AI and water: how much drinking water does ChatGPT consume?

Every interaction with an AI chatbot has an ecological impact, particularly due to water consumption. Data centers, the facilities where these AI systems run, often cool their servers with drinking water. This cooling system leads to evaporation, so the total water consumption per question to a chatbot, while limited, can add up to huge amounts when we consider the total number of questions on a daily basis.

The dynamics of cooling in data centers are crucial. Servers generate heat and require cold water to remain at a safe temperature. After this water is used, the heated water flows to a cooling tower, where a portion evaporates and is lost. The challenges of water consumption are particularly evident in the Netherlands, where the drinking water network is utilized for cooling on hot days.

Amazon and OpenAI have presented figures on water consumption per question. For instance, a question to ChatGPT is estimated to cost about 0.32 milliliters of water, but critics point to the obscured calculations, as only the consumption-explosive water in the data center is taken into account. Recent research indicates that when including the water needed for energy production, water consumption can rise to 16.9 milliliters per answer from GPT-3, with the majority of the water consumption being indirect.

Dutch data centers, such as the new Equinix complex in Amsterdam, are responsible for significant amounts of drinking water. This specific data center is allowed to use up to 1.1 billion liters of drinking water annually. This is not unique; 79 percent of the colocation data centers in the Netherlands rely on drinking water during hot days. During periods of high temperature, the pressure on drinking water supplies continues to rise, calling for innovative and water-saving solutions such as regenerative cooling systems.

Steps are planned in the future to improve efficiency regarding water consumption in data centers. For example, Microsoft is working on new data centers that do not require water for cooling through direct chip cooling, an innovative system that experiences minimal evaporation. However, the trade-off between energy and water consumption presents complexity for data center operators.

Read the full article from AI Insider.