How much water does AI use per prompt?

A 2026 report by the United Nations University Institute for Water, Environment and Health reveals AI’s growing water footprint. By 2030, AI data centers may use water equivalent to 1.3 billion people in Sub-Saharan Africa, with GPT-5 training alone consuming 1 billion litres—enough to meet the annual domestic needs of 81,000 people or fill 237 Olympic-sized pools. GPT-4 training required 600 million litres, underscoring the industry’s escalating demand. Water consumption occurs in three key ways: on-site server cooling, off-site electricity generation, and microchip manufacturing. AI chips generate extreme heat, requiring cooling towers, while electricity production for data centers also relies heavily on water. Manufacturing specialized GPUs and microchips demands additional water resources. ChatGPT, with 2.5 billion daily prompts, uses approximately 500 ml of water per 10-50 prompts, equivalent to a single water bottle. In contrast, Google’s Gemini consumes just 0.26 ml per prompt, or five drops. AI video generation consumes up to 4.1 litres per output, while image generation requires about 29 ml. The report highlights the environmental trade-offs of AI’s rapid growth, with data centers consuming millions of litres daily for cooling and energy production. As AI adoption accelerates, its water usage could strain global resources, particularly in regions with limited water availability.