sustainability

DOI:10.1145/ 3724499
a major roadblock to sustainability
Uncovering and addressing the secret water and create social conflicts, as
freshwater resources suitable for
footprint of AI models. human use are extremely limited
and unevenly distributed.
BY PENGFEI LI, JIANYI YANG, MOHAMMAD A. ISLAM, As acknowledged in Google's
AND SHAOLEI REN sustainability report9 and the recent

Making
U.S. datacenter energy report,25
the expansion of AI products and
services is a key driver of the rapid
increase in datacenter water con-
sumption. Even excluding the water
usage in leased third-party colo-

AI Less
cation facilities, one technology
company's self-owned datacenters
alone directly withdrew 29 billion
liters and consumed (that is, evapo-
rated) more than 23 billion liters of
freshwater for onsite cooling in 2023,

‘Thirsty’
nearly 80% of which was potable
water.9,a This amount of annual water
consumption even rivals that of a
major household-name beverage
company.21 Importantly, the compa-
ny's datacenter water consumption
increased by ∼ ​ ​20% from 2021 to 2022
and by ∼ ​ ​17% from 2022 to 2023,9 and
another technology company's data-
center water consumption saw ∼ ​ ​34%
and ​∼​22% increases over the same
periods, respectively.18 Furthermore,
according to the recent U.S. datacen-
ter energy report, the total annual
onsite water consumption by U.S.
has enabled remarkable
A R T I F ICI A L I N T EL L IGENCE (A I)
datacenters in 2028 could double or
breakthroughs in numerous areas of critical even quadruple the 2023 level, reach-
importance, including tackling global challenges such ing approximately 150–280 billion
liters and further stressing the water
as climate change. On the other hand, many AI models, infrastructures.25
IMAGE BY AND RIJ BORYS ASSOCIAT ES WITH ASSISTANCE OF SH UT TERSTOCK.A I

especially large generative ones like GPT-4, are trained AI represents the fastest-expand-
ing workloads in datacenters.9,25 For
and deployed on energy-hungry servers in warehouse- example, a recent study suggests that
scale datacenters, accelerating the datacenter energy global AI could consume 85–134TWh
consumption at an unprecedented rate.25 As a result, of electricity in 2027,6 whereas a
more aggressive projection by the
AI's carbon footprint has been undergoing scrutiny, recent U.S. datacenter energy report
driving the recent progress in AI carbon efficiency.24,31 predicts that AI servers' electricity
consumption in the U.S. alone will
However, AI's water footprint—many millions of liters surpass 150–300TWh in 2028.25 Even
of freshwater consumed for cooling the servers and for
electricity generation—has largely remained under the a The detailed difference between water with-
radar and keeps escalating. If not properly addressed, drawal and water consumption is presented
in the section Water Withdrawal versus Water
AI's water footprint can potentially become Consumption.

54 COMM UNICATIO NS O F THE ACM | J U LY 2025 | VO L . 68 | NO. 7

 sustainability

JU LY 2 0 2 5 | VO L. 6 8 | N O. 7 | C OM M U N IC AT ION S OF T HE ACM 55
sustainability

considering the lower estimate, the structure. The urgency can also be
combined scope-1 and scope-2 water reflected in part by the recent com-
withdrawal of global AI is projected mitment to “Water Positive by 2030”
to reach 4.2–6.6 billion cubic meters from industry leaders, including
in 2027, which is more than the total Google9 and Microsoft,18 and by the
annual water withdrawal of four to inclusion of water footprint as a key
six Denmarks or half the U.K.b Simul- metric into the world's first interna-
taneously, a total of 0.38–0.60 billion tional standard on sustainable AI to
cubic meters of water will be evapo- be published by the ISO/IEC.19
rated and considered “consumption” In this article, we advocate for
The combined due to the global AI demand in 2027. a holistic approach to sustainable
scope-1 and scope-2 Moreover, these global estimates will
be exceeded by the total water with-
AI that extends beyond the carbon
footprint to also address the water
water withdrawal drawal and consumption attributed footprint. Specifically, we present a

