How AI uses our drinking water - BBC World Service

Overview

A BBC World Service report examines how generative AI—particularly chatbots such as ChatGPT—can substantially increase both water and energy demand. This raises concerns about added pressure on local water sources and power grids.

Water use claims (OpenAI’s Sam Altman)

• Altman’s estimate: An average ChatGPT interaction uses about 1/15 of a teaspoon of water, based on around 1 billion messages per day.
• Expert scepticism: The BBC notes that experts are not confident in verifying the figure. They say the estimate may vary depending on:
  • which model sizes are used, and
  • what assumptions are made.

Why AI uses water

The report explains that water demand comes from several stages:

• Compute and training

  • Running prompts and training models requires powerful compute chips housed in large data centres.
  • These generate heat and require cooling.

• Cooling and liquid systems

  • Liquid cooling is increasingly common because data centres and related infrastructure are becoming more energy-intensive.
  • It often depends on clean (mostly drinking) water to reduce risks like clogs and corrosion.

• Typical cooling cycle

  • Coolant flows over chips, becomes heated, and then hot water is cooled via cooling towers using fans and evaporation.
  • Up to ~80% can evaporate, meaning that portion is effectively lost from the local water cycle.

Community and political pushback

Communities worried about water strain and rising electricity demand have protested data centre expansion in countries including:

• Spain
• India
• Chile
• Uruguay
• Parts of the United States

Indirect water use via electricity and manufacturing

AI impacts water use not only directly (cooling data centres), but also indirectly through the broader supply chain:

• Electricity generation
  • Many power plants (coal, gas, nuclear) rely on heat water into steam to drive turbines.
• Rising electricity demand
  • The International Energy Agency (IEA) projects that electricity demand for AI-optimised data centres could rise by 400% by 2030, reaching about 300 TWh—roughly equivalent to UK annual electricity use.
• Semiconductor manufacturing
  • Chip production and related supply chains also require water.
  • As a result, AI’s effects can be both direct and indirect across the hardware lifecycle.

Difficulty measuring true AI-driven impact

• Major tech companies often report water use at the data-centre level.
• However, they do not clearly break down how much water use is specifically attributable to AI.

Responses and efforts to reduce impact

• Water neutrality pledges
  • Companies including Google, Microsoft, and Meta have pledged water neutrality by 2030, though the report notes major gaps remain.
• Cooling trials and redesigns
  • Some initiatives focus on cooling systems designed to minimise or eliminate evaporation.
• Alternative locations
  • Proposals include relocating data centres offshore, including to colder regions (e.g., the Arctic).
  • More extreme ideas—such as placing data centres in space—are mentioned as very early-stage and with major hurdles.
• Overall perspective
  • The commentary stresses that GenAI is still young, and society will need to learn how to minimise water and energy impacts as adoption grows.

Presenters / Contributors

• Sam Altman (OpenAI CEO)
• BBC interview/coverage experts referenced in commentary (not individually named in the subtitles)
• NTT (representative mentioned; name not provided in the subtitles)