The Hidden Hydrodynamics of AI: How Much Potable Water is Consumed by a Data Center in a Single Day of Training and Operation?

Artificial Intelligence (AI) is often celebrated as a “clean” technology – a software product, ethereal, residing in the cloud. However, this perception ignores the dense physical infrastructure required to power everything from a simple chatbot to the training of large language models (LLMs) like GPT-4.

The AI and Water Paradox

The true cost of AI is not just in $kWh$ of energy; it resides, critically, in water. We are talking about a scarce resource used on two fronts: the direct cooling of data centers and, even more significantly, the virtual water embedded in the generation of all the electrical energy these centers consume.

In this post, we will unravel the hidden water cycles behind AI, presenting engineering metrics and the urgency of a more sustainable approach in the technology sector.

Direct Water Consumption: The Heart of the Data Center

The data center (DC) is the physical home of AI, and its greatest engineering challenge is managing heat. Each server rack can dissipate the heat equivalent of several residential heaters, requiring a robust cooling system where water is the preferred thermal agent.

The Water Anatomy of a Data Center

In modern DCs, the most common cooling method is evaporative cooling, which utilizes the famous Cooling Towers.

The Cooling Tower Process:

1. Hot water (heated by the servers) is pumped to the tower.

   
   

2. There, a small portion of the water evaporates into the atmosphere.

   
   

3. Evaporation is a highly efficient cooling process, as the change of state (liquid to gas) removes a large amount of latent heat from the system.

   
   

4. The problem: The water that evaporates needs to be replaced (makeup water). Additionally, another portion is drained (blowdown water) to prevent the buildup of dissolved solids (minerals) that cause corrosion and scaling.

The Engineering Metric: WUE

To quantify water efficiency, engineers use the Water Usage Effectiveness (WUE) metric.

WUE = Annual Volume of Total Water Used in the DC (Liters)

Annual IT Energy (kWh)

WUE is measured in Liters per Kilowatt-hour (L/kWh) or Gallons per Kilowatt-hour (gal/kWh).

A DC in a warm climate, which relies heavily on evaporative cooling, may have a much higher (worse) WUE than a DC in a cold climate that utilizes Free Cooling for most of the year.

Technical Detail: Makeup water, in most cases, must be of high quality (potable or highly treated) to prevent damage to expensive cooling systems and ensure maximum heat transfer efficiency.

Indirect Water Consumption: The Virtual Water of Energy

Direct consumption is only the tip of the iceberg. AI's largest water footprint lies in the virtual water embedded in the electrical energy generation consumed by the data center.

The Water Intensity of Electricity

The way energy is generated dictates its water cost:

• Thermal Power Plants (Fossil/Gas): Use large volumes of water to cool the condenser during the steam cycle. Water intensity can vary widely but is often high.

• Nuclear: Require intensive and constant use of water to cool the reactor and condenser.

• Hydroelectric: Although water is not "consumed" in the traditional sense, the evaporation of water from large reservoirs is allocated to energy consumption, which can represent significant losses in dry regions.

• Renewables (Wind/Solar Photovoltaic): Have the lowest water intensity, consuming water mainly for panel cleaning and manufacturing.

A DC operating on power from a thermal power plant emits a much larger indirect water footprint than a DC powered by wind energy in the same proportion.

The Complete Cycle: From AI Response to Water Source

To understand consumption on a human scale, we need to apply these metrics to the daily use of a Large Language Model (LLM).

Water Consumed Per Response (Query)

Researchers have attempted to quantify the water consumption per interaction with chatbots. Although the exact data varies greatly (depending on the DC's location and efficiency), studies indicate a surprising consumption:

The Shocking Factor: It is estimated that a conversation of 20 to 50 questions and answers on an advanced chatbot can consume the equivalent of approximately 0.5 Liters of potable water.

The Engineering Rationale:

This estimate is derived from the sum of direct and indirect water consumption:

Water per Query ≈ Energy per Query x (Direct WUE + Indirect WUE)

Number of Queries/Day

Model training is even more impactful. Reports suggest that training a large model, such as GPT-3, in an evaporative DC, may have required hundreds of thousands of liters of water. This water is consumed in a single day, not for final use, but for the training that enables the model to function.

The Location Factor: An Engineering Challenge

Data center engineering is being forced to consider water scarcity. The decision to build a DC in a dry, hot climate (dependent on evaporative cooling) or in a cold, humid region (favorable to free cooling) is now an ethical and sustainability imperative.

Innovation is focused on seeking alternatives such as:

• Water Reuse: Utilizing gray water or treated effluent.

• Liquid Immersion Cooling: Where components are submerged in dielectric fluids that do not conduct electricity. This method is significantly more water and energy efficient, as it eliminates the need for large air-based evaporative cooling systems.

  
  

Engineering and Water Sustainability

AI is not a technology free of environmental costs. The water consumption of a data center in a single day, whether for direct cooling or the virtual water of energy, is massive and unsustainable in the long term in water-stressed regions.

The challenge falls to Engineering: to develop closed-loop cooling systems and push for a data center energy matrix that is 100% based on low water intensity renewable sources.

When we look at the screen and interact with an AI, we must be aware of the glass of water we are, invisibly, consuming.

Want to see more content like this? Access our main Blog page!

#AI #ArtificialIntelligence #DataCenters #WaterConsumption #TechSustainability #WUE #WaterUsageEffectiveness #AIsWaterFootprint #DataCenterCooling #VirtualWater #LLMs #LargeLanguageModels #GPT3 #GPT4 #EnvironmentalEngineering #WaterEfficiency #CoolingSystems #LiquidImmersion #SustainableDataCenter #AIsWaterCost #DataCenterWaterUse #WUEOptimization #TechnologyAndWater #E_S_A #SustainableAutomatedElectric