The Water Math

Abstract

The public conversation about data center water use in India usually assumes the building is cooled by evaporating water in cooling towers. HyperNext campuses are not cooled that way. Every HyperNext data center is liquid-cooled. The heat is carried off the chips by a sealed, recirculating glycol coolant loop, not by water, and that heat is rejected through dry coolers rather than evaporative towers. Because nothing is evaporated away, the cooling system consumes effectively no freshwater. The limited water a campus uses for ordinary building services is treated on site in our own sewage treatment and water treatment plants and put straight back into use, so the same water is used again and again and net freshwater draw stays very low. A conventional evaporative-cooled facility at the same IT load would consume tens of millions of cubic metres of water a year. The HyperNext design avoids almost all of it. This paper sets out the water math behind that design and argues for proportionality in the debate. Data centers should be measured, reported, and held to standards, and those standards should reflect how the facility is actually cooled.

Contents

1. 011. Why we have to write this paper
2. 022. What a hyperscale AI data center actually consumes
3. 033. Comparable urban water uses
4. 044. Other industrial water users
5. 055. The economic productivity of water
6. 066. Liquid cooling and dry heat rejection
7. 077. What the conversation should be about
8. 088. Headlines and recommendations
9. 099. References and sources

Key findings

• Every HyperNext data center is liquid-cooled. Heat is removed by a sealed glycol coolant loop and rejected through dry coolers, so there are no cooling towers and no evaporative water loss. The cooling system consumes effectively no freshwater.
• Water used for general building services is collected, treated on site in HyperNext's own sewage treatment and water treatment plants, and reused. The same water is recycled repeatedly, so net freshwater draw is a small fraction of a conventional facility.
• Because the glycol loop is sealed and heat is rejected by dry coolers, there is no evaporation, so the cooling water-usage effectiveness (WUE) is effectively zero, against 1.5 to 2.5 litres per kWh for a conventional evaporative-cooled facility at the same load.
• A conventional evaporative-cooled facility at a 1.2 GW IT load would consume on the order of 45 to 55 million cubic metres of water a year. The HyperNext glycol-cooled, dry-rejection design avoids almost all of that.
• Because HyperNext runs on Khavda captive solar and wind, it also avoids the thermal-generation water that grid power would have required, on the order of 24 to 38 million cubic metres a year for an equivalent coal plant. The water saving sits on both sides, cooling and generation.