Content uploaded by Mohammad Atiqul Islam
Author content
All content in this area was uploaded by Mohammad Atiqul Islam on Oct 24, 2024
Content may be subject to copyright.
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) models, especially large ones such as GPT-3,
has been undergoing public scrutiny. Unfortunately, however, the equally important and enormous water
(withdrawal and consumption) footprint of AI models has remained under the radar. For example, training
GPT-3 in Microsoft’s state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwa-
ter, but such information has been kept a secret. More critically, the global AI demand may be accountable
for 4.2 – 6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water with-
drawal of 4 – 6 Denmark or half of the United Kingdom. This is very concerning, as freshwater scarcity has
become one of the most pressing challenges shared by all of us in the wake of the rapidly growing popula-
tion, depleting water resources, and aging water infrastructures. To respond to the global water challenges,
AI models can, and also must, take social responsibility and lead by example by addressing their own water
footprint. In this paper, we provide a principled methodology to estimate the water footprint of AI models,
and also discuss the unique spatial-temporal diversities of AI models’ runtime water efficiency. Finally, we
highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly
sustainable AI.
1 Introduction
•“Water is a finite resource, and every drop matters.” — Facebook (now Meta) Sustainability Report 2020 [1].
•“Fresh, clean water is one of the most precious resources on Earth ... Now we’re taking urgent action to
support water security and healthy ecosystems.” — Google’s Water Commitment 2023 [2].
•“Water is a human right and the common development denominator to shape a better future. But water
is in deep trouble.” — U.N. Secretary-General Ant´
onio Guterres at the U.N. Water Conference 2023 [3].
•“Historic droughts threaten our supply of water ... As the source of both life and livelihoods, water
security is central to human and national security.” — U.S. White House Action Plan on Global Water Security
2022 [4].
Artificial intelligence (AI) models have witnessed remarkable breakthroughs and success in numerous
areas of critical importance to our society over the last decade, including in the ongoing combat against
several global challenges such as climate change [5]. Increasingly many AI models are trained and deployed
on power-hungry servers housed inside warehouse-scale data centers, which are often known as energy
hogs [6]. Consequently, despite the numerous benefits and potential of AI, the environmental footprint of
AI models, in particular carbon footprint, has been undergoing public scrutiny, driving the recent progress
in AI carbon efficiency [7–11]. Unfortunately, however, the enormous water footprint of AI models — many
millions of liters of freshwater withdrawn or consumed for electricity generation and for cooling the servers
— has largely remained under the radar. If not properly addressed, the increasing water usage can become
a potential major roadblock to the socially responsible and environmentally sustainable AI in the future.
Despite the water cycle through our planet’s natural ecosystem, clean freshwater resource available and
suitable for use is extremely limited and unevenly distributed across the globe. In fact, freshwater scarcity
is one of the most pressing challenges shared by all of us in the wake of the rapidly growing population
and extended megadroughts [12, 13]. Severe water scarcity has already been affecting 4 billion people,
or approximately two-thirds of the global population, for at least one month each year [13, 14]. Without
integrated and inclusive approaches to addressing the global water challenge, nearly half of the world’s
1Corresponding author: Shaolei Ren (shaolei@ucr.edu), University of California, Riverside.
1
arXiv:2304.03271v3 [cs.LG] 29 Oct 2023
population will endure severe water stress by 2030 [4], and roughly one in every four children worldwide
will be living in areas subject to extremely high water stress by 2040 [14].
Warehouse-scale data centers — physical “homes” where many AI models, especially large ones like
GPT-3 and GPT-4 for language services, are physically trained and deployed — are known to be energy-
intensive, collectively accounting for about 1-2% of the global electricity usage [6]. As such, it is well-known
that data centers are responsible for a significant scope-2 carbon footprint associated with location-based
electricity generation [7, 15–17]. Nonetheless, what is much less known is that data centers are also ex-
tremely “thirsty” — even excluding the embodied water usage due to supply chains (e.g., scope-3 water for
chip manufacturing), data centers use an enormous amount of water for both on-site cooling and off-site
electricity generation [18,19], which, similar to the scope definitions for carbon emissions, are referred to as
scope-1 and scope-2 water usage, respectively [20].
Even putting aside the water usage in leased third-party colocation facilities, Google’s self-owned data
centers alone directly withdrew 25 billion liters and consumed nearly 20 billion liters of scope-1 water for on-
site cooling in 2022, the majority of which was potable water [21].2Overall, Google’s data center water usage
(both withdrawal and consumption) in 2022 increased by 20% compared to 2021 [21,22], and Microsoft’s
total water usage even saw a 34% increase for the same period [23]. Such significant increases are likely
attributed in part to the growing demand for AI.
