AI’s Hidden Water Footprint

Every time you use an AI chatbot like ChatGPT, it consumes water. Using ChatGPT’s latest model, GPT-4:

• Having it generate a 100-word email consumes about 500ml of water (17 oz).¹
• 10 to 50 queries consume about 2 litres of water (½ gallon).²

Additionally, training a model like GPT-3 in Microsoft’s US data centres can consume 5.4 million litres of water.² Given that the latest model GPT-4 is larger and more advanced, it’s likely that its numbers are even higher.

The global AI demand in 2027 is projected to require 4.2 and 6.6. billion cubic meters of water withdrawal, surpassing the annual water use of 4–6 countries the size of Denmark, or about half of the United Kingdom’s total annual withdrawals.²

This hidden water cost comes from the infrastructure needed to power and cool AI systems and the manufacturing of chips and servers. What’s more, the water used often has to be drinking quality because impurities can damage the servers.²

How AI Uses Water

Understanding AI’s water impact requires distinguishing between water withdrawal and water consumption.

Water withdrawal: Freshwater taken from the ground or surface water sources and used temporarily or permanently for industrial purposes.

Water consumption: The amount of water evaporated, transpired, or otherwise removed from the immediate water environment.

AI’s water footprint spans three main areas:

On-Site Water for Data Centre Cooling (Scope 1)

• Data centres require cooling systems that often rely on fresh water to manage server heat:
  • Cooling towers evaporate water to dissipate heat, with about 80% of the withdrawn water consumed.²
  • Air cooling with evaporation uses water to cool air in hot or dry climates, consuming roughly 70% of the water withdrawn.²
  • Dry cooling avoids water use but increases energy consumption, potentially shifting the water burden to electricity generation.

Off-Site Water for Electricity Generation (Scope 2)

• Data centers are accountable for off-site water usage associated with electricity consumption. In many countries, thermoelectric power is among the top sectors in terms of water withdrawal and consumption.
• In the US, generating electricity withdraws 43.8 liters and consumes 3.1 liters per kWh on average.²

Supply Chain Water for Manufacturing (Scope 3)

• Manufacturing AI chips and servers uses huge amounts of often ultrapure water, with low recycling rates and high potential for toxic waste.
• For example, Apple reports that 99% of its water footprint comes from its supply chain.²

Key Challenges

Lack of reporting: There is no requirement for data centres to report how much water they are consuming.^3,4

“Water positive” initiatives: Companies such as Google and Microsoft claim they will be “water positive” by 2030, pledging to return more water than they consume.^5,6 However, there is little accountability or evidence to support these promises, and water replenishment often happens in different locations than where it was extracted, leaving the plight of areas experiencing water stress unresolved.²

Carbon-water tradeoffs: Efforts to reduce carbon emissions are often in conflict with reducing water use.⁷ For instance, solar energy in hot periods can increase water needs.

For more insights, check out our Takeaways page.

To print your own postcards for distribution, click here.

References