Consumes a bottle of water with every message

This water consumption includes not only the water used to cool servers in data centers, but also the water consumed by power plants that produce the electricity these systems need.

THE TWO INVISIBLE SIDES OF WATER CONSUMPTION

Two main sources of water lie behind every AI query. The first is water used directly for server cooling. Evaporative cooling systems are typically used for this purpose: massive fans spray water onto hot pipes, and the evaporative water carries away the heat, cooling the environment. However, this water is drawn directly from local water sources.

The second source is water consumed during electricity generation. Coal, natural gas, and nuclear power plants use large amounts of water. Hydroelectric and concentrated solar power plants can also cause water loss through evaporation. In contrast, wind turbines and traditional solar panels use almost no water.

LOCATION AND TIMING ARE CRITICAL

A data center's location and operating hours significantly impact water usage. For example, a facility in cool, humid Ireland can operate with virtually no water for months using outdoor air cooling, while in hot, dry Arizona, the same process can consume significant amounts of water in July. In winter, data center water needs can be halved compared to summer.

There are alternative systems being developed to address these issues. For example, Microsoft is working on a completely closed-loop, evaporation-free, liquid-cooled system. Another solution involves cooling servers directly by immersing them in special, electrically non-conductive liquids. However, these methods have not yet become widespread due to high costs and compatibility issues.

THE “INVISIBLE FOOTPRINT” OF ARTIFICIAL INTELLIGENCE

Different AI models consume different amounts of energy, and therefore water. For example, a medium-length query with GPT-5 uses approximately 19.3 watt-hours of energy, while GPT-4o uses only 1.75 watt-hours. This significantly varies the water consumption between them.

With a simple calculation:

• GPT-5 query → 19.3 × 2 ml/watt-hour = 39 ml water
• GPT-4o query → 1.75 × 2 ml/watt-hour = 3.5 ml of water

These numbers can be even lower in efficient systems. For example, Google's Gemini system uses only 0.26 milliliters of water for an average query. However, these calculations vary depending on query length for comparability.

HOW IS WATER CONSUMPTION MEASURED?

Three simple steps are all it takes to calculate your own AI footprint:

• Learn the model's energy usage.
• Determine how many ml of water is used for each watt-hour (1.3-2 ml is reasonable).
• Energy × Water factor = Total water consumption (ml)

This way, you can evaluate the environmental impact of your technology use more consciously.

WATER COMPARISON WITH DAILY EXPENDITURES

OpenAI reports processing approximately 2.5 billion AI queries per day. The water consumption of these queries varies significantly depending on the model used:

• All GPT-4o queries: 8.8 million liters per day
• All GPT-5 queries: 97.5 million liters per day
• All Google Gemini queries: approximately 650,000 liters

By comparison, lawn and garden watering in the US alone uses 34 billion liters of water daily. So, AI systems still consume much less water.

However, optimizing these systems and transparently disclosing their environmental impact is crucial. Site selection, recycling systems, and efficient equipment can make all the difference.