The first 5 pages, shown as text so search engines and readers can find this document.
Page 1
Making AI Less “Thirsty”: Uncovering and Addressing the
Secret Water Footprint of AI Models
Pengfei Li
UC Riverside
Jianyi Yang
UC Riverside
Mohammad A. Islam
UT Arlington
Shaolei Ren1
UC Riverside
Abstract
The growing carbon footprint of artificial intelligence (AI) has been undergoing public scrutiny. Nonethe-
less, 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 King-
dom. 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 ad-
dressing 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.
1 Introduction
Artificial intelligence (AI) has enabled remarkable breakthroughs in numerous areas of critical importance,
including tackling global challenges such as climate change. On the other hand, many AI models, especially
large generative ones like GPT-4, are trained and deployed on energy-hungry servers in warehouse-scale
data centers, accelerating the data center energy consumption at an unprecedented rate [1]. As a result,
AI’s carbon footprint has been undergoing scrutiny, driving the recent progress in AI carbon efficiency [2,3].
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. If not properly
addressed, AI’s water footprint can potentially become a major roadblock to sustainability and create social
conflicts as freshwater resources suitable for human use are extremely limited and unevenly distributed.
As acknowledged in Google’s sustainability report [4] and the recent U.S. data center energy report [1],
the expansion of AI products and services is a key driver of the rapid increase in data center water consump-
tion. Even excluding the water usage in leased third-party colocation facilities, one technology company’s
self-owned data centers alone directly withdrew 29 billion liters and consumed (i.e., evaporated) more than
23 billion liters of freshwater for on-site cooling in 2023, nearly 80% of which was potable water [4].2 This
amount of annual water consumption even rivals that of a major household-name beverage company [5].
Importantly, the company’s data center water consumption increased by ∼20% from 2021 to 2022 and by
∼17% from 2022 to 2023 [4], and another technology company’s data center water consumption saw ∼34%
and ∼22% increases over the same periods, respectively [6]. Furthermore, according to the recent U.S. data
center energy report, the total annual on-site water consumption by U.S. data centers in 2028 could double
or even quadruple the 2023 level, reaching approximately 150 – 280 billion liters and further stressing the
water infrastructures [1].
AI represents the fastest expanding workloads in data centers [1, 4]. For example, a recent study sug-
gests that the global AI could consume 85 – 134 TWh of electricity in 2027 [7], whereas a more aggressive
projection by the recent U.S. data center energy report predicts that AI servers’ electricity consumption in
the U.S. alone will surpass 150 – 300 TWh in 2028 [1]. Even considering the lower estimate, the combined
scope-1 and scope-2 water withdrawal of global AI is projected to reach 4.2 – 6.6 billion cubic meters in
1Corresponding author: Shaolei Ren (shaolei@ucr.edu), University of California, Riverside.
2The detailed difference between water withdrawal and water consumption is presented in Section 2.1.
1
arXiv:2304.03271v5 [cs.LG] 26 Mar 2025
Page 2
2027, which is more than the total annual water withdrawal of 4 – 6 Denmark or half of the United King-
dom.3 Simultaneously, a total of 0.38 – 0.60 billion cubic meters of water will be evaporated and considered
“consumption” due to the global AI demand in 2027. Moreover, these global estimates will be exceeded by
the total water withdrawal and consumption attributed to AI in the U.S. alone in 2028 if the projection in [1]
comes to fruition.
Despite its profound environmental and societal impact, the increasing water footprint of AI has received
disproportionately less attention from the AI community as well as the general public. For example, while
the scope-2 carbon emissions are routinely included as part of AI model cards, even scope-1 direct water
usage (either withdrawal or consumption) is missing, let alone scope-2 water usage. This may impede inno-
vations to enable water sustainability and build truly sustainable AI. Crucially, water and carbon footprints
are complementary to, not substitutable of, each other for understanding the environmental impacts. In-
deed, 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 [1, 9].
To ensure that the growth in AI does not exacerbate the global water stresses or outweigh the environ-
mental benefits it provides, it is a critical time to uncover and address AI’s hidden water footprint amid the
increasingly severe freshwater scarcity crisis, worsened extended droughts, and quickly aging public water
infrastructure. The urgency can also be reflected in part by the recent commitment to “Water Positive by
2030” from industry leaders, including Google [4] and Microsoft [6], and by the inclusion of water foot-
print as a key metric into the world’s first international standard on sustainable AI to be published by the
ISO/IEC [10].
