CoChat
Sign Up
Cited by 176Green IT and SustainabilityEnergy Efficiency in ComputingEnvironmental Impact and Sustainability

Making AI Less 'Thirsty'

Open access
Li, Pengfei, Yang, Jianyi, Islam, Mohammad A., Ren, Shaolei·2025·Communications of the ACM
DOI 10.1145/3724499
Open this paper in CoChat — free accountOpen PDF

Abstract

Uncovering and addressing the secret water footprint of AI models

AI snapshot

Written from the paper’s full text.

TL;DR

This article argues that AI's water footprint has been largely overlooked relative to its carbon footprint, and presents a methodology for estimating AI's operational and embodied water consumption. Using GPT-3 as a case study, the authors estimate that training it in Microsoft's U.S. datacenters consumed about 5.4 million liters of water, and they recommend transparency, water-aware scheduling, and joint water–carbon optimization.

Key findings

  • Training GPT-3 (175B parameters) in Microsoft's U.S. datacenters can consume a total of 5.4 million liters of water, including 700,000 liters of scope-1 onsite water consumption; GPT-3 also consumes a 500ml bottle of water for roughly 10–50 medium-length responses depending on when and where it is deployed.
  • Water efficiency varies spatially and temporally: onsite WUE rises with outside wet-bulb temperature (cooling towers) or hot/dry conditions (air cooling), and offsite WUE varies with the grid's fuel mix, so "when" and "where" a model is trained or served materially changes its water footprint.
  • Carbon and water efficiency can conflict: minimizing one footprint may increase the other, so "follow the sun" for carbon may mean "unfollow the sun" for water; the authors argue the two footprints are complementary, not substitutable.
  • Contextual figures cited: one technology company's self-owned datacenters withdrew 29 billion liters and consumed more than 23 billion liters of freshwater for onsite cooling in 2023 (nearly 80% potable), with 20% growth from 2021–2022 and 17% from 2022–2023; U.S. datacenter onsite water consumption in 2028 could reach approximately 150–280 billion liters; global AI water withdrawal is projected at 4.2–6.6 billion cubic meters in 2027, with 0.38–0.60 billion cubic meters consumed.

Methods

  • The paper proposes a general methodology for estimating AI's water consumption footprint (and, by substituting water withdrawal efficiency for WUE, the withdrawal footprint), covering operational water (scope-1 onsite cooling plus scope-2 offsite electricity generation) and embodied scope-3 water amortized over server lifespan.
  • Case study: GPT-3 with 175 billion parameters, using an estimated training energy of 1287 MWh, Microsoft's location-wise annualized PUE and WUE, and the Reig et al. electricity water intensity factors (EWIF); for inference, a conversation task with ~800 words of input and 150–300 words of output and an assumed 0.004 kWh server energy per request.
  • The authors exclude Singapore and Taiwan from the training-location analysis because EWIF data are unavailable there, and they derive recommendations on scheduling and reporting from the spatial-temporal variation of WUE.

Limitations

  • Stated: the embodied (scope-3) water footprint is excluded from the case study due to lack of public data on scope-3 water usage; the training location for GPT-3 is not public, so multiple Microsoft datacenter locations are considered; Singapore and Taiwan are excluded for lack of EWIF data.
  • Stated: the inference estimate is conservative — actual consumption could be several times higher, since a smaller model (Llama-3-70B) already uses ~0.010 kWh per medium-sized request and Falcon-180B ~0.016 kWh, and third-party colocation datacenters may have several times higher scope-1 water consumption; the U.S. EWIF used (3.14 L/kWh) is lower than the recently reported 4.35 L/kWh.
  • Evident from the text: the analysis relies on third-party and company-reported averages (Microsoft PUE/WUE, Meta, Google, Reig et al.) rather than direct measurement, and the paper notes that limited data exist for scope-3 chip-manufacturing water use.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3724499

Cites

20 of 20 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

7 of 49 cited sources checked · 2 supported · 1 flagged · accuracy not yet measured

  • Premisecitation
    § Background

    For the Falcon-180B model, the server-level energy consumption reaches approximately 0.016 kWh per medium-sized request.

