Understanding AI's Environmental Footprint

As artificial intelligence becomes integrated into our daily lives, from simple chatbots to complex research tools, its environmental impact is becoming a critical topic of discussion. Every query sent to a large language model (LLM) triggers a chain of events in a data center that consumes both water and electricity.

💧 The Hidden Water Cost

Data centers require immense amounts of water for cooling. High-performance GPUs used to train and run AI models generate significant heat. To maintain optimal operating temperatures, data centers use cooling towers that evaporate water.

Recent studies suggest that a single conversation with an AI (roughly 10-50 exchanges) can "drink" a 500ml bottle of water. This varies significantly based on:

• Model Size: Larger models like GPT-4 or Claude Opus require more compute per query.
• Location: Data centers in hotter climates or those using less efficient cooling tech consume more water.
• Time of Day: Cooling is more water-intensive during peak afternoon heat.

⚡ Energy Consumption

The electricity used by AI models is divided into two phases:

1. Training: Consumes massive amounts of power over weeks or months.
2. Inference (Usage): The power used to generate a response to your specific query.

While training often gets the headlines, the cumulative energy used for inference as billions of people use these tools daily is staggering.

🌍 How This Calculator Works

Our calculator uses industry-averaged data points from peer-reviewed research (such as Li et al., 2023 "Making AI Less Thirsty") and corporate sustainability reports.

Key Assumptions:

• Average Water per Query: Ranges from ~3ml for efficient models to ~200ml for extremely intensive ones.
• Average Energy per Query: Measured in Watt-hours (Wh), ranging from ~10Wh to ~100Wh.

Comparisons:

To make these numbers relatable, we compare your AI footprint to:

• Showers: A standard 10-minute shower uses about 91 liters of water.
• Smartphone Charges: A full charge of a modern smartphone uses about 17.5 Watt-hours.
• Boiling Kettles: Boiling 1 liter of water takes approximately 110 Watt-hours.

🌱 How to Reduce Your Impact

1. Be Precise: Use well-crafted prompts to get the right answer in fewer exchanges.
2. Choose Efficient Models: For simple tasks, use "mini" or "flash" models which are much more efficient.
3. Batch Queries: Instead of many small questions, group your inquiries.