In this paper, we advocate for a holistic approach to sustainable AI that extends beyond the carbon foot-
print to also address the water footprint. Specifically, we present a principled methodology to estimate AI’s
total water footprint, including both operational water and embodied water. By taking the GPT-3 model
with 175 billion parameters as an example [11], we show that training GPT-3 in Microsoft’s U.S. data cen-
ters can consume a total of 5.4 million liters of water, including 700,000 liters of scope-1 on-site water con-
sumption. Additionally, GPT-3 needs to “drink” (i.e., consume) a 500ml bottle of water for roughly 10 – 50
medium-length responses, depending on when and where it is deployed.
Next, we show that WUE (Water Usage Effectiveness, a measure of water efficiency) varies both spatially
and temporally, suggesting that judiciously deciding “when” and “where” to train a large AI model can
significantly cut the water footprint. We also emphasize the need for increasing transparency of AI’ water
footprint, including disclosing more information about operational data and keeping users informed of the
runtime water efficiency. Finally, we highlight the necessity of holistically addressing water footprint along
with carbon footprint to enable truly sustainable AI — the water footprint of AI can no longer stay under the
radar.
2 Background
2.1 Water Withdrawal vs. Water Consumption
There are two related but different concepts — water withdrawal and water consumption, both of which are
important for understanding the impacts on water stress and availability [12, 13].
• Water withdrawal: It refers to freshwater taken from the ground or surface water sources, either tem-
porarily or permanently, and then used for agricultural, industrial, or municipal uses (normally excluding
water used for hydroelectricity generation) [12]. As water is a finite shared resource, water withdrawal
indicates the level of competition as well as dependence on water resources among different sectors.
• Water consumption: It is defined as “water withdrawal minus water discharge”, and means the amount
of water “evaporated, transpired, incorporated into products or crops, or otherwise removed from the im-
mediate water environment” [13]. Water consumption reflects the impact on downstream water availability
and is crucial for assessing watershed-level scarcity [12].
These two types of water usage correspond to two different water footprints, i.e., water withdrawal foot-
3The scope definition of water usage [8] is in line with that of carbon emissions and is discussed in Section 2.2. Our scope-2 water
withdrawal (and consumption when applicable) is for location-based electricity generation throughout the paper. Large data centers
often adopt sustainability programs (e.g., renewable purchasing agreements) to offset their location-based electricity usage and thus
may have lower market-based carbon and water footprints.
2
Page 3
print (WWF) [8,14] and water consumption footprint (WCF), respectively [15]. By default, water footprint
refers to the water consumption footprint unless otherwise specified.
2.2 How Does AI Use Water?
AI’s water usage spans three scopes: on-site water for data center cooling (scope 1), off-site water for elec-
tricity generation (scope 2), and supply-chain water for server manufacturing (scope 3).
2.2.1 Scope-1 Water Usage
Nearly all the server energy is converted into heat, which must then be removed from the data center server
room to avoid overheating. This process involves two sequential stages: server-level cooling followed by
facility-level cooling.
In the server-level cooling stage, heat is transferred from the servers to the facility or a heat exchanger,
typically using either air or liquid cooling methods (e.g., direct-to-chip cooling or immersion cooling),
which do not evaporate or consume water. In general, new data centers dedicated to AI training often
rely on liquid cooling due to the high server power densities.
In the facility-level cooling stage, heat is rejected from the data center facility to the outside environment.
While there are various cooling methods, water-intensive cooling towers and water evaporation-assisted
air cooling are two common approaches used in many data centers, including those operated by major
technology companies [1, 4].Data Center
Chilled
Water
Warm
Water
Pump
Cooling
Tower
Server Rack
CRAHWater
Source
Cooling
Tower
Heat
Exchanger
ChatGPT AlphaGO
Power
Plant
Scope-1 Water
Scope-2 Water
Figure 1: An example of data center’s operational water us-
age: on-site scope-1 water usage for data center cooling (via
cooling towers in the example), and off-site scope-2 water
usage for electricity generation. The icons for AI models
are only for illustration purposes.
