#also sorry i didn't include sources. last time i cited two hours of research on ''ai'' ppl universally ignored it to soapbox misinfo at me

[opha]

commercially available image generators consume an average 3 watt-hours per generation. running the AC for an hour is a little over 1000. we usually measure power in kilowatt-hours, btw, because watt-hours are so negligible in scope that you can pedal a bike for an hour and have enough power for 33 image generations, or running a single 25 watt light bulb for 4 hours.

as for water, about 500ml for anywhere between 5 and 50 generations (depends on location and season). all of that comes from data centers, which often (but not always) reuse the water and about a third of which are using non-potable wastewater for cooling in the first place. you know, the stuff you personally use up to the tune of 30-50 gallons per day by flushing the toilet, showering, running the washing machine, etc.

training large language models does consume much more energy, but it's still pretty small fish compared to any other tech industry, and that's not something that is constantly happening.

please don't be fooled by contextless figures and inaccurate analogies slung by obvious clickbait articles like "you'll be astonished how much power it takes to generate a single image!" and other people who have a vested interest in fearmongering misinformation. there are plenty of real problems in this sector, like the use of labor exploitation in the global south in fine-tuning LLMs.

See also, "We're in a drought; conserve water!" Meanwhile, bottled water companies and golf courses for rich folk empty the aquifers.