How Much Water Does Your AI Query Use?

When we talk about the environmental impact of AI, CO₂ emissions are usually cited. But there's a lesser-known, more invisible cost: water consumption. Every query sent to a large language model requires energy for computation and, above all, for server cooling. Water is the silent engine behind every response we receive.

Data center cooling: the real water drain

Data centers hosting AI models are enormous industrial complexes that generate heat constantly. Evaporative cooling systems, among the most widespread in the world, consume hundreds of millions of liters of water per day in the United States alone. During drought periods, this demand comes into direct conflict with the needs of local communities and agriculture.

According to recent studies, a single chat session with an AI assistant can require up to 2 liters of water to process its requests. Multiplied by billions of daily interactions, the result is a global water consumption that grows exponentially, paralleling the spread of language models.

How much does a request actually consume?

The calculation isn't straightforward because it depends on many factors: the model's size, the provider's infrastructure, the cooling system adopted, and the geographic region where the data center operates. A lightweight model like GPT-3.5 requires fewer resources than a GPT-4 or Claude 3.5 Sonnet, but the difference in water terms remains in the order of magnitude between a few milliliters and several liters per interaction.

Here's an indicative estimate of water consumption for some common operations:

• A short text response: approximately 10-50 ml of water
• A complex generation with multi-step reasoning: 200-500 ml
• An hour of intensive use with large models: up to 2 liters

These numbers may seem negligible individually, but when scaled to a global level — with billions of daily queries — the impact becomes significant and urgent to consider.

Why the tech sector doesn't talk about it enough

Transparency around data center water consumption is still scarce. Many providers publish carbon emissions data but omit detailed information on water consumption, often considered a secondary figure in sustainability reports. This opacity makes it difficult for users and regulators to assess the true environmental cost of AI.

Some companies are starting to change course. Google, Microsoft, and OpenAI have announced initiatives to improve water reporting and invest in dry and recirculation cooling technologies. However, experts stress that current efforts are still far below what would be needed to offset the rapid growth in demand.

What we can do as users and organizations

Awareness is the first step. Reducing redundant queries, avoiding using AI for tasks that don't need it, and preferring smaller models when possible are concrete actions every user can adopt. Companies, meanwhile, should request transparent data from providers on water and energy consumption, including these indicators in procurement assessments.

Furthermore, supporting investments in more efficient cooling technologies and data centers designed for areas with water abundance can make a difference. The transition toward sustainable AI is not just a technological issue, but also one of conscious choices that start with each of us.

Practical takeaway

The next time you send a query to an AI, remember that behind that response there's a real water cost. Ask meaningful questions, avoid repetitions, and choose tools that publish transparent environmental data. Small, responsible digital habits can, collectively, make a big impact on the sustainability of our technological future.