    “For the Falcon-180B model, which is comparable in size to GPT-3-175B, the server-level energy consumption reaches approximately 0.016kWh per medium-sized request.30”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [30] [30] Stojkovic, J. et al. DynamoLLM: Designing LLM inference clusters for performance and energy efficiency. In Proceedings of the IEEE Intern. Symp. on High-Performance Computer Architecture (2025). Google Scholar a [...] token phase that produces outputs. b [...] words of input and 150–300 words of output. c [...] and accounting for non-GPU server overhead. d [...] 0.016kWh per medium-sized request. source

      Table III in Section III (p. 3) reports 7.94–10.34 Wh for Falcon-180B medium-sized requests at medium system load, depending on GPU frequency. That is approximately 0.008–0.010 kWh, not 0.016 kWh per request. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Premisecitation
    § Background

    When considering service-level objectives for LLM response times in enterprise-grade Nvidia DGX H100 systems for conversation tasks, the server-level inference energy consumption for Llama-3-70B is already approximately 0.010 kWh per medium-sized request when using a state-of-the-art LLM inference solution and accounting for non-GPU server overhead.

    “when considering service-level objectives (SLOs) for LLM response times in enterprise-grade Nvidia DGX H100 systems for conversation tasks, the server-level inference energy consumption for a much smaller model (for example, Llama-3-70B) is already approximately 0.010kWh per medium-sized request when using a state-of-the-art LLM inference solution and accounting for non-GPU server overhead.30”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [30] [30] Stojkovic, J. et al. DynamoLLM: Designing LLM inference clusters for performance and energy efficiency. In Proceedings of the IEEE Intern. Symp. on High-Performance Computer Architecture (2025). Google Scholar a [...] token phase that produces outputs. b [...] words of input and 150–300 words of output. c [...] and accounting for non-GPU server overhead. d [...] 0.016kWh per medium-sized request. source

      Section III reports measurements on an NVIDIA DGX H100 using vLLM and gives SLO-constrained energy figures for medium-sized Llama3-70B requests. Table III reports 4.28–6.45 Wh per request at medium load, not approximately 10 Wh (0.010 kWh). The cited passages do not establish that adding non-GPU server overhead raises those figures to 10 Wh, or give that figure specifically for Conversation requests. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Premisecitation
    § Background

    Some subsequent models like GPT-4 could consume substantially more energy and water than GPT-3 for processing the same request, especially under the reasoning mode.

    “some subsequent models like GPT-4 could consume substantially more energy and water than GPT-3 for processing the same request, especially under the reasoning mode.3,26”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [3] [3] Banse, A., Rince, S., and Defour, V. EcoLogits Calculator (2024); https://tinyurl.com/2au2xaml. Go to Citation Google Scholar

      Not checked in the free run.

    • cites [26] [26] Shumba, N. et al. A water efficiency dataset for African data centers. In Proceedings of NeurIPS Workshop on Tackling Climate Change with Machine Learning (2024). Go to Citation Google Scholar source

      The Abstract (p. 1) and §4.1 (p. 5) estimate substantially greater water and energy use for GPT-4 than for Llama-3-70B on the same writing tasks. They do not compare GPT-4 with GPT-3 or evaluate reasoning mode, so those parts of the claim are unsupported by this paper. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Findingown data
    § Introduction

    Water usage effectiveness (WUE) varies both spatially and temporally.

    “we show that water usage effectiveness (WUE), a measure of water efficiency, varies both spatially and temporally”
    Quote in the paperSubstantial claim
  • Findingnone given
    § Introduction

    GPT-3 needs to consume a 500ml bottle of water for roughly 10-50 medium-length responses, depending on when and where it is deployed.

    “GPT-3 needs to “drink” (that is, consume) a 500ml bottle of water for roughly 10–50 medium-length responses, depending on when and where it is deployed.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 12
  • May combine two claims. 18

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § Background

    A medium-sized request has approximately 800 words of input and 150-300 words of output.

    “we consider a medium-sized request, each with approximately 800 words of input and 150–300 words of output.30”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [30] [30] Stojkovic, J. et al. DynamoLLM: Designing LLM inference clusters for performance and energy efficiency. In Proceedings of the IEEE Intern. Symp. on High-Performance Computer Architecture (2025). Google Scholar a [...] token phase that produces outputs. b [...] words of input and 150–300 words of output. c [...] and accounting for non-GPU server overhead. d [...] 0.016kWh per medium-sized request. source

      Section III, Table IV (p. 3) classifies requests by input and output length, so it supports the idea of a medium-sized request. But its thresholds are in tokens—fewer than 1,024 input tokens and fewer than 350 output tokens—not approximately 800 input words and 150–300 output words. The paper does not establish that word-count conversion. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation
    § Background

    GPT-3 was trained and deployed by OpenAI in Microsoft's datacenters, with an estimated training energy of 1287 MWh.