Cooling tower. As illustrated in Figure 1, some
water is evaporated (i.e., “consumed”) in the cool-
ing tower to dissipate heat into the environment,
while the remaining water moves along an open
loop to the heat exchanger to further absorb the
server heat. Additionally, non-evaporated water
can be recycled only a few times (typically 3–10 cy-
cles, depending on water quality) before discharge,
requiring continuous clean freshwater replenish-
ment to prevent mineral and salt buildup. Thus, to
keep the cooling tower working, new water must be
constantly added to make up for the evaporated wa-
ter and discharged water. Importantly, clean fresh-
water (potable water in many cases [4]) is needed
to avoid pipe clogs and/or bacterial growth.
For cooling towers, water withdrawal refers to
the amount of added water, including both evapo-
rated water and discharged water, while water consumption exclusively indicates the amount of evaporated
water. With good water quality, roughly 80% of water withdrawal is evaporated and considered “consump-
tion” [4]. On average, depending on the weather conditions and operational settings, data centers can evap-
orate approximately 1 – 9 liters per kWh of server energy: 1 L/kWh for Google’s annualized global on-site
water efficiency [4] and 9 L/kWh for a large commercial data center during the summer in Arizona [16].
Air cooling with water evaporation assistance. When the climate condition is appropriate, data centers
may use “free” outside air to directly reject the heat to the outside environment. Nonetheless, water evapo-
ration is still needed when the outside air is too hot (e.g., higher than 85 degrees Fahrenheit); additionally,
water is also needed for humidity control when the outside air is too dry [17]. The added water is consid-
ered “withdrawal”, out of which about 70% is consumed based on Meta’s report [18]. Generally, outside air
cooling is more water-efficient than cooling towers on average. However, hot weather raises the evaporative
water demand and maximum water consumption, potentially stressing local water supplies during peak
demand on hot days. Additionally, the application of outside air cooling may have challenges in hot regions
and/or for many colocation facilities that are located in business districts.
Some data centers may opt for dry coolers, which consume no on-site water year-round [19]. However,
this approach typically increases cooling energy consumption compared to water-based cooling methods,
potentially exacerbating the overall stress on water resources due to higher scope-2 water consumption.
3
Page 4
2.2.2 Scope-2 Water Usage
In many countries, thermoelectric power is among the top sectors in terms of water withdrawal and water
consumption [8]. Thus, similarly to scope-2 carbon emissions, data centers are accountable for off-site scope-
2 water usage associated with electricity consumption, which forms part of the “true water cost of data
centers,” as highlighted by the recent U.S. data center energy report [1].
Different power plants use different amounts of water for each kWh generation, depending on the cooling
techniques. Typically, water withdrawal due to hydropower generation is excluded, but water consumption
due to increased water evaporation rates from hydropower generation is included [1]. For electricity gener-
ation, the U.S. national average water withdrawal and consumption are estimated at about 43.8 L/kWh [20]
and 3.1 L/kWh [8], respectively. Meta’s self-reported scope-2 water consumption for its global data center
fleet was 3.7 L/kWh (i.e., 55,475 megaliters divided by 14,975,435 MWh) in 2023 [18].
2.2.3 Scope-3 Water Usage
AI chip and server manufacturing uses a huge amount of water [21, 22]. For example, ultrapure water is
needed for wafer fabrication and water is also needed for keeping semiconductor plants cool. Importantly,
the discharged water may contain toxic chemicals and/or hazardous wastes. While water recycling at semi-
conductor plants can effectively reduce water withdrawal, the recycling rate in many cases remains low, e.g.,
the average recycling rate for wafer plants and semiconductor plants in Singapore are 45% and 23%, respec-
tively [22]. Although largely obscure, scope-3 water usage is likely significant [21]. For instance, Apple
reports that its supply chain accounts for 99% of its total water footprint [23].
It is important to recognize that, unlike agriculture whose water footprint is mostly green (i.e., water
stored in soil and used by plants), 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.
3 Estimating AI’s Water Footprint
We present a general methodology for estimating AI’s water consumption footprint. To obtain the water
withdrawal footprint, we simply replace the WUE with water withdrawal efficiency.
3.1 Operational Water Footprint
We collectively refer to on-site scope-1 water and off-site scope-2 water as the operational water.
• On-site WUE. We denote the on-site scope-1 WUE at time t by ρs1,t, which is defined as the ratio of
the on-site water consumption to server energy consumption and varies over time depending on the outside
temperature (see [9] for an example of on-site WUE based on cooling towers). Concretely, ρs1,t 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.
