Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models

2023/04/06 by Pengfei Li, Jianyi Yang, Li, Pengfei +5 · 87 voices · 163 citations
Computer Science · Environmental Science · #Business #Carbon footprint #Computer science #Economics #Environmental economics #Environmental resource management #Environmental science #Flood Risk Assessment and Management #Footprint #Geography #Geology #Greenhouse gas #Hydrology and Watershed Management Studies #Law #Natural resource economics #Political science #Scrutiny #Water resources #Water scarcity #Water use #Water-Energy-Food Nexus Studies #cs.AI #cs.LG

paper · pdf · doi:10.48550/arxiv.2304.03271

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2023/04/06 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/05

Abstract

The growing carbon footprint of artificial intelligence (AI) has been undergoing public scrutiny. Nonetheless, the equally important water (withdrawal and consumption) footprint of AI has largely remained under the radar. For example, training the GPT-3 language model in Microsoft's state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret. More critically, the global AI demand is projected to account for 4.2-6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of 4-6 Denmark or half of the United Kingdom. This is concerning, as freshwater scarcity has become one of the most pressing challenges. To respond to the global water challenges, AI can, and also must, take social responsibility and lead by example by addressing its own water footprint. In this paper, we provide a principled methodology to estimate the water footprint of AI, and also discuss the unique spatial-temporal diversities of AI's runtime water efficiency. Finally, we highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly sustainable AI.

Cited by

• Enhancing SLMs for Sustainable Code Optimization in Radio-Astronomy
• Programmable and nonvolatile computing with composition tuning in thin film lithium niobate
• Small Bottle, Big Pipe: Quantifying and Addressing the Impact of Data Centers on Public Water Systems
• Misinformation by Omission: The Need for More Environmental Transparency in AI
• Transfeminist AI Governance
• Balancing Bits and Drops: Stress-Adjusted Water Management for Data Centers
• Spatio-Temporal Shifting to Reduce Carbon, Water, and Land-Use Footprints of Cloud Workloads
• From FLOPs to Footprints: The Resource Cost of Artificial Intelligence
• Toward Sustainable Generative AI: A Scoping Review of Carbon Footprint and Environmental Impacts Across Training and Inference Stages
• Smart but Costly? Benchmarking LLMs on Functional Accuracy and Energy Efficiency
• Data Centers Water Footprint: The Need for More Transparency
• Pensar con la mirada: cocreación humano-algorítmica y agencia distribuida en Visions of Destruction
• Green-LLM: Optimal Workload Allocation for Environmentally-Aware Distributed Inference
• The Cloud Next Door: Investigating the Environmental and Socioeconomic Strain of Datacenters on Local Communities
• Geography for AI sustainability and sustainability for GeoAI
• Life Cycle Assessment of Pre-training the Lucie 7B Open-Source Large Language Model on the Jean Zay Supercomputer
• A Survey on Parameter-Efficient Fine-Tuning for Foundation Models in Federated Learning
• Sustainable Metal-Organic Framework Water Harvesters in the Artificial Intelligence Era
• How Hyper-Datafication Impacts the Sustainability Costs in Frontier AI
• GPTFootprint: Increasing Consumer Awareness of the Environmental Impacts of LLMs
• Data center [wikipedia]

