How Much Water Does AI Use?

Artificial intelligence runs on powerful computers, but those computers need more than electricity. They also generate heat, and data centers often use water to keep their servers operating safely.

This means that every AI system has a water footprint. However, there is no single number that explains exactly how much water one AI prompt uses. The amount depends on the AI model, the length of the request, the data center’s cooling system, the local climate and the source of electricity.

AI systems run inside large data centers that require electricity and cooling.

Why Does AI Need Water?

AI models run on powerful processors, including graphics processing units and other specialized computer chips. These chips consume electricity and produce heat during both training and everyday use.

Data centers must remove this heat to prevent equipment from overheating. Some facilities use air cooling, while others use chilled-water systems, cooling towers or liquid-cooling technology.

In evaporative cooling systems, some water changes into vapor and is released into the atmosphere. This water is considered consumed because it does not immediately return to the local water supply.

AI can also have an indirect water footprint. Power plants may use water to generate electricity, especially facilities that rely on steam or cooling systems.

The total water footprint of AI can include:

• Water used directly to cool computer servers.
• Water used to generate the electricity that powers data centers.
• Water required to manufacture chips, servers and other equipment.

Water may be used to remove heat from powerful computer equipment.

How Much Water Does Training an AI Model Use?

Training a large AI model can require substantial computing power over many days or weeks. During this process, thousands of computer chips may operate continuously.

One widely cited research estimate found that training GPT-3 in Microsoft’s U.S. data centers could directly evaporate approximately 700,000 liters of clean freshwater, or about 185,000 gallons, under the conditions examined by the researchers.

This figure is an estimate, not a universal measurement. Different AI models use different amounts of computing power. A larger model may require more electricity and cooling, while a smaller and more efficient model may require considerably less.

The location of the data center also makes a difference. A facility in a hot climate may need more cooling than one in a cooler region. The type of cooling system and the source of electricity also affect the final estimate.

Training an AI model can require powerful computers to operate for long periods.

How Much Water Does One AI Question Use?

An individual AI request generally uses much less water than training an entire AI model. However, it is difficult to calculate the exact amount used by one question.

Some estimates suggest that a conversation involving several questions and answers could be associated with hundreds of milliliters of water under particular operating conditions.

This does not mean that every AI conversation uses exactly one bottle of water. A short request processed in an efficient data center may have a much smaller footprint. A long request that produces a detailed response, image or video may require more computing power and cooling.

The most accurate conclusion is that water use varies from one AI request to another.

Every AI request requires computing power, although the amount varies by task.

Why Do AI Water-Use Estimates Vary?

Several factors affect how much water an AI system uses.

1. The Size of the AI Model

Large AI models usually require more computing resources. A request sent to a small, specialized model may use less energy than the same request sent to a large general-purpose model.

2. The Length of the Request

A short text response generally requires less computation than a long answer, a coding task, an image or a video.

3. The Cooling System

Some data centers rely heavily on evaporative cooling. Others use air cooling, closed-loop systems or liquid cooling. Newer cooling technologies may reduce or even eliminate the need for water evaporation during normal operations.

4. The Local Climate

A data center in a hot or dry region may require more cooling than one in a cooler area. Water use can also have a greater local impact in a community affected by drought.

5. The Source of Electricity

Electricity generation can add an indirect water footprint. The amount depends on how the electricity is produced and how the power plant is cooled.

AI water use depends on the model, task, location, cooling system and energy source.

How Much Water Do AI Companies Use?

Technology companies often report water use for their overall data-center operations instead of separating water used specifically by AI products.

As AI demand grows, companies are building more data centers and expanding their computing infrastructure. These facilities may use large amounts of water for cooling and other operations.

However, company-wide figures usually cover more than AI. They may also include cloud computing, online services, data storage, websites and other digital products.

For this reason, AI’s water impact is best understood at the infrastructure level rather than by focusing only on one prompt.

Companies are exploring more efficient cooling systems and alternative water sources.

Is AI’s Water Use a Serious Environmental Problem?

AI’s water use is not equally harmful everywhere. The impact depends on where and how the data center operates.

A data center using recycled water in a region with abundant supplies may create less pressure than one using freshwater in a drought-prone community.

The main concern is the rapid growth of AI. Billions of requests may be processed by AI systems, and more data centers are being built to meet demand.

Technology companies are working on ways to reduce water consumption, including:

• Using air cooling or closed-loop liquid cooling.
• Building data centers in areas with lower water stress.
• Using recycled or non-drinking water.
• Improving computer-chip efficiency.
• Designing smaller and more efficient AI models.
• Publishing clearer information about water use.
• Replenishing water in the communities where companies operate.

What Can Individual Users Do?

Individual users are not the main source of AI’s water footprint. The biggest decisions are made by technology companies, data-center operators, chip manufacturers and policymakers.

Still, users can reduce unnecessary computing by:

• Writing clear prompts to avoid repeated attempts.
• Using smaller tools when they are sufficient.
• Avoiding unnecessary image and video generation.
• Reusing useful answers instead of asking the same question repeatedly.
• Choosing services that publish environmental information.
• Supporting companies that invest in efficient technology.

Users can reduce unnecessary computing by making their AI requests clear and purposeful.

Conclusion

AI does use water, mainly because data centers must remove heat from powerful computers. Electricity generation and the manufacturing of computer equipment can also require water.

Training a large AI model may consume hundreds of thousands of liters of water under certain conditions. Individual AI requests use much less, but the exact amount varies depending on the model, request length, cooling system, location and energy source.

The most important issue is not whether one prompt uses a few drops or several milliliters. The bigger concern is the combined effect of billions of requests, expanding data centers and local water shortages.

As AI continues to grow, companies will need to make water use more transparent and invest in cooling technologies that reduce pressure on freshwater resources.

Sources

• Research on the water footprint of AI
• Microsoft data-center cooling information
• Google environmental reporting