    “GPT-3 was trained and deployed by OpenAI in Microsoft's datacenters, with an estimated training energy of 1287MWh.20”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [20] [20] Patterson, D. et al. Carbon Emissions and Large Neural Network Training (2021); arXiv:2104.10350 [cs.LG] Go to Citation Google Scholar source

  • May combine two claims.

    Premisecitation
    § Background

    A conversation task typically includes a CPU-intensive prompt phase that processes user input and a memory-intensive token phase that produces outputs.

    “a conversation task, which typically includes a CPU-intensive prompt phase that processes user input (also known as prompt) and a memory-intensive token phase that produces outputs.30”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [30] [30] Stojkovic, J. et al. DynamoLLM: Designing LLM inference clusters for performance and energy efficiency. In Proceedings of the IEEE Intern. Symp. on High-Performance Computer Architecture (2025). Google Scholar a [...] token phase that produces outputs. b [...] words of input and 150–300 words of output. c [...] and accounting for non-GPU server overhead. d [...] 0.016kWh per medium-sized request. source

      Section II, Background (p. 2) describes a compute-intensive prefill phase that processes input tokens and a memory-intensive decode phase that generates output tokens. It does not characterize the first phase as CPU-intensive; that hardware-specific part of the claim is unsupported. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation
    § Background

    The official estimate indicates that GPT-3 consumes an order of 0.4 kWh of electricity to generate 100 pages of content, equivalent to roughly 0.004 kWh per page.

    “The official estimate indicates that GPT-3 consumes an order of 0.4kWh of electricity to generate 100 pages of content, equivalent to roughly 0.004kWh per page.4”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [4] [4] Brown, T. et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33, Larochelle H. Curran Associates, Inc. (2020), 1877-1901; https://tinyurl.com/yqlf8fzk Google Scholar a [...] with 175 billion parameters as an example, b [...] GPT-3 model with 175 billion parameters. c [...] equivalent to roughly 0.004kWh per page. source

  • May combine two claims.

    Findingcitation
    § Introduction

    Training GPT-3 in Microsoft's U.S. datacenters can consume a total of 5.4 million liters of water, including 700,000 liters of scope-1 onsite water consumption.

    “By taking the GPT-3 model with 175 billion parameters as an example,4 we show that training GPT-3 in Microsoft's U.S. datacenters can consume a total of 5.4 million liters of water, including 700,000 liters of scope-1 onsite water consumption.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given
    § Background

    Water efficiency exhibits spatial-temporal diversity: onsite water efficiency changes with variations of outside weather conditions, and offsite water efficiency changes with variations of the grid's energy fuel mixes meeting time-varying demands.

    “Water efficiency exhibits a spatial-temporal diversity—onsite water efficiency changes due to variations of outside weather conditions, and off-site water efficiency changes due to variations of the grid's energy fuel mixes to meet time-varying demands (Figure 2).”
    • Also stated in: § Conclusion · plain
      “We also discuss that the scope-1 and scope-2 water efficiencies vary spatially and temporally—judiciously deciding “when” and “where” to run a large AI model can significantly cut the water footprint.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Background

    There is misalignment between the scope-2 water consumption intensity factor and the carbon emission rate, so minimizing one footprint might increase the other footprint.

    “This conflict can also be shown in Figure 2a and Figure 2b, where we see misalignment between the scope-2 water consumption intensity factor and carbon emission rate: Minimizing one footprint might increase the other footprint.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Conclusion

    A large AI model can consume millions of liters of water for training, as shown using GPT-3 as an example.

    “using GPT-3 as an example, we show that a large AI model can consume millions of liters of water for training.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § Introduction

    The company's datacenter water consumption increased by 20% from 2021 to 2022 and by 17% from 2022 to 2023.