Discussions

• What's the deal with AI datacenters using water for cooling? [lemmy, 159 points, 83 comments]
• Most quotes in the press about the water use of AI are referring to this analysis, which is primarily referring to the water used for power generation, and then guessing how much power AI stuff uses. [bsky, 52 points, 1 comments]
• Here's the one that modelled chatgpt specifically arxiv.org/abs/2304.03271 [bsky, 44 points, 6 comments]
• “Global AI demand is projected to account for 4.2 – 6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of Denmark or half of the United Kingdom. [bsky, 36 points, 6 comments]
• Our new AI water paper is out! Updates include water withdrawal vs. consumption and scope definition. In 2027, global AI May withdraw 4-6 billion cubic meters of scope 1&2 water, equivalent to to 4-6 [bsky, 25 points, 2 comments]
• A modern server with a modern inference setup could probably serve Mistral NeMo 12B at ~1Wh for ~400 tokens. At 0,55l/kWh (cooling water only) (arxiv.org/pdf/2304.03271), that's 0,00055l for 400T. Be [bsky, 10 points, 1 comments]
• the paper does go into it in some detail to be fair arxiv.org/pdf/2304.03271 [bsky, 8 points, 1 comments]
• "An average user’s conversational exchange with ChatGPT basically amounts to dumping a large bottle of fresh water out on the ground... Microsoft’s water consumption has increased 30% y/y to keep AI s [bsky, 8 points, 0 comments]
• データセンターで蒸発することで失う水…サーバーの冷却を効率化する気化システム “米カリフォルニア大学リバーサイド校とテキサス大学アーリントン校の研究チームが、AIによる水資源消費について調査したレポート｢Making AI Less ‘Thirsty’｣” arxiv.org/pdf/2304.032... の中で “ChatGPT（GPT-3）を1人のユーザーが使うとして、25から50個の基本的 [bsky, 6 points, 3 comments]
• The original paper, "Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models" arxiv.org/pdf/2304.03271 [bsky, 6 points, 0 comments]
• retracting some earlier posts. it’s still true that training a model requires astonishing water consumption (MILLIONS of liters) but each individual query to a GPT model does indeed use a non-trivia [bsky, 5 points, 0 comments]
• "GPT training in U.S. data centers can evaporate 700k L of freshwater" [hn, 3 points, 2 comments]
• " More critically, the global AI demand is projected to account for 4.2 – 6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of 4 – 6 Denmark or [bsky, 3 points, 1 comments]
• A report on data center's extensive water usage: "More critically, the global AI demand is projected to account for 4.2 – 6.6 billion cubic meters of water withdrawal in 2027, which is more than the t [bsky, 3 points, 0 comments]
• The calculation I have seen is 0.004kWh per request for an old-school GPT-3.5 class model, for what it is worth. arxiv.org/pdf/2304.03271 [bsky, 3 points, 0 comments]
• The paper that this story is based on goes into details about how these cooling systems work, and includes references to where they're getting their data from. arxiv.org/pdf/2304.03271 [bsky, 3 points, 0 comments]
• If you’re already concerned about the energy demand implications of generative AI, just wait until you hear about its impact on water usage… arxiv.org/pdf/2304.03271 [bsky, 3 points, 1 comments]
• Une unité existe (le Water Usage Efficiency), mais son utilité reste assez limitée car elle décrit l'efficacité au niveau data center. Cet article détaille bien son fonctionnement et tente l'exercice [bsky, 3 points, 1 comments]
• Completely unrelatedly, just reading this scary article about how chatgpt-ing your lil emails is using a 500ml bottle of water at a time 💀 thanks a lot guys arxiv.org/abs/2304.03271 [bsky, 2 points, 0 comments]
• I think this is the paper at the root of all the memes arxiv.org/pdf/2304.03271 [bsky, 2 points, 1 comments]
• Here’s an interesting paper examining the under-reported consumption of fresh water by AI models (like Chat GPT.) it’s not just a carbon footprint. arxiv.org/pdf/2304.03271 [bsky, 2 points, 1 comments]
• Aviam que et sembla aquest paper arxiv.org/abs/2304.03271 [bsky, 2 points, 1 comments]
