Skip to main content

Overestimating AI's water footprint


Last week, a journalist requested that I comment on a new preprint article discussing the water footprint of AI. Water is a crucial subject, as the discussion about the environmental impact of IT has mainly centered around energy and carbon emissions, a limitation I addressed in a 2021 article.

I agree with the preprint’s conclusions that data center water consumption is often an overlooked aspect, with data center operators traditionally treating it as a trade secret. We need to better understand the full environmental footprint so we can model whether strategies like demand response or load shifting are worth it (maybe not). Although major companies like Microsoft and Google are becoming more transparent about their water usage, comprehensive data on the water footprint of large data centers is still lacking

However, the study’s reported figures likely overestimate actual consumption. They describe an average on-site Water Usage Effectiveness (WUE) of 3.8-5.2 L/kWh based on a WSJ article from 2015 referencing a presentation from 2009 (the lower number) and an online post in Chinese (the higher number). Their model utilizes generic commercial cooling tower specifications and weather data. While the actual figures are not provided (a footnote, strangely not in the methodology section, explains a range of “0.5L/kWh and 5L/kWh depending on weather conditions”), the paper’s August 2022 on-site WUE graphs indicate ranges of ~3.8-6.2 L/kWh, depending on location.

Image of graphs
Hourly carbon efficiency and on-site WUE for the first week of August 2022, from Figure 5 in “Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models” (source).

This is an interesting approach because data center operators do not publish granular data, so they have tried to infer it. However, the use of generic equipment likely misses the efficiencies that exist in hyperscale data centers.

Research from 2016 suggests the average US data center WUE is 1.8 L/kWh, while Facebook (Meta), an industry leader in data center water footprint, reports 0.26 L/kWh (pg39). Microsoft’s WUE ranges from 0.1 L/kWh (EMEA) to 1.65 L/kWh (Asia Pacific), with the Americas sitting in the middle (0.55 L/kWh). Google’s efficiency is likely similar, making the graphed values excessive.

Graph of Facebook WUE
Annual data center water usage effectiveness, from Meta’s Sustainability report 2021. The axis appears mislabeled and should be L/kWh of IT power.

It is also crucial to contextualize data center water usage in relation to other industries. Billions of liters may seem like a lot, but understanding location-specific water stress is vital. While the paper mentions this, it does not thoroughly explore the topic alongside its results. In comparison to other industries, data centers consume relatively little water. All water counts, but how much it counts has to be considered relative to the local resources. Unlike carbon, zero water is not necessarily the goal.

Being a preprint, the article may undergo significant changes before publication, if it is published at all1. However, this is another example of selecting the high end of a range of possible input values to suit a particular narrative. We’ve seen similar scare tactics used when discussing “digital sobriety” in relation to extreme projections for future IT energy usage.

Perhaps a revised version could compare the estimates from the commercial cooling specs alongside the industry average and reported values from Facebook. That would be an interesting illustration. The methodology also needs clarification - different numbers are cited at different points, the ranges are wide, and the graphs in the appendix aren’t helpful in understanding the actual values used in the calculations.

Researchers must exercise caution when selecting model inputs, avoiding the temptation to adopt the highest possible value. They should show the range. Overestimates only undermine the credibility of vital environmental research topics.

  1. I also find it interesting how popular it is to publicize preprints before they have been peer reviewed. Academia is painfully slow, so I understand the desire to get things out, but there is a reason for independent review. ↩︎