Stop acting like AI uses a lot of water
Look, I get it. AI hype is extremely annoying, and increasingly inescapable. But that’s not a reason to take every anti-AI argument as uncritical truth.
Specifically, the thing I want to talk about here is the claim that AI is using a lot of water, and how you’ve probably been mislead by journalists who are not contextualizing what water use means
TL; DR: When you compare the water use estimates to the kinds of water consumption we use for food,
- The total estimate of global water use for AI in 2027 that’s commonly cited is equivalent to ~Dallas’s beef consumption today
- 3600 GPT-4 queries uses about the same amount of water as 1 quarter-pounder burger
How much water does AI use?
The source of most of the claims about AI water use seems to be a preprint titled “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models” by Pengfei Li et al [link]. This source seems credible so I will take its estimates seriously. Note that I am using scientific notation for all numbers because units of volume are annoying (1km³ is unintuitively equal to 1e9 m³)
Li et al cite an estimate of 2.2e9 m³ for “total annual water withdrawal” in 2022, of which 0.18e9 m³ “was ‘lost’ due to evaporation and hence considered ‘consumption.’” They go on to extrapolate based on estimates that the consumption could rise to 0.38e9–0.60e9 m³ by 2027. They also estimate that training GPT-3 took 5400m³ of water to train (they write this as 5.4 million liters) and each GPT-3 run takes between 1e-5 and 5e-5 m³ (they write this as 500mL for 10–50 responses).
What I believe is the same team also wrote in the Washington Post that a 100 word email using GPT-4 requires 500mL, or 5e-4 m³, about 10x more than GPT-3.
Important note: one common objection to this line of criticism is that datacenters do not consume water, it is instead recycled in a loop. However the paper that originates this criticism clearly acknowledges this and accounts for it, providing separate numbers for withdrawal and consumption of water. I will be using the consumption estimates from this point onward.
Is that a lot?
The thing about the previous section is that it contained a lot of decontextualized numbers. What does a billion cubic meters of freshwater look like? I could like, draw an image of this, but in my view this is the wrong way to look at the problem. Instead, let’s take a look at how much water we use for various tasks.
By far the largest user of water in modern society is agriculture (see my article from last year for more on that). Agriculture requires a lot of water for irrigation, and a lot of food that we (well, you, I’m vegetarian) eat is produced by an inefficient process in which a lot of food in the form of plant material is converted into not a lot of food in the form of meat. So, let’s compare to that!
Explaining water use to an American: ok so picture a burger
For the purposes of this writeup, I’m introducing two new units of water, the “burger” (water needed to make 0.25 lbs of beef) and the “year of american beef diet” (YABD). The burger can be computed as 1.8m³ (1847 gal/4) and the YABD can be computed as 470m³ (1847 gal * 67).
Rewriting the previous numbers, we find that the 2027 estimate is equivalent to 810k-1.27m YABD, or the beef-based water consumption of a city between the sizes of San Francisco and Dallas (city limits only). This is obviously not nothing, but keep in mind this is estimated global water consumption for an entire industry.
The GPT numbers are even more stark, GPT-3 took 12 YABD to train (yes, 12. with no zeros after it. GPT-3 took as much water to train as 12 Americans consume in beef per year), or 3000 burgers, about as much as McDonalds sells in 20 seconds (extremely rough calculation, based on this article and then dividing the amount by 2 to be sure). Each GPT-3 query consumes (using the high estimate) 27 microburgers, or in other words, one burger consumes as much water as 36000 GPT-3 queries. Plugging this in for GPT-4, we get that one burger consumes as much water as 3600 GPT-4 queries.
Interestingly, the Washington Post article makes these comparisons, but only for the full model training; and tries to sell 100 pounds of beef = GPT-3 training and 4439 pounds of rice = LLaMA training (no idea why two different units) as a lot. I have no idea why I am supposed to think this is a lot; the first of these is well under a typical individual person’s water consumption and this model only needs to be trained once.
Sidenote: water use is not the same as carbon emissions
One of the comparisons throughout Li et al is that of water use to carbon emissions. This is, in my view, a somewhat flawed analogy. Carbon emissions are global, as the atmosphere mixes together and global warming is, well, global. On the other hand, every watershed on the planet is pretty much a distinct water usage zone. You using more water in America isn’t making someone in Africa thirsty because the challenge is not that there isn’t enough water, it’s that there isn’t enough water in Africa. Shipping water over seas is just not a thing that’s really done, except as a scam. Ideally, water use can be aligned with places that have more than enough water; though of course this doesn’t always happen in practice (second plug for my western water crisis article).
Conclusion
There’s a lot of reasons to criticize the current hype bubble around AI. Water is not one of them.