• Off-site WUE. We denote the off-site scope-2 WUE at time t as ρs2,t, which is defined as the ratio
of off-site water consumption for each kWh of electricity consumption and measures the electricity water
intensity factor (EWIF). While there are different methods to estimate ρs2,t, a common one is weighted
averaging: ρs2,t =
P
k bk,t×EW IFkP
k bk,t where bk,t denotes the amount of electricity generated from fuel type k
at time t for the grid serving the data center under consideration, and EW IFk is the EWIF for fuel type
k [24, 25]. Thus, variations in energy fuel mixes of electricity generation result in temporal variations of the
off-site WUE. Moreover, the off-site WUE also varies across regions due to different energy fuel mixes [1,8].
• Operational water footprint. Consider a time-slotted model t = 1, 2, · · · , T , where the length of each
time slot depends on how frequently we want to assess the operational water footprint. At time t, suppose
that an AI model uses energy et which can be measured using power meters and/or servers’ built-in tools,
and the data center hosting the AI model has a power usage effectiveness (PUE) of θt that accounts for
the non-IT energy overhead. Then, the total operational water footprint of the AI model can be written as
W aterOperational = PT
t=1 et · [ρs1,t + θt · ρs2,t].
3.2 Embodied Water Footprint
Similar to accounting for the embodied carbon footprint [26], the total scope-3 water footprint is amortized
over the lifespan of a server. Specifically, if W represents the total water used to manufacture the AI servers
4
Page 5
Table 1: Estimate of GPT-3’s operational water consumption footprint. “*” denotes data centers under construction
as of July 2023, whose PUE and WUE are projected by Microsoft.
Location PUE
On-site
WUE
(L/kWh)
Off-site
EWIF
(L/kWh)
Water for Training (million L) Water for Each Request (mL) # of Requests
for 500ml
Water
On-site
Water
Off-site
Water
Total
Water
On-site
Water
Off-site
Water
Total
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 servers are expected to operate for a period of T0, then the embodied water footprint over a period
of T is calculated as W aterEmbodied = T ·W
T0
By adding up the operational and embodied water footprints, we can obtain the total water footprint
as W aterT otal = PT
t=1 et · [ρs1,t + θt · ρs2,t] + T ·W
T0 . In practice, to obtain a rough estimate, we can use the
average values for the annualized WUE and the estimated AI server energy consumption.
3.3 Case Study: Estimating GPT-3’s Operational Water Consumption Footprint
The core of ChatGPT, a popular online service, is a large language model (LLM) based on subsequent ver-
sions of GPT-3. We present a case study to estimate the operational water consumption for the full GPT-3
model with 175 billion parameters [11]. We exclude embodied water footprint due to the lack of public data
for scope-3 water usage. We choose GPT-3 as Microsoft publishes its location-wise WUE and PUE [27, 28].
The results are summarized in Table 1.
3.3.1 Training
GPT-3 was trained and deployed by OpenAI in Microsoft’s data centers, with an estimated training energy
of 1287 MWh [29]. In line with the practice of estimating the carbon footprint, we use the most recent
annualized average on-site PUE and WUE for each location, as reported by Microsoft [27, 28]. For power
plant water efficiency, different references may provide different estimates of EWIF. Thus, for consistency
across regions, we use the EWIF provided by [8] to estimate scope-2 water consumption, as it employs the
same methodology for calculating EWIF. Moreover, a large number of Microsoft’s data centers are located
in the U.S., where the average EWIF provided by [8] is 3.14 L/kWh and significantly lower than 4.35 L/kWh
reported by the recent study [1]. The specific location for training GPT-3 is not public. Thus, we consider
Microsoft’s different data center locations, while excluding Singapore and Taiwan as EWIF data for these
regions is not available in [8].
3.3.2 Inference
As a representative usage scenario for an LLM, we consider a conversation task, which typically includes
a CPU-intensive prompt phase that processes the user’s input (a.k.a., prompt) and a memory-intensive
token phase that produces outputs [30]. More specifically, we consider a medium-sized request, each with
approximately ≤800 words of input and 150 – 300 words of output [30]. 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 [11]. While no details are provided, the estimate likely considers only the GPU energy
used during token generation.
To account for both the prompt phase and the non-GPU energy consumption of servers, we assume a
5