    “the company's datacenter water consumption increased by 20% from 2021 to 2022 and by 17% from 2022 to 2023,9”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [9] [9] Google. Environmental Report. 2024; https://sustainability.google/reports/. Google Scholar a [...] in Google's sustainability report b [...] nearly 80% of which was potable water. c [...] 17% from 2022 to 2023, d [...] fastest-expanding workloads in datacenters. e [...] from industry leaders, including Google f [...] operated by major technology companies. g [...] freshwater (potable water in many cases h [...] is evaporated and considered “consumption”. i [...] annualized global onsite water efficiency j [...] to minimize onsite water consumption,

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Introduction

    Global AI could consume 85-134 TWh of electricity in 2027.

    “a recent study suggests that the global AI could consume 85–134TWh of electricity in 2027,6”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] [6] de Vries, A. The growing energy footprint of artificial intelligence. Joule (Oct. 2023); https://tinyurl.com/2cmwv5m3 Go to Citation Google Scholar source

  • May combine two claims.

    Premisecitation
    § Introduction

    Optimizing for carbon efficiency does not necessarily result in, and may even worsen, water efficiency.

    “optimizing for carbon efficiency does not necessarily result in, and may even worsen, water efficiency, which varies with the fuel mixes for electricity generation and outside weather in a unique way.10,25”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [10] [10] Islam, M.A. et al. Exploiting spatio-temporal diversity for water saving in geo-distributed data centers. IEEE Transactions on Cloud Computing 6, 3 (2018), 734--746;https://tinyurl.com/2c43smrv Google Scholar a [...] and outside weather in a unique way. b [...] the outside temperature (see Islam et al. c [...] and water footprints are not substitutable. source
    • cites [25] [25] Shehabi, A., et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory LBNL-2001637 (Dec. 2024). Google Scholar a [...] consumption at an unprecedented rate. b [...] the recent U.S. datacenter energy report, c [...] stressing the water infrastructures. d [...] fastest-expanding workloads in datacenters. e [...] U.S. alone will surpass 150–300TWh in 2028. f [...] in 2028 if the projection in Shehabi et al. g [...] and outside weather in a unique way. h [...] operated by major technology companies. i [...] the recent U.S. datacenter energy report. j [.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisereasoning
    § Introduction

    AI's water footprint has largely remained under the radar and keeps escalating.

    “However, AI's water footprint—many millions of liters of freshwater consumed for cooling the servers and for electricity generation—has largely remained under the radar and keeps escalating.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The majority of AI's water footprint is blue water extracted from rivers, lakes, or groundwater.

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § Background

    To cut the water footprint, it may be more appealing to 'unfollow the sun' to avoid high-temperature hours of a day when WUE is high.

    “Nonetheless, to cut the water footprint, it may be more appealing to “unfollow the sun” to avoid high-temperature hours of a day when WUE is high.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § Acknowledgment

    Large datacenters often adopt sustainability programs (for example, renewable purchasing agreements) to offset their location-based electricity usage and thus may have lower market-based carbon and water footprints.

    “Large datacenters often adopt sustainability programs (for example, renewable purchasing agreements) to offset their location-based electricity usage and thus may have lower market-based carbon and water footprints.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § Background

    The EWIF for the U.S. (3.14 L/kWh) is conservative and significantly lower than the recently reported 4.35 L/kWh.

    “our EWIF for the U.S. (3.14L/kWh) is conservative and significantly lower than the recently reported 4.35 L/kWh.25”
    • Also stated in: § Background · plain
      “the average EWIF provided by Reig et al.23 is 3.14L/kWh and significantly lower than 4.35L/kWh noted by the recent report.25”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [23] [23] Reig, P., Luo, T., Christensen, E., and Sinistore, J. Guidance for calculating water use embedded in purchased electricity. World Resources Institute (2020). Google Scholar a [...] that is, water withdrawal footprint (WWF) b [...] of water withdrawal and water consumption. c [...] and 3.1L/kWh, d [...] regions due to different energy fuel mixes. e [...] we use the EWIF provided by Reig et al. f [...] the average EWIF provided by Reig et al. g [...] regions is not available in Reig et al. h [...] intensity factor vs. carbon emission rate. i [...] consumption intensity for each fuel type. j

      Not checked in the free run.

    • cites [25] [25] Shehabi, A., et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory LBNL-2001637 (Dec. 2024). Google Scholar a [...] consumption at an unprecedented rate. b [...] the recent U.S. datacenter energy report, c [...] stressing the water infrastructures. d [...] fastest-expanding workloads in datacenters. e [...] U.S. alone will surpass 150–300TWh in 2028. f [...] in 2028 if the projection in Shehabi et al. g [...] and outside weather in a unique way. h [...] operated by major technology companies. i [...] the recent U.S. datacenter energy report. j [.