of global AI is to AI in the U.S. alone in 2028 if the

projection in Shehabi et al.25 comes
principled methodology to estimate
AI's total water footprint, including
projected to reach to fruition.
Despite its profound environmen-
both operational water and embod-
ied water. By taking the GPT-3 model
4.2–6.6 billion cubic tal and societal impact, the increas- with 175 billion parameters as an
meters in 2027, ing water footprint of AI has received
disproportionately less attention
example,4 we show that training GPT-
3 in Microsoft's U.S. datacenters can
more than the from the AI community as well as the consume a total of 5.4 million liters
total annual water general public. For example, while
scope-2 carbon emissions are rou-
of water, including 700,000 liters of
scope-1 onsite water consumption.
withdrawal of four tinely included as part of AI model Additionally, GPT-3 needs to “drink”

to six Denmarks or cards, even scope-1 direct water us-

age (either withdrawal or consump-
(that is, consume) a 500ml bottle of
water for roughly 10–50 medium-
half the U.K. tion) is missing, let alone scope-2 length responses, depending on
water usage. This may impede inno- when and where it is deployed.
vations to enable water sustainability Next, we show that water usage
and build truly sustainable AI. Cru- effectiveness (WUE), a measure of
cially, water and carbon footprints water efficiency, varies both spa-
are complementary to, not substitut- tially and temporally, suggesting
able of, each other for understanding that judiciously deciding when and
the environmental impacts. Indeed, where to train a large AI model can
optimizing for carbon efficiency does significantly cut the water footprint.
not necessarily result in, and may We also emphasize the need for
even worsen, water efficiency, which increasing transparency of AI's water
varies with the fuel mixes for electric- footprint, including disclosing more
ity generation and outside weather in information about operational data
a unique way.10,25 and keeping users informed of the
To ensure growth in AI does not runtime water efficiency. Finally, we
exacerbate the global water stresses highlight the necessity of holistically
or outweigh the environmental ben- addressing water footprint along
efits it provides, it is a critical time with carbon footprint to enable truly
to uncover and address AI's hidden sustainable AI—the water footprint of
water footprint amid the increasing- AI can no longer stay under the radar.
ly severe freshwater scarcity crisis,
worsened extended droughts, and Background
quickly aging public water infra- Water withdrawal versus water con-
sumption. There are two related but
different concepts: water withdrawal
b The scope definition of water usage23 is in line
with that of carbon emissions. Our scope-2
and water consumption, both of
water withdrawal (and consumption when ap- which are important for understand-
plicable) is for location-based electricity gen- ing the impacts on water stress and
eration throughout the article. Large datacen- availability.13,22
ters often adopt sustainability programs (for ˲ Water withdrawal refers to fresh-
example, renewable purchasing agreements)
to offset their location-based electricity usage
water taken from ground- or surface-
and thus may have lower market-based carbon water sources, either temporarily
and water footprints. or permanently, and then used for

56 COMM UNICATIO NS O F THE AC M | J U LY 2025 | VO L . 68 | NO. 7

 sustainability

agricultural, industrial, or municipal Figure 1. Example of a datacenter’s operational water usage: onsite scope-1 water usage
uses (normally excluding water used for datacenter cooling (via cooling towers in the example), and offsite scope-2 water us-
for hydroelectricity generation).22 As age for electricity generation. The icons for AI models are only for illustration purposes.
water is a finite shared resource, wa-
ter withdrawal indicates the level of
Heat
competition as well as dependence Scope-1 Water Exchanger
on water resources among different
sectors. Cooling
˲ Water consumption is defined Tower
as “water withdrawal minus water
Scope-2 Water Warm
discharge”; it refers to the amount Water
of water “evaporated, transpired,
Cooling
incorporated into products or crops, Tower Chilled
Water
or otherwise removed from the im- Power
mediate water environment.”13 Water Plant
consumption reflects the impact on
downstream water availability and is
crucial for assessing watershed-level Water ChatGPT AlphaGO
scarcity.22 Source

