Key Points
- A single 100-word email generated by GPT-4 consumes roughly 519 milliliters of water, equivalent to a standard bottle, primarily due to evaporative cooling at data centers.
- Training a large language model like GPT-3 can require millions of liters of water, with regional data exposing that a conversation of 20-50 questions with an AI can evaporate half a liter.
- The water is mostly “consumed” (lost to evaporation) rather than circulated, pulling from local drinking supplies and stressing drought-prone areas like Arizona and Spain.
- Meta’s Llama 2 training in West Sacramento, for example, used over 10 million liters of water during a severe California drought, raising ethical concerns about resource allocation.
- The global AI data center boom is projected to withdraw up to 6.6 billion cubic meters of water annually by 2027, intensifying competition with agriculture and municipal needs.
Why It Matters
As AI becomes embedded in daily life, its hidden hydrological cost could trigger regulatory clashes and community pushback, forcing tech firms to choose between innovation and water security in a warming world.
Read More
Explore more coverage: At the Arctic Games, Canada and Greenland vs. Trump Feels Like Its Own Sport, Mob Burns Congo Ebola Center Amid Rare Strain Outbreak, Rare Rubens notebook sheet goes on display in artist’s home city of Antwerp, I tried Amazon’s Bee wearable and am both intrigued and slightly creeped out