In addition, the combined on-site scope-1 and off-site scope-2 global water withdrawal of Google, Mi-
crosoft, and Meta reached an estimate of 2.2 billion cubic meters in 2022, equivalent to the total annual
water withdrawal (including municipal, industrial, and agricultural usage) of two Denmark [24].3This
includes 1.5 billion cubic meters of water withdrawal in the U.S., accounting for about 0.33% of the total
U.S. annual water withdrawal. Simultaneously, out of the total global water withdrawal, approximately
0.18 billion cubic meters (including 0.13 billion cubic meters in the U.S.) was “lost” due to evaporation and
hence considered “consumption”, which was even more than the total annual water withdrawal of Liberia
(a country of about 5 million people in West Africa) [24]. Note that [24] does not provide country-wide
water consumption data. As a rule of thumb, the ratio of water consumption to water withdrawal varies
from 5 to 15% in urban areas and from 10 to 50% in rural areas [26]. The growing tension over the enormous
water usage between data centers and human needs may create new environmental risks and even social
conflicts (e.g., the protest against Google’s planned data center construction in Uruguay amid its extended
drought [27]).
AI represents one of the most prominent and fastest expanding workloads in data centers [7,28,29]. For
example, a recent study suggests that the global AI demand might consume 85–134 TWh of electricity in
2027 [30]. If this estimate materializes, the combined scope-1 and scope-2 operational water withdrawal
of global AI may reach 4.2 – 6.6 billion cubic meters in 2027, which is more than the total annual water
withdrawal of 4 – 6 Denmark or half of the United Kingdom that is currently under the threat of droughts
[31]. If the U.S. hosts half of the global AI workloads, the operation of AI may take up about 0.5 – 0.7% of
its total annual water withdrawal. Additionally, the total scope-1 and scope-2 water consumption of global
AI may exceed 0.38 – 0.60 billion cubic meters, i.e., roughly evaporating the annual water withdrawal of half
of Denmark or 2.5 – 3.5 Liberia. Therefore, AI models can, and also must, take social responsibility and lead
by example in the collective efforts to combat the global water scarcity challenge by cutting their own water
footprint.
Despite its profound environmental and societal impact, however, the enormous water footprint of AI
models 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 [32],
even scope-1 water usage (either withdrawal or consumption) is missing, let alone scope-2 water usage.
This may impede innovations to enable water sustainability and build truly sustainable AI. Importantly,
water and carbon footprints are complementary to, not substitutable of, each other for understanding the
environmental impacts [33]. Indeed, optimizing for carbon efficiency does not necessarily result in, and
2The detailed difference between water withdrawal and water consumption is presented in Section 2.1.
3As data centers predominantly rely on the electric grid instead of being directly powered by renewables [23,25], our scope-2 water
withdrawal (and consumption if applicable) is for location-based electricity generation throughout the paper. The detailed calculation
method is available in the appendix. Nonetheless, Google, Microsoft, and Meta often adopt alternative sustainability programs (e.g.,
renewable purchasing agreements) to offset their location-based electricity usage and thus have lower or zero market-based carbon and
water footprints. The current ESG reports typically include both location-based and market-based carbon emissions.
2
may even worsen, water efficiency, which varies with the energy fuel mixes for electricity generation and
outside weather conditions in its own unique way [34].
On the other hand, unlike many other workloads, the highly flexible nature of AI workloads opens up
novel scheduling opportunities to minimize the water footprint of AI. (1) Spatial flexibility: Both AI model
training and inference can be processed in almost any data center with little impact on latency due to recent
advances in data center networking [35]. (2) Temporal flexibility: AI models can be trained intermittently
by a certain deadline. (3) Performance flexibility: For the same inference service, a set of heterogeneous AI
models with distinct computing resource consumption and accuracy performance are often available via
model pruning and compression (e.g., GPT-3 has eight sizes, ranging from 125 million parameters to 175
billion parameters [36]).
Therefore, it is truly a critical time to uncover and address AI models’ secret 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” by increasingly many companies, including Google [22], Microsoft [37] and Meta [38].
In this paper, we make the first-of-its-kind efforts to uncover the secret water footprint of AI models.