These two types of water usage Server Rack CRAH Data

Center
correspond to two different water
footprints, that is, water withdrawal
footprint (WWF)5,23 and water con-
sumption footprint (WCF), respec- proaches used in many datacenters, energy: 1L/kWh for Google's annual-
tively.27 By default, water footprint including those operated by major ized global onsite water efficiency 9
refers to the WCF unless otherwise technology companies.9,25 and 9 L/kWh for a large commercial
specified. Cooling tower. As illustrated in datacenter during the summer in
How does AI use water? AI's water Figure 1, some water is evaporated Arizona.11
usage spans three scopes: onsite (that is, “consumed”) in the cooling Air cooling with water evapora-
water for datacenter cooling (scope tower to dissipate heat into the envi- tion assistance. When the climate
1), offsite water for electricity gen- ronment, while the remaining water condition is appropriate, datacen-
eration (scope 2), and supply-chain moves along an open loop to the heat ters may use “free” outside air to
water for server manufacturing exchanger to further absorb server directly reject the heat to the outside
(scope 3). heat. Additionally, non-evaporated environment. Nonetheless, water
Scope-1 water usage. Nearly all water can be recycled only a few evaporation is still needed when the
server energy is converted into heat, times (typically 3–10 cycles, depend- outside air is too hot (for example,
which must then be removed from ing on water quality) before dis- higher than 85°F); additionally,
the datacenter server room to avoid charge, requiring continuous clean water is also needed for humidity
overheating. This process involves freshwater replenishment to prevent control when the outside air is too
two sequential stages: server-level mineral and salt buildup. Thus, to dry.14 The added water is considered
cooling followed by facility-level keep the cooling tower working, new “withdrawal,” out of which about
cooling. water must be constantly added to 70% is consumed based on Meta's
In the server-level cooling stage, make up for the evaporated water report.15 Generally, outside air cool-
heat is transferred from the servers and discharged water. Importantly, ing is more water-efficient than
to the facility or a heat exchanger, clean freshwater (potable water in cooling towers on average. However,
typically using either air- or liquid- many cases9) is needed to avoid pipe hot weather raises the evaporative
cooling methods (for example, clogs and/or bacterial growth. water demand and maximum water
direct-to-chip cooling or immersion For cooling towers, water with- consumption, potentially stressing
cooling), which do not evaporate drawal refers to the amount of added local water supplies during peak
or consume water. In general, new water, including both evaporated demand on hot days. Additionally,
datacenters dedicated to AI training water and discharged water, while the application of outside air cooling
often rely on liquid cooling due to water consumption exclusively indi- may have challenges in hot regions
the high server power densities. cates the amount of evaporated wa- and/or for many colocation facilities
In the facility-level cooling stage, ter. With good water quality, roughly located in business districts.
heat is rejected from the datacenter 80% of water withdrawal is evapo- Some datacenters may opt for dry
facility to the outside environment. rated and considered “consump- coolers, which consume no onsite
While there are various cooling tion.”9 On average, depending on the water year-round.29 However, this
methods, water-intensive cooling weather conditions and operational approach typically increases cooling
towers and water evaporation-assist- settings, datacenters can evaporate energy consumption compared to
ed air cooling are two common ap- approximately 1–9L/kWh of server water-based cooling methods, poten-

JU LY 2 0 2 5 | VO L. 6 8 | N O. 7 | C OM M U N IC AT ION S OF T HE ACM 57
sustainability

tially exacerbating the overall stress is blue water extracted from riv-
on water resources due to higher ers, lakes, or groundwater, which is
scope-2 water consumption. directly accessible for human use but
Scope-2 water usage. In many often more limited in availability.
countries, thermoelectric power
is among the top sectors in terms Estimating AI's Water Footprint
of water withdrawal and water We present a general methodology
consumption.23 Thus, similarly to for estimating AI's water consump-
scope-2 carbon emissions, data- tion footprint. To obtain the water
centers are accountable for off-site withdrawal footprint, we simply
Just as addressing scope-2 water usage associated with replace the WUE with water with-
scope-2 carbon electricity consumption, which
forms part of the “true water cost
drawal efficiency.
Operational water footprint. We
emissions is of datacenters,” as highlighted by collectively refer to onsite scope-1

important for the recent U.S. datacenter energy

report.25
water and off-site scope-2 water as
the operational water.
mitigating climate Different power plants use dif-
ferent amounts of water for each
˲ Onsite WUE. We denote the on-
site scope-1 WUE at time ​t​ by ​​ρs1,t ​  ​​​,
change, it is equally kWh generation, depending on which is defined as the ratio of the
crucial to address the cooling techniques. Typically,
water withdrawal due to hydropower
onsite water consumption to server
energy consumption and varies
scope-2 water generation is excluded, but water over time depending on the outside
consumption to consumption due to increased
water evaporation rates from hy-
temperature (see Islam et al.10 for
an example of onsite WUE based
reduce AI’s “true dropower generation is included.25 on cooling towers). Concretely, ​​ρs1,t ​  ​​​

water cost.” For electricity generation, the U.S.

