Artificial Intelligence Uses up to 50 ml of Water per Answer

In recent years, artificial intelligence technologies have rapidly entered various aspects of our lives. However, most people have little understanding of the infrastructure behind these systems. Every time we ask a question to an AI chatbot, the data centers that power these systems consume a substantial amount of water.

According to calculations by the International Energy Agency, an average 100-megawatt data center in the U.S. uses about 2 million liters of water per day. This water is primarily used for cooling servers. Cooling is typically achieved through evaporative cooling towers. Furthermore, the electricity used to power these data centers is often generated at thermal power plants, which themselves require large quantities of water for cooling, further increasing indirect water usage.

According to one analysis, each time we ask a question to an AI chatbot, a data center is essentially “forced to drink” water. More precisely, a single data center uses about 500 ml of water to generate 10 to 50 average-length GPT-3 responses. This means each response consumes approximately 10 to 50 ml of water. Image generation requires much more water than text, and video generation demands even more than images.

To prevent servers from overheating, data centers need cooling systems. One of the main methods is evaporative cooling, which consumes water. Additionally, data centers indirectly use water through their electricity consumption, since power plants also require cooling.

Some tech companies have implemented or proposed water-efficient cooling technologies. Experts say that there's often a trade-off: cooling methods that use more water typically consume less electricity, whereas water-saving methods tend to use more energy. It’s a complex balancing act.

Many companies are shifting to more water-efficient systems, but such solutions are not always economically or technically viable everywhere. Read 'Zamin' on Telegram!