In the US, average residential water delivery in 2015 was around 300 liters per person. Please do not share factoids about how AI is gobbling up all the water because it consumes a fraction of a liter. https://pubs.usgs.gov/of/2017/1131/ofr20171131.pdf
@Alon Also much (I believe almost all) of data center water use is non-consumptive use for cooling.
@BenRossTransit @Alon The "data center water use" factoids going around seem to actually be based on attributing the cooling water consumption of power plants to the data centers.
@BenRossTransit @wollman @Alon The electricity usage of AI might actually be significant, but the water thing I assumed was some sort of joke. That said, an energy comparison establishing equivalent EV distance driven for a typical training-amortized AI query would be useful, I think.
@Colinvparker @BenRossTransit @Alon The problem with any of these is how much waste you assume away. ICE cars are about 25% efficient, but the average thermal power station is twice that at best -- and you almost always are looking at marginal costs ("joules per inference") and not capital ("tCO2eq to produce one H100 and associated metal and plastic components including amortized shares of their production facilities").
@wollman @BenRossTransit @Alon That’s why I specifically don’t want to look at manufacturing or hardware lifecycle costs, or at ICE engines. Electric cars and AI data centers both draw power off the grid, I just want to compare total energy usage by EVs/miles driven with total energy usage by AI/number of queries, which gives me a potentially useful miles/query figure of merit.
@Alon they didn't say AI they said Al, short for alfalfa.