• So according to this study, each response generated by GPT-3 used ~0.004kWH of electricity / ~12.5mL of water. They estimate that 500mL of water is used up by 10-50 inferences. So why am I seeing nume [bsky, 2 points, 1 comments]
• There's also arxiv.org/abs/2304.03271 [bsky, 2 points, 0 comments]
• AI models are water guzzlers Training a model like GPT-3 in Microsoft's US data centres could directly consume 700,000 liters of clean freshwater. If the training occurred in Microsoft's Asian data [bsky, 2 points, 0 comments]
• "have you looked into literally any research on how much water it takes for chat-gpt to correct your while loop?" — @jennschiffer.com arxiv.org/abs/2304.03271 [bsky, 2 points, 1 comments]
• yeah! but the water dimension is also the one most easily coopted, because it's "just a fraction of a shower", without adressing things like necessity, frequency, etc. While the focus on the water di [bsky, 2 points, 1 comments]
• Making AI Less Thirsty: Uncovering and Addressing the Secret AI Water Footprint [hn, 2 points, 0 comments]
• [2/8] Les chercheurs évaluent la dépense énergétique moyenne d’une requête sur GPT-3 à 0,004 kWh, soit environ 16,9 mL d’eau consommée aux États-Unis. Il faudrait donc près de 30 requêtes pour atteind [bsky, 2 points, 1 comments]
• ? Around 80% of the water withdrawn is ultimately evaporated (consumed) for cooling. The remainder is recirculated until it is too contaminated & gets discharged. This water consumption might be immat [bsky, 2 points, 1 comments]
• This is the paper I see cited most often. A closed loop system transports heat from the computers to a cooling tower, then the heat is transferred either to the surrounding air to blow away or a water [bsky, 1 points, 1 comments]
• Intéressant : le journaliste a l'air aussi surpris que moi du chiffre avancés. J'ai trouvé l'étude en queston arxiv.org/pdf/2304.03271 Pour une requête de 1100 mots (800 in, 300 out), ils estiment [bsky, 1 points, 1 comments]
• Sobat Altra boleh banget baca paper lengkapnya di sini 👇 "Uncovering and Addressing the Secret Water Footprint of AI Models" arxiv.org/pdf/2304.03271 "Estimating the Carbon Footprint of BLOOM : A 17 [bsky, 1 points, 1 comments]
• "every viral post" is demonstrably untrue This was my introduction to this topic - it's peer-reviewed! arxiv.org/abs/2304.03271 All opposition to people seeking to confirm their biases is not just mor [bsky, 1 points, 1 comments]
• References: "Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models" (source for water estimates for MS) arxiv.org/pdf/2304.03271 Water use by energy source: visua [bsky, 1 points, 1 comments]
• arxiv.org/abs/2304.03271 [bsky, 1 points, 1 comments]
• yup, arxiv.org/pdf/2304.03271 I just found that different authors have different data for water intensity of electricity in the Netherlands or Denmark. 3x-6x difference is quite significant. Meanwhil [bsky, 1 points, 2 comments]
• [2304.03271] Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models https://arxiv.org/abs/2304.03271 #commons #uusitalous #p2p #neweconomy #regeneration #dework #c [bsky, 1 points, 0 comments]
• Untrue according to published estimates. 500ml of fresh water per 30-50 prompt session & more co2 you think. OpenAI, Google, and Microsoft are not transparent enough to give us the data needed to per [bsky, 1 points, 0 comments]
• Har sett uppskattningar på ~3–4 Wh/sida text. Det går nog att få ned med mindre modeller, MoE etc. Men det blir inte bra om »AI« skall byggas in i allting, samtidigt som det ofta är tveksam vad det ti [bsky, 1 points, 1 comments]
• This post references a screenshot, which references popsci articles, which reference a different popsci article, which references actual research. The actual research does not corroborate the claims f [bsky, 1 points, 0 comments]
• as a scientist, stem major, working with other scientists it’s not easy either :/ they can comprehend it but they won’t give it up for its efficiency but here is the groundbreaking research that unc [bsky, 1 points, 1 comments]
• Training LLMs and using them uses a huge amount of water for cooling. "Training GPT-3 in Microsoft’s state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater": arxi [bsky, 1 points, 1 comments]
