Why is Everyone So Wrong About AI Water Use??

AI Summary

Hank Green addresses the widespread confusion and misinformation surrounding AI's water consumption, highlighting the stark contrast between Sam Altman's claim of 0.000085 gallons per ChatGPT query and Morgan Stanley's projection of 1 trillion liters annually by 2028. He explains that this discrepancy arises from the complexity of resource consumption analysis, particularly the difference between direct query-time water use and the full lifecycle footprint, which includes the massive water and electricity demands for training AI models. Green argues that Altman's figure is misleading because it omits the substantial water used in the continuous, energy-intensive training process, which can account for approximately 50% of an AI model's total resource use. He also delves into the water used by thermoelectric power plants that generate electricity for data centers, noting that while these plants withdraw vast amounts of water (40% of US freshwater withdrawals), much of it is returned, and it's not municipal drinking water. Green emphasizes that the *type* and *location* of water use are critical, distinguishing between municipal water, industrial water, and ultra-pure water needed for chip manufacturing. Ultimately, he reveals that AI's total global water use (260 billion gallons annually) is dwarfed by other industrial and agricultural uses, specifically US corn production (20 trillion gallons annually), 40% of which is for ethanol. Green concludes that while AI water use will be significant in some areas, the projected increase in *power demand* is a much larger environmental and economic concern, and he expresses skepticism about the long-term viability of the current AI buildout and the potential for an economic bubble.

Claims Extracted (12)