallons consumed indirectly through power generation in 2023 — roughly twelve times the direct figure — so a data centre’s true water intensity is set as much by how its power is generated as by how it is cooled. Second, and decisively, cooling demand is a direct function of heat: because the majority of the electricity a facility consumes ends up as waste heat that must be rejected, water demand scales with compute. The industry has begun to notice — Microsoft reports that closed-loop, near-zero-evaporation cooling can save around 125 million litres per datacentre per year (Microsoft, 2024) — but the structural point stands: AI’s growth is a heat-rejection problem expressed in litres.
3. Water needs energy: the desalination side
Now run the coupling the other way. Where freshwater is scarce, water is made with electricity. Modern seawater reverse osmosis typically consumes on the order of 2.5–4 kWh per cubic metre — against a thermodynamic floor near 1 kWh/m³ that no membrane can beat — and energy is the single largest line in a desalination plant’s operating cost (Elimelech & Phillip, Science, 2011). Energy intensity rises with salinity, so the hot, high-salinity waters of the Gulf sit at the demanding end of that range, before the additional energy of pumping water hundreds of kilometres inland and a thousand metres uphill to reach cities.
This is not a niche case. Global desalination already runs to roughly 20,000 plants producing on the order of 40 billion cubic metres a year, projected to grow by roughly a third by 2030 (International Desalination Association / IWA), and a handful of Gulf states operate the largest capacity on earth — for them it is not an industrial choice but an existential system. In these economies, “water security” and “energy security” are the same sentence. That is exactly why a leading Gulf utility has set a target to produce all of its desalinated water from clean energy and waste heat by 2030 (DEWA): they have understood, from the water side, what the data-centre operators are learning from the compute side — that the leverage is at the heat interface.
4. The same problem from two ends: the waste-heat interface
Put the two sides together and they are the same problem viewed from opposite ends. A data centre has an abundance of low-grade heat it is paying (in water and energy) to throw away. A desalination or district system has a demand for exactly that kind of energy. The tightest, most under-exploited point in the whole coupled system is heat rejection and recovery — the interface where compute’s waste becomes water’s input. Intervening there is unusual because it pays on all three axes at once: it cuts the water a data centre evaporates, reduces the energy a water system burns, and lowers the combined carbon of both. (My own work at this interface — a patent-pending waste-heat recovery method — is one illustration of the intervention point, not the point itself; the point is that the coupling has a lever, and almost no one is holding it because it falls between three mandates.)
5. Why this is a governance and risk problem, not a technical one
The engineering to cool a server or desalinate seawater is mature. What is missing is not technology but coordination under uncertainty — which is a governance discipline. Three observations follow.
First, siting and cascading failure. AI infrastructure is concentrating in water-stressed regions chosen for cheap land and power, not water. That couples digital, energy and water risk in the same geography, so a drought, a heat wave or a grid event no longer hits one system — it propagates. An energy failure becomes a water failure becomes a public-health failure. This is the signature of a coupled system, and it is invisible to any single regulator.
Second, the macro layer is now watching. Water scarcity has moved from an environmental footnote to a macroeconomic variable. A December 2025 study by the Bank for International Settlements and the World Bank, covering 169 countries from 1990 to 2020, found that a one-standard-deviation increase in water scarcity is associated with roughly 0.12–0.16% lower GDP growth, 0.39–0.42% lower investment growth, and 2.9–3.5% higher inflation (BIS Working Paper 1314 / World Bank Policy Research Working Paper 11261). When the BIS models water as a driver of growth and inflation, the coupling has stopped being a sustainability concern and become a question of economic stability.
Third, the disclosure gap. For years, most data-centre operators did not even meter their water intake; desalination’s true energy-and-carbon cost is often buried in a water tariff. You cannot govern what no one measures across sectoral lines. The first governance act is simply to make the coupled system legible.
6. Implications for decision-makers
The recommendation is not a new technology; it is a change of unit of analysis — from three sectors to one coupled system — applied to the decisions that are being made now:
- Site and permit data centres, power and water as one decision, with an explicit coupled water-energy-heat budget, not three separate approvals.
- Procure for heat recovery by default in co-located compute and water/thermal assets, so waste heat is treated as an asset, not an emission.
- Disclose across the boundary: require water-usage effectiveness for compute and energy-and-carbon intensity for water, in comparable terms.
- Stress-test the coupling, not each system alone: model the drought-plus-heatwave-plus-load event that hits all three at once.
- Assign ownership. Today the interface falls between an energy mandate, a water mandate and a digital mandate, so no one governs it. Someone must.
This is the logic behind PSR³, a framework that treats critical infrastructure as one coupled system for resilience and governance rather than a stack of independent pillars. Its claim is narrow and, I think, defensible: in the age of AI, the resilience of energy, water and compute is a single problem, and it will be won or lost at the seams between them.
Closing
The AI build-out is usually told as a story about chips and power. It is at least as much a story about water and heat — and all four are the same story, because they are thermodynamically one system. The organisations that will build durable advantage are the ones that stop optimising the pillars and start governing the coupling. That is the whole thesis, and it is where the next decade of critical-infrastructure value will be decided.
Sources
- International Energy Agency (2025), Energy and AI (special report, April 2025). Global data-centre electricity ≈415 TWh (2024) rising to ≈945 TWh by 2030; a 100 MW data centre consumes ≈2 million litres/day for cooling (≈6,500 households).
- Lawrence Berkeley National Laboratory (2024), 2024 United States Data Center Energy Usage Report (for the U.S. Department of Energy). U.S. data centres consumed ≈17.4 billion gallons directly (cooling) in 2023, plus ≈211 billion gallons indirectly (electricity generation) — indirect ≈12× direct; direct projected to double–quadruple by 2028.
- P. Li, J. Yang, M. A. Islam & S. Ren, “Making AI Less Thirsty” — arXiv:2304.03271 (2023); Communications of the ACM 68(7):54–61 (2025). Global AI water withdrawal projected at 4.2–6.6 billion m³/year by 2027.
- Microsoft (2024), Efficiency and Sustainability at Microsoft Datacenters. Closed-loop / near-zero-evaporation cooling saves ≈125 million litres per datacentre per year.
- M. Elimelech & W. A. Phillip (2011), “The Future of Seawater Desalination,” Science 333(6043):712–717. Energy is the dominant desalination cost; SWRO practical energy (≈2.5–4 kWh/m³) approaches a thermodynamic minimum near 1 kWh/m³.
- International Desalination Association / IWA — global installed desalination capacity (≈20,000 plants, ≈40 billion m³/year).
- Dubai Electricity & Water Authority (DEWA) — public target: 100% of desalinated water from clean energy and waste heat by 2030.
- Bank for International Settlements & World Bank (December 2025), “The Economics of Water Scarcity” — BIS Working Paper No. 1314 / World Bank Policy Research Working Paper 11261.
- World Resources Institute (Aqueduct) — overlap of AI-facility siting with water-stressed basins.