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given
    § Background

    To cut the carbon footprint, it is preferable to 'follow the sun' when solar energy is more abundant.

    “To cut the carbon footprint, it is preferable to “follow the sun” when solar energy is more abundant.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Background

    Datacenters have increasingly adopted climate-conscious cooling system designs to minimize onsite water consumption.

    “Although datacenters have increasingly adopted climate-conscious cooling system designs to minimize onsite water consumption,9,14,29”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [9] [9] Google. Environmental Report. 2024; https://sustainability.google/reports/. Google Scholar a [...] in Google's sustainability report b [...] nearly 80% of which was potable water. c [...] 17% from 2022 to 2023, d [...] fastest-expanding workloads in datacenters. e [...] from industry leaders, including Google f [...] operated by major technology companies. g [...] freshwater (potable water in many cases h [...] is evaporated and considered “consumption”. i [...] annualized global onsite water efficiency j [...] to minimize onsite water consumption,

      Not checked in the free run.

    • cites [14] [14] Meta. Sustainability: Water (2023); https://tinyurl.com/2c65p7u3. Google Scholar a [...] control when the outside air is too dry. b [...] to minimize onsite water consumption,

      Not checked in the free run.

    • cites [29] [29] Solomon, S. Sustainable by design: Next-generation datacenters Consume zero water for cooling (2024); https://tinyurl.com/2xmxxxnz Google Scholar a [...] which consume no onsite water year-round. b [...] to minimize onsite water consumption,

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Background

    Some technology companies such as Meta have begun to include scope-2 water consumption in their sustainability reports.

    “some technology companies such as Meta have begun to include scope-2 water consumption in their sustainability reports.15”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] [15] Meta. Sustainability Report (2024); https://tinyurl.com/2yagb6vk. Google Scholar a [...] 70% is consumed based on Meta's report. b [...] divided by 14,975,435MWh) in 2023. c [...] in their sustainability reports.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Background

    Apple reports that its supply chain accounts for 99% of its total water footprint.

    “Apple reports that its supply chain accounts for 99% of its total water footprint.2”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [2] [2] Apple. Environmental Responsibility Report. 2024; https://tinyurl.com/jwg93lu. Google Scholar a [...] for 99% of its total water footprint. b [...] scope-3 supply-chain water footprint,

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Background

    The average recycling rates for wafer plants and semiconductor plants in Singapore are 45% and 27%, respectively.

    “the average recycling rates for wafer plants and semiconductor plants in Singapore are 45% and 27%, respectively.28”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28] [28] Singapore Public Utilities Board. Sectoral Water Efficiency Benchmark; https://tinyurl.com/2935uhn8 Google Scholar a [...] manufacturing uses a huge amount of water. b [...] in Singapore are 45% and 27%, respectively.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning
    § Background

    Reporting scope-2 water consumption as a standard practice would make offsite water consumption visible to AI model developers and end users and can unlock new opportunities for demand-side flexibility, thereby alleviating the overall strain on water resources.

    “This approach makes the offsite water consumption visible to AI model developers as well as end users and can unlock new opportunities for demand-side flexibility, thereby alleviating the overall strain on water resources.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: A comprehensive understanding and reporting of AI's scope-2 water consumption associated with electricity generation is limited.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § Background

    If informed of real-time water efficiency, some water-conscious users may prefer to use AI inference during water-efficient hours and/or in water-efficient datacenters, which can reduce AI's water footprint by enabling demand-side flexibility.

    “Likewise, if informed of the real-time water efficiency, some water-conscious users may prefer to use AI inference during water-efficient hours and/or in water-efficient datacenters, which can reduce AI's water footprint by enabling demand-side flexibility.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Background

    Our estimate of inference water consumption for GPT-3 is on the conservative side, and the actual water consumption could be several times higher.

    “Our estimate of inference water consumption for GPT-3 is on the conservative side, and the actual water consumption could be several times higher.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The official estimate of GPT-3's energy consumption likely considers only the GPU energy used during token generation.

  • May combine two claims.