national average water withdrawal
increases significantly for cooling
towers when the outside wet bulb
and consumption are estimated at temperature increases, and increas-
about 43.8L/kWh32 and 3.1L/kWh,23 es for outside air cooling when the
respectively. Meta's self-reported outside dry bulb temperature is too
scope-2 water consumption for its hot or the humidity is too low.
global datacenter fleet was 3.7L/kWh ˲ Offsite WUE. We denote the
(that is, 55,475 megaliters divided by off-site scope-2 WUE at time t​ ​ as ​​
14,975,435MWh) in 2023.15 ρ​ s2,t​​​, which is defined as the ratio of
Scope-3 water usage. AI chip and offsite water consumption for each
server manufacturing uses a huge kWh of electricity consumption and
amount of water.7,28 For example, measures the electricity water inten-
ultrapure water is needed for wafer sity factor (EWIF). While there are
fabrication and water is also needed different methods to estimate ρ​  ​​ s2,t​​​, a
to keep semiconductor plants cool. common one is weighted averaging: ​​
​∑ ​​b​ ​ × EWI ​F​​
Importantly, the discharged water ρ​s2,t​ = ​_ k k,t
​∑ ​​ ​b​ ​
k k,t
k
​​ where ​​bk,t​  ​​​ denotes the
may contain toxic chemicals and/or amount of electricity generated from
hazardous wastes. While water re- fuel type k​ ​at time t​ ​for the grid serv-
cycling at semiconductor plants can ing the datacenter under consider-
effectively reduce water withdrawal, ation, and E ​ WI ​Fk​  ​​​is the EWIF for fuel
the recycling rate in many cases re- type k​ ​.1,8 Thus, variations in energy
mains low; for example, the average fuel mixes of electricity generation
recycling rates for wafer plants and result in temporal variations of the
semiconductor plants in Singapore offsite WUE. Moreover, the offsite
are 45% and 27%, respectively.28 WUE also varies across regions due
Although largely obscure, scope-3 to different energy fuel mixes.23,25
water usage is likely significant.7 For ˲ Operational water footprint.
instance, Apple reports that its sup- Consider a time-slotted model ​
ply chain accounts for 99% of its total t = 1, 2, ⋯ , T​, where the length of
water footprint.2 each time slot depends on how
It is important to recognize that frequently we want to assess the
unlike agriculture, whose water foot- operational water footprint. At time ​
print is mostly green (that is, water t, suppose an AI model uses energy e​ ​​ t​​​
stored in soil and used by plants), which can be measured using power
the majority of AI's water footprint meters and/or servers' built-in tools,

58 COM MUNICATIO NS O F TH E ACM | J U LY 2025 | VO L . 68 | NO. 7

 sustainability

Table 1. Estimate of GPT-3’s operational water consumption footprint. “*” denotes datacenters under construction as of July 2023,
whose PUE and WUE are projected by Microsoft.

Water for Training Water for Each Request

(million L) (mL)
# of Requests
Onsite WUE Offsite EWIF Onsite Offsite Total Onsite Offsite Total for 500ml
Location PUE (L/kWh) (L/kWh) Water Water Water Water Water Water Water
U.S. Average 1.170 0.550 3.142 0.708 4.731 5.439 2.200 14.704 16.904 29.6
Arizona 1.180 1.630 4.959 2.098 7.531 9.629 6.520 23.406 29.926 16.7
Georgia* 1.120 0.060 2.309 0.077 3.328 3.406 0.240 10.345 10.585 47.2
Illinois 1.350 0.740 2.233 0.952 3.880 4.833 2.960 12.060 15.020 33.3
Iowa 1.160 0.140 3.104 0.180 4.634 4.814 0.560 14.403 14.963 33.4
Texas 1.280 0.250 1.287 0.322 2.120 2.442 1.000 6.590 7.590 65.9
Virginia 1.140 0.140 2.385 0.180 3.499 3.679 0.560 10.875 11.435 43.7
Washington 1.150 0.950 9.501 1.223 14.063 15.285 3.800 43.706 47.506 10.5
Wyoming 1.110 0.130 2.574 0.167 3.677 3.845 0.520 11.429 11.949 41.8
Australia* 1.120 0.012 4.259 0.015 6.138 6.154 0.048 19.078 19.126 26.1
Denmark* 1.160 0.010 3.180 0.013 4.747 4.760 0.040 14.754 14.794 33.8
Finland* 1.120 0.010 4.542 0.013 6.548 6.561 0.040 20.350 20.390 24.5
India* 1.430 0.000 3.445 0.000 6.340 6.340 0.000 19.704 19.704 25.4
Indonesia* 1.320 1.900 2.271 2.445 3.858 6.304 7.600 11.992 19.592 25.5
Ireland 1.190 0.020 1.476 0.026 2.261 2.287 0.080 7.027 7.107 70.4
Mexico* 1.120 0.056 5.300 0.072 7.639 7.711 0.224 23.742 23.966 20.9
Netherlands 1.140 0.060 3.445 0.077 5.054 5.131 0.240 15.708 15.948 31.4
Sweden 1.160 0.090 6.019 0.116 8.986 9.101 0.360 27.927 28.287 17.7