• This study from 2023 lays it out. It is very real. arxiv.org/pdf/2304.03271 [bsky, 1 points, 0 comments]
• i don't get it if i go back to two of the references arxiv.org/pdf/2304.03271 www.city.ac.uk/news-and-eve... it is obvious that the meaning of "water consumption" is completely different if feel for [bsky, 1 points, 0 comments]
• 3/n the researcher interviewed by The Verge uses the estimate 0.0031 liters per Wh, but that's for all electricity in the US, not specific to data centers. This estimate is an order of magnitude more [bsky, 1 points, 1 comments]
• Not trying to stop anything. You just highlighted 2 reasons why I think I will be fine without it. 1. The fact you got from the 4 AIs are wrong. At least according to this paper: arxiv.org/abs/2304. [bsky, 1 points, 0 comments]
• Feel free to read the study: arxiv.org/pdf/2304.03271 BTW: it also sites using 1 in 10 WORINKG Americans, so the 10% is 16 million figure (October 2024, about 161.5 million people were employed in th [bsky, 1 points, 1 comments]
• The article links the research paper it’s based on at the end. It includes evaporated water in its definition of “consumed” water because it is no longer available in the immediate water environment. [bsky, 1 points, 1 comments]
• Uncovering and Addressing the Secret Water Footprint of AI Models [hn, 1 points, 0 comments]
• Secret Water Footprint of AI Models [hn, 1 points, 0 comments]
• Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models arxiv.org/abs/2304.03271 [bsky, 0 points, 0 comments]
• Here's the research paper: arxiv.org/abs/2304.03271 It claims 700,000 liters of water was used to train GPT3. It takes ~6800 liters to produce a pound of beef. So you're looking at around ~100lbs of b [bsky, 0 points, 2 comments]
• Cooling does use up water in vast quantities as a lot is evaporated and for the rest, the cycles are finite. AI in the US will need 2 billion litres in 2028 alone according to a paper from UC Riversid [bsky, 0 points, 1 comments]
• Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models arxiv.org/abs/2304.03271 [bsky, 0 points, 0 comments]
• Would you care to provide a citation for that as, as of last year, the information seems pretty clear otherwise: arxiv.org/abs/2304.03271 [bsky, 0 points, 1 comments]
• For anyone interested I've gone down a rabbit hole... The impact of AI on the water eco system... arxiv.org/pdf/2304.03271 [bsky, 0 points, 0 comments]
• @dtelder Svår och intressant fråga. Får nog hänvisa till den där studien som nämns i artikeln, eftersom den handlar mycket om just det: https://arxiv.org/pdf/2304.03271 [bsky, 0 points, 1 comments]
• The bottled water estimates are from a 2023 study from UC-Riverside. But the real numbers are likely much worse. They studied older, smaller AI models running in big tech owned data centers. Today A [bsky, 0 points, 1 comments]
• Making AI Less “Thirsty”: Uncovering & Addressing the Secret Water Footprint of AI Models "The growing carbon footprint of AI has been undergoing public scrutiny. Nonetheless, the equally important wa [bsky, 0 points, 0 comments]
• Les dades referents al 80-20 surten d'aquest estudi: arxiv.org/pdf/2304.03271 [bsky, 0 points, 1 comments]
• This is shaolei ren older one from when data was low arxiv.org/pdf/2304.03271 And it was all the way back in gpt 3 but he has more recentilt said that with moee updated data he think the newest models [bsky, 0 points, 0 comments]
• I can't find a good source for the 20 gallons for each use, but the figure for amortized water consumption including hardware manufacture, training, and cooling water for power plants as well as the d [bsky, 0 points, 1 comments]
• arxiv.org/abs/2304.03271 [bsky, 0 points, 0 comments]
• Approximate total annual freshwater consumption by 2027: 6.6km^3. [bsky, 0 points, 1 comments]
• arxiv.org/pdf/2304.03271 www.washingtonpost.com/technology/2... Here's more reading if anybody that isn't that guy wants to check it out. I'm resolving to be less combative on here. [bsky, 0 points, 0 comments]