    Premisecitation
    § Background

    Onsite WUE increases significantly for cooling towers when the outside wet bulb temperature increases, and increases for outside air cooling when the outside dry bulb temperature is too hot or the humidity is too low.

    “increases significantly for cooling towers when the outside wet bulb temperature increases, and increases for outside air cooling when the outside dry bulb temperature is too hot or the humidity is too low.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [10] [10] Islam, M.A. et al. Exploiting spatio-temporal diversity for water saving in geo-distributed data centers. IEEE Transactions on Cloud Computing 6, 3 (2018), 734--746;https://tinyurl.com/2c43smrv Google Scholar a [...] and outside weather in a unique way. b [...] the outside temperature (see Islam et al. c [...] and water footprints are not substitutable. source

  • May combine two claims.

    Premisecitation
    § Background

    The U.S. national average water withdrawal and consumption for electricity generation are estimated at about 43.8 L/kWh and 3.1 L/kWh, respectively.

    “the U.S. national average water withdrawal and consumption are estimated at about 43.8L/kWh32 and 3.1L/kWh,23 respectively.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [32] [32] U.S. Energy Information Administration. Open Data; https://tinyurl.com/2c5qq6wk. Google Scholar a [...] are estimated at about 43.8L/kWh b [...] consumption intensity for each fuel type. c [...] Virginia, starting from Apr. 4, 2022.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [23] [23] Reig, P., Luo, T., Christensen, E., and Sinistore, J. Guidance for calculating water use embedded in purchased electricity. World Resources Institute (2020). Google Scholar a [...] that is, water withdrawal footprint (WWF) b [...] of water withdrawal and water consumption. c [...] and 3.1L/kWh, d [...] regions due to different energy fuel mixes. e [...] we use the EWIF provided by Reig et al. f [...] the average EWIF provided by Reig et al. g [...] regions is not available in Reig et al. h [...] intensity factor vs. carbon emission rate. i [...] consumption intensity for each fuel type. j

      Not checked in the free run.

25 of 25

  • The paper relies on this without support.

    The majority of AI's water footprint is blue water extracted from rivers, lakes, or groundwater.

    “the majority of AI's water footprint is blue water extracted from rivers, lakes, or groundwater, which is directly accessible for human use but often more limited in availability.”

  • The paper relies on this without support.

    The official estimate of GPT-3's energy consumption likely considers only the GPU energy used during token generation.

    “While no details are provided, the estimate likely considers only the GPU energy used during token generation.”

  • The paper relies on this without support.

    A comprehensive understanding and reporting of AI's scope-2 water consumption associated with electricity generation is limited.

    “a comprehensive understanding and reporting of AI's scope-2 water consumption associated with electricity generation remain limited.”

  • The paper relies on this without support.

    There is limited data available for embodied water usage by chip manufacturing.

    “there is limited data available for embodied water usage by chip manufacturing.”
4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Ahmed K. Islam M.A. Ren S. and Quan G. Exploiting temporal diversity of water efficiency to make data center less “thirsty”. In Proceedings of the 11th Intern. Conf of Autonomic Computing (2014).

Cited by

Sorted by citations

Related

Authors
Li, P., Yang, J., Islam, M. A., & Ren, S.
Affiliations
  • University of California, Riverside (US)
  • University of Houston (US)
  • The University of Texas at Arlington (US)
ORCID
Venue
Communications of the ACM
Publisher
Association for Computing Machinery (ACM)
Published
2025-06-27
Volume
68
Issue
7
Pages
54-61
ISSN
  • 0001-0782 (ISSN-L)
  • 1557-7317
DOI
10.1145/3724499
OpenAlex
W4411368489
Semantic Scholar
05fae03311d24e3f4a82e8021153f3d68d3bdf63
Cited by
176
References
15
Citations by year
2026: 138 · 2025: 38
FWCI
39.38
Type
journal article
Language
en
Open access
Licence
Locations
Communications of the ACM (publishedVersion, cc-by-nd)
Funders
NSF (2324915, 2324916, 2152357)
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Green IT and Sustainability
  • Physical Sciences › Computer Science › Hardware and Architecture › Energy Efficiency in Computing
  • Physical Sciences › Environmental Science › Environmental Engineering › Environmental Impact and Sustainability
Keywords
energy consumption of AI, water footprint, AI training, environmental impact of AI, water scarcity, energy efficiency

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-30.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-30.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review