and the datacenter hosting the AI popular online service, is a large number of Microsoft's datacenters
model has a power usage effective- language model (LLM) based on are located in the U.S., where the av-
ness (PUE) of ​​θt​  ​​​that accounts for the subsequent versions of GPT-3. We erage EWIF provided by Reig et al.23
non-IT energy overhead. Then, the present a case study to estimate the is 3.14L/kWh and significantly lower
total operational water footprint of operational water consumption for than 4.35L/kWh noted by the recent
the AI model can be written as W ​ ater- the full GPT-3 model with 175 billion report.25 The specific location for
Operational = ​∑ t=1 T
​​​ ​et​  ​​  · ​[​ρs1,t
​  ​​  + ​θ​ t​​  · ​ρ​ s2,t​​]​​. parameters.4 We exclude embodied training GPT-3 is not public. Thus,
Embodied water footprint. Simi- water footprint due to the lack of we consider Microsoft's different
lar to accounting for the embodied public data for scope-3 water us- datacenter locations while excluding
carbon footprint,12 the total scope-3 age. We choose GPT-3, as Microsoft Singapore and Taiwan, as EWIF data
water footprint is amortized over publishes its location-wise WUE and for these regions is not available in
the lifespan of a server. Specifically, PUE.16,17 The results are summarized Reig et al.23
if W
​ ​represents the total water used in Table 1. Inference. As a representative us-
to manufacture the AI servers and Training. GPT-3 was trained and age scenario for an LLM, we consider
the servers are expected to operate deployed by OpenAI in Microsoft's a conversation task, which typically
for a period of T​  ​​ 0​​​, then the embodied datacenters, with an estimated train- includes a CPU-intensive prompt
water footprint over a period of T ​ ​ is ing energy of 1287MWh.20 In line phase that processes user input (also
calculated as W ​ aterEmbodied = ​T_ 0
​T​  ​​ ​​
·W
with the practice of estimating the known as prompt) and a memory-
By adding the operational and carbon footprint, we use the most intensive token phase that produces
embodied water footprints, we recent annualized average onsite outputs.30 More specifically, we
obtain the total water footprint as ​ PUE and WUE for each location, as consider a medium-sized request,
WaterTotal = ​∑ t=1 T
​​​ ​et​  ​​  · ​[​ρs1,t ​  ​​  + ​θ​ t​​  · ​ρ​ s2,t​​]​  reported by Microsoft.16,17 For power each with approximately ​≤​800 words
_
+ ​ ​T​  ​​ ​​. In practice, to obtain a rough
T ·W
plant water efficiency, different refer- of input and 150–300 words of out-
0

estimate, we can use the average val- ences may provide different esti- put.30 The official estimate indicates
ues for the annualized WUE and the mates of EWIF. Thus, for consistency that GPT-3 consumes an order of
estimated AI server energy consump- across regions, we use the EWIF 0.4kWh of electricity to generate
tion. provided by Reig et al.23 to estimate 100 pages of content, equivalent to
Case Study: Estimating GPT-3's scope-2 water consumption, as it roughly 0.004kWh per page.4 While
operational water-consumption employs the same methodology for no details are provided, the estimate
footprint. The core of ChatGPT, a calculating EWIF. Moreover, a large likely considers only the GPU energy

JU LY 2 0 2 5 | VO L. 6 8 | N O. 7 | C OM M U N IC AT ION S OF T HE ACM 59
sustainability