• add the water footprint to the carbon footprint and we have AI accelerate our ongoing environmental catastrophe arxiv.org/pdf/2304.03271 [bsky, 0 points, 0 comments]
• OK I managed to find a different article with that claim that had sources. And guess what. We're back to " For example, training the GPT-3 language model in Microsoft’s state-of-the-art U.S. data cent [bsky, 0 points, 1 comments]
• "GPT training in U.S. data centers can evaporate 700k L of freshwater" Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models arxiv.org/abs/2304.03271 [bsky, 0 points, 0 comments]
• You're so right for this pro-environment AI image. By the way, Training GPT-3 in Microsoft’s U.S. data centers can directly evaporate 700,000 liters of clean fresh water. arxiv.org/pdf/2304.03271 [bsky, 0 points, 0 comments]
• Tomorrow may rain, so ... unfollow the sun arxiv.org/pdf/2304.03271 [bsky, 0 points, 0 comments]
• arxiv.org/pdf/2304.032... [bsky, 0 points, 0 comments]
• I found the original paper, and unless I'm missing something or doing my math wrong, they actually seem to say it's 0.425 oz per *page* of content, though idk what a full page necessarily means. arxi [bsky, 0 points, 2 comments]
• From glancing at this paper arxiv.org/pdf/2304.03271 it looks like it gets evaporated by the cooling, either at the data center or at the power plant if it’s a « boil some water » one. [bsky, 0 points, 1 comments]
• "Each 100-word AI prompt uses up roughly one 500ml bottle of water due to the cooling needs of datacenters, researchers have estimated." arxiv.org/pdf/2304.03271 [bsky, 0 points, 0 comments]
• Data centres use cold water to cool their computers — and they use immense amounts. One study projected global AI demand will withdraw between 4.2 and 6.6 billion cubic meters of water in 2027. arxiv. [bsky, 0 points, 1 comments]
• Focusing in on water usage for ChatGPT, the range I found was 462ml per query in IL, where I live, to 50x less www.washingtonpost.com/technology/2... arxiv.org/pdf/2304.03271 [bsky, 0 points, 1 comments]
• Li, Pengfei, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. 2023. “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models.” arXiv. doi.org/10.48550/arX... . [bsky, 0 points, 1 comments]
• Firstly, a lot of the seems to stem from this paper, which is requoted in a lot of places. arxiv.org/pdf/2304.03271 [bsky, 0 points, 1 comments]
• Lol but anyways here's the paper if you really can't be bothered to click a link: https:// arxiv.org/pdf/2304.03271 [bsky, 0 points, 1 comments]
• As for water use, training GPT 3 (which is 175 billion parameters, much costlier to train than better AND smaller models like LLAMA 3.1 8b) evaporated 185,000 gallons of water: arxiv.org/pdf/2304.0327 [bsky, 0 points, 1 comments]
• I'd love to know your sources, I can only find the opposite: arxiv.org/pdf/2304.03271 [bsky, 0 points, 2 comments]
• Fuente: arxiv.org/abs/2304.03271 [bsky, 0 points, 0 comments]
• Try reading 2.2 arxiv.org/pdf/2304.03271 [bsky, 0 points, 1 comments]
• Anyway, the paper - arxiv.org/pdf/2304.03271 [bsky, 0 points, 0 comments]
• Here’s one: the use of “thirsty” implies excessive water use. Pengfei Li, Jianyi Yang, Mohammad A. Islam & Shaolei Ren. 2023. Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footp [bsky, 0 points, 0 comments]
• And it’s not like there isn’t use for the excess heat. On the bright side, reading the paper suggests the easiest way to reduce water use is on the power generation side: less thermoelectric coal, ga [bsky, 0 points, 0 comments]

Related

• The Environmental Impact of AI: A Case Study of Water Consumption by Chat GPT
• The Hidden AI Race: Tracking Environmental Costs of Innovation
• Investigating awareness and behavioural intentions from AI water footprint digital posters in higher education settings
• Beyond AI as an environmental pharmakon: Principles for reopening the problem-space of machine learning's carbon footprint
• Recommendations for public action towards sustainable generative AI systems