Figure 2. (a) The U.S. eGRID-level scope-2 water consumption intensity factor vs. carbon emission rate.23,33 The dashed line represents a
linear regression model, showing that the eGRID-level scope-2 carbon emission and water consumption efficiencies are not aligned. (b)
A five-day snapshot of scope-2 carbon emission rate and water consumption intensity in Virginia, starting from Apr. 4, 2022. The values
are calculated based on the fuel mixes, carbon emission rate and water consumption intensity for each fuel type.23,32,33 The scope-2
carbon and water efficiencies only have a weak Pearson correlation coefficient of 0.06 in Virginia. (c) A five-day snapshot of energy fuel
mixes serving Virginia, starting from Apr. 4, 2022.32

3.0 0.40 Coal Oil Hydro

16 Natural Gas Wind Other
60
AKMS Nuclear Solar
2.7

Carbon (kg/kWh)
0.35
Water (L/kWh)

Water (L/kWh)
12

Percentage
2.4 40
NYUP NWPP
8 0.30
AZNM 2.1 20
4 0.25
MROE 1.8
HIOA 0
0 1.5 0.20
0.1 0.3 0.5 0.7 0.9 MON TUE WED THU FRI MON TUE WED THU FRI
Carbon (kg/kWh)
(a) Carbon/water efficiency (b) Hourly carbon/water efficiency (c) Hourly energy fuel mixes

used during token generation. To tive and significantly lower than the sustainable AI. As an initial step to
account for both the prompt phase recently reported 4.35 L/kWh.25 raise awareness among end users
and the non-GPU energy consump- While no official information is about the water resource impacts
tion of servers, we assume a per-re- available on resource consumption, of their AI usage, we recommend
quest server energy consumption of some subsequent models such as tracking and reporting AI's water
0.004kWh for our conversation task. GPT-4 could consume substantially consumption in AI model cards and/
The PUE, WUE, and EWIF are the more energy and water than GPT-3 or through cloud dashboards.
same as those used for estimating for processing the same request, Moreover, a comprehensive
the training water consumption. especially under the reasoning understanding and reporting of AI's
Our estimate of inference water mode.3,26 With continued efforts to scope-2 water consumption asso-
consumption for GPT-3 is on the reduce AI's computational demand ciated with electricity generation
conservative side, and the actual and improve overall water efficiency, remain limited. While datacenters
water consumption could be several the water consumption per request have increasingly adopted climate-
times higher. Specifically, when may decrease in the future. However, conscious cooling-system designs
considering service-level objectives the total water consumption is likely to minimize onsite water consump-
(SLOs) for LLM response times in to continue rising due to the grow- tion,9,14,29 these efforts mostly focus
enterprise-grade Nvidia DGX H100 ing demand for AI services and the on scope-1 water usage while largely
systems for conversation tasks, the increasing scale of AI applications.25 overlooking scope-2 impacts. Just as
server-level inference energy con- addressing scope-2 carbon emissions
sumption for a much smaller model Our Recommendations is important for mitigating climate
(for example, Llama-3-70B) is already We provide our recommendations change, it is equally crucial to address
approximately 0.010kWh per medi- to address AI's water footprint from scope-2 water consumption to reduce
um-sized request when using a state- scheduling and policy perspectives, AI's “true water cost,” as noted by the
of-the-art LLM inference solution making future AI more environmen- recent U.S. datacenter energy report.25
and accounting for non-GPU server tally sustainable. To better reflect the true impacts
overhead.30 For the Falcon-180B More transparency and compre- of datacenters on water resources,
model, which is comparable in size hensive reporting. Despite its grow- some tech companies such as Meta
to GPT-3-175B, server-level energy ing importance, AI's water footprint have begun to include scope-2 water
consumption reaches approximately has received relatively less attention. consumption in their sustainability
0.016kWh per medium-sized re- For example, while AI model cards reports.15 We recommend reporting
quest.30 Furthermore, we emphasize routinely include carbon emissions scope-2 water consumption as a stan-
that Microsoft's datacenters already and serve as an important report- dard practice. This approach makes
have some of the lowest onsite WUE ing framework for understanding offsite water consumption visible to
in the industry. If the same model AI's environmental impacts, they AI model developers as well as end us-
is deployed in a third-party coloca- currently omit information on AI's ers and can unlock new opportunities
tion datacenter, the scope-1 direct water consumption. The lack of for demand-side flexibility, alleviating
water consumption may be several transparency may obstruct efforts to the overall strain on water resources.
times higher. Additionally, our EWIF drive innovations that enhance wa- Finally, despite the enormous
for the U.S. (3.14L/kWh) is conserva- ter sustainability and support truly scope-3 supply-chain water foot-

60 COM MUNICATIO NS O F TH E AC M | J U LY 2025 | VO L . 68 | NO. 7

 sustainability

print,2 there is limited data available Conclusion data centers: A case study in hot-arid climates.
Resources, Conservation and Recycling 181 (2022),
for embodied water usage by chip In this article, we uncover AI's water 106194;https://tinyurl.com/289qqpu3
manufacturing. We recommend usage as a critical concern for socially 12. Luccioni, A.S., Viguier, S., and Ligozat, A-L. Estimating
the carbon footprint of BLOOM, a 176B parameter
further research on scope-3 water responsible and environmentally language model. J. Mach. Learn. Res. 24, 1, Article 253
consumption to achieve a compre- sustainable AI. We present a prin- (Mar. 2024), 15.
13. Macknick, J., Newmark, R., Heath, G., and Hallett,
hensive understanding of AI's overall cipled methodology to estimate AI's K.C. A review of operational water consumption
water footprint and to foster corpo- water footprint. Then, using GPT-3 as and withdrawal factors for electricity generating
technologies. NREL Tech. Report: NREL/TP-
rate water stewardship. an example, we show that a large AI 6A20-50900 (2011).
'When' and 'where' matter. Judi- model can consume millions of liters 14. Meta. Sustainability: Water (2023); https://tinyurl.
com/2c65p7u3.
ciously deciding when and where of water for training. We also discuss 15. Meta. Sustainability Report (2024); https://tinyurl.
com/2yagb6vk.
to train a large AI model can sig- that the scope-1 and scope-2 water 16. Microsoft. Microsoft in Your Community; https://
nificantly affect the water footprint. efficiencies vary spatially and tem- tinyurl.com/257z85pr.
17. Microsoft. Microsoft's Sustainability Targets (2023);
Water efficiency exhibits a spatial- porally—judiciously deciding when https://tinyurl.com/2cgo3aff.
temporal diversity—onsite water and where to run a large AI model can 18. Microsoft. Environmental Sustainability Report (2024);
https://tinyurl.com/2eum2dhg.
efficiency changes due to variations significantly cut the water footprint. 19. Naden, C. AI for sustainability. Intern.
of outside weather conditions, and In addition, we recommend increased Electrotechnical Commission e-tech (2023); https://
tinyurl.com/2bj6m6dz.
offsite water efficiency changes due transparency and comprehensive 20. Patterson, D. et al. Carbon Emissions and Large
to variations of the grid's energy fuel reporting of AI's water footprint, and Neural Network Training (2021); arXiv:2104.10350
[cs.LG]
mixes to meet time-varying demands highlight the necessity of holistically 21. PepsiCo. ESG - Water (2023); https://tinyurl.
(Figure 2). Therefore, we can dy- addressing water footprint along com/2d5qfzuk.
22. Reig, P. What's the difference between water use
namically schedule AI training and with carbon footprint to enable truly and water consumption? World Resources Institute
inference in a water-wise manner to sustainable AI. Commentary (2013); https://tinyurl.com/2n9uqrtz.
23. Reig, P., Luo, T., Christensen, E., and Sinistore, J.
cut the water footprint. For example, AI's water footprint can no longer Guidance for calculating water use embedded in
we may schedule AI training at stay under the radar and must be purchased electricity. World Resources Institute
(2020).
midnight and/or in a datacenter with addressed as a priority as part of the 24. Schwartz, R., Dodge, J., Smith, N.A., and Etzioni, O.
Green AI. Commun. ACM 63, 12 (Nov. 2020), 54–63;
better water efficiency. Likewise, collective efforts to combat global 10.1145/3381831
if informed of the real-time water water challenges. 25. Shehabi, A., et al. 2024 United States Data Center
Energy Usage Report. Lawrence Berkeley National
efficiency, some water-conscious Laboratory LBNL-2001637 (Dec. 2024).
users may prefer to use AI inference Acknowledgment 26. Shumba, N. et al. A water efficiency dataset for
African data centers. In Proceedings of NeurIPS
during water-efficient hours and/or The work of Pengfei Li and Shaolei Workshop on Tackling Climate Change with Machine
in water-efficient datacenters, which Ren was supported in part by NSF Learning (2024).
27. Siddik, Md. A.B., Shehabi, A., and Marston, L. The
can reduce AI's water footprint by CCF-2324916. The work of Moham- environmental footprint of data centers in the United
enabling demand-side flexibility. mad A. Islam was supported in part States. Environmental Research Letters 16, 6 (2021),
064017.
'Follow the Sun' or 'Unfollow the by NSF ECCS-2152357 and CCF- 28. Singapore Public Utilities Board. Sectoral Water
Sun.' To cut the carbon footprint, 2324915. Efficiency Benchmark; https://tinyurl.com/2935uhn8
29. Solomon, S. Sustainable by design: Next-generation
it is preferable to “follow the sun” datacenters Consume zero water for cooling (2024);
when solar energy is more abun- https://tinyurl.com/2xmxxxnz
References 30. Stojkovic, J. et al. DynamoLLM: Designing LLM
dant. Nonetheless, to cut the water 1. Ahmed, K., Islam, M.A., Ren, S., and Quan, G. inference clusters for performance and energy
Exploiting temporal diversity of water efficiency to efficiency. In Proceedings of the IEEE Intern. Symp.
footprint, it may be more appealing make data center less “thirsty”. In Proceedings of the on High-Performance Computer Architecture (2025).
to “unfollow the sun” to avoid high- 11th Intern. Conf of Autonomic Computing (2014). 31. Strubell, E., Ganesh, A., and McCallum, A. Energy
2. Apple. Environmental Responsibility Report. 2024; and policy considerations for deep learning in NLP.
temperature hours of a day when https://tinyurl.com/jwg93lu. In Proceedings of the 57th Ann. Meeting of the Assoc.
WUE is high. This conflict can also 3. Banse, A., Rince, S., and Defour, V. EcoLogits for Computational Linguistics (2019), 3645–3650;
Calculator (2024); https://tinyurl.com/2au2xaml. 10.18653/v1/P19-1355
be shown in Figure 2a and 2b, where 4. Brown, T. et al. Language models are few-shot 32. U.S. Energy Information Administration. Open Data;
learners. In Advances in Neural Information
we see misalignment between the Processing Systems 33, LarochelleH. Curran
https://tinyurl.com/2c5qq6wk.
33. U.S. EPA. eGRID Data Explorer; https://tinyurl.com/
scope-2 water consumption intensity Associates, Inc. (2020), 1877-1901; https://tinyurl. yn72page.
com/yqlf8fzk
factor and carbon emission rate: 5. Cohen, E. and Ramaswami, A. The water withdrawal
Minimizing one footprint might footprint of energy supply to cities. J. of Industrial
Pengfei Li is a Ph.D. candidate at the University of
Ecology 18, 1 (2014), 26-39; https://tinyurl.
increase the other footprint. This com/24ffjzox California, Riverside, CA, USA.
observation further corroborates the 6. de Vries, A. The growing energy footprint of artificial Mohammad A. Islam is an assistant professor at the
intelligence. Joule (Oct. 2023); https://tinyurl. University of Texas, Arlington, TX, USA.
previous finding that the environ- com/2cmwv5m3
Jianyi Yang is an assistant professor at the University of
mental impacts of carbon and water 7. Frost, K. and Hua, I. Quantifying spatiotemporal
Houston, TX, USA.
impacts of the interaction of water scarcity and water
footprints are not substitutable.10,25 use by the global semiconductor manufacturing Shaolei Ren is an associate professor at the University of
Therefore, to judiciously achieve a industry. Water Resources and Industry 22 (2019); California, Riverside, CA, USA.
https://tinyurl.com/247zadpz
balance between “follow the sun” 8. Gao, P.X., Curtis, A.R., Wong, B., and Keshav, S. It’s
for carbon efficiency and “unfollow not easy being green. SIGCOMM Comput. Commun.
Rev. (2012).
the sun” for water efficiency, we need 9. Google. Environmental Report. 2024; https://
to reconcile the potential water- sustainability.google/reports/.
10. Islam, M.A. et al. Exploiting spatio-temporal diversity
This work is licensed under a
carbon conflicts by using holistic for water saving in geo-distributed data centers.
Creative Commons Attribution-NoDerivs
IEEE Transactions on Cloud Computing 6, 3 (2018),
approaches that are carbon-efficient 734-746;https://tinyurl.com/2c43smrv
International 4.0 License.
and water-wise. 11. Karimi, Leila et al. Water-energy tradeoffs in © 2025 Copyright held by the owner/author(s).

JU LY 2 0 2 5 | VO L. 6 8 | N O. 7 | C OM M U N IC AT ION S OF T HE ACM 61