Specifically, we present a principled methodology to estimate the total water (both withdrawal and con-
sumption) footprint, including both operational water and embodied water. By taking the GPT-3 model
(with 175 billion parameters) for language services as a concrete example [36], we show that training GPT-
3 in Microsoft’s state-of-the-art U.S. data centers can consume a total of 5.4 million liters of water, includ-
ing 700,000 liters of scope-1 on-site water consumption. Additionally, GPT-3 needs to “drink” (i.e., con-
sume) a 500ml bottle of water for roughly 10-50 responses, depending on when and where it is deployed.
These numbers may increase for the newly-launched GPT-4 that reportedly has a substantially larger model
size [39].
Next, we show that WUE (Water Usage Effectiveness, a measure of water efficiency) varies both spa-
tially and temporally, implying that judiciously deciding “when” and “where” to train a large AI model can
significantly cut the water footprint. We also point out the need for increasing transparency of AI models’
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 models can no longer stay
under the radar.
2 Background
2.1 Water Withdrawal vs. Water Consumption
There are two related but different types of water usage — water withdrawal (a.k.a. water abstraction) and
water consumption, both of which are important for holistically understanding the impacts on water stress
and availability [40,41].
•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) [42]. As water is a finite shared resource, water withdrawal
indicates the level of competition as well as dependence on water resources among different sectors. In an
emergency, a country may only have enough water withdrawal for 48 hours, and the change in water quality
after withdrawal contributes to water stress levels for downstream usage [40].
•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” [41]. Water consumption reflects the impact of water use on downstream
water availability and is crucial for evaluating watershed-level water scarcity [40].
These two types of water usage also correspond to two different types of water footprints, i.e., water
withdrawal footprint (WWF) [43] and water consumption footprint (WCF), respectively [6,19]. By default,
water footprint refers to the water consumption footprint unless otherwise specified.
3
2.2 How Does AI Use Water?
Following the scope definition for carbon emissions [25], we describe AI’s water usage according to on-site
water for server cooling (scope 1), off-site water for electricity generation (scope 2), and supply-chain water
for server manufacturing (scope 3).
2.2.1 Scope-1 Water Usage
Computing servers, especially those equipped with multiple graphic processing units (GPUs) hosting AI
workloads, are energy-intensive. Nearly all the server energy is converted into heat, which must then be
removed from the data center server room to avoid overheating. There are two basic types of cooling systems
— cooling towers and outside air cooling — which both use water and we describe as follows.
Data Center
Chilled
Water
Warm
Water
Pump
Cooling
Tower
Server Rack
CRAH
Water
Source
Cooling
Tower
Chiller
Heat
Exchanger
ChatGPT AlphaGO
Power
Plant
On-site Water
Off-site Water
Figure 1: An example of data center’s operational water
usage: on-site scope-1 water for server 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 towers. Many data centers, includ-
ing some of Google’s data centers [22], use cool-
ing towers as the heat rejection mechanism. While
the specific designs may differ, a common system
contains two water loops as illustrated in Figure 1:
one closed loop between the chiller and data center
server room, and one open loop between the cool-
ing tower and the chiller. The closed loop does not
withdraw or consume water — the water circulat-
ing inside is pumped from the chiller into the data
center to cool down the air handling unit’s supply
air in order to maintain a proper server inlet temper-
ature, and the warm water that absorbs heat from
the hot air returns to the chiller. In other words, the
closed-loop water simply transfers the heat from the
hot air exiting the servers’ outlets to the chiller unit,
which then must be cooled by further rejecting the
heat into the outside environment. The chiller may be turned off and operate in a natural “bypass” mode
for energy saving when the outside temperature is sufficiently low.
While air-cooled chillers are available, a more efficient design is to use an open loop that carries water
to move the heat from the chiller to a cooling tower. Along the open loop, some water gets evaporated (i.e.,
“consumed”) in the cooling tower to dissipate heat into the environment, while the remaining water moves
to the chiller unit to further absorb heat. Additionally, the remaining non-evaporated water in the open loop
can only be cycled up to a few times (e.g., 3 – 10, depending on the water quality) and must be discharged to
avoid high concentrations of salt and minerals. Thus, to keep the cooling tower working, new water must be
constantly added to make up for the evaporated water and discharged water. Importantly, clean freshwater
(potable water in many cases [21]) is needed to avoid pipe clogs and/or bacterial growth.
For cooling towers, water withdrawal refers to the amount of water added to the cooling tower (including
both evaporated water and discharged water), while water consumption exclusively indicates the amount
of evaporated water. Typically, given good water quality (that allows more water cyles before discharging),
roughly 80% of water withdrawal is evaporated and considered “consumption” [21]. On average, depend-
ing on the weather conditions and operational settings, data centers can evaporate about 1 – 9 liters per
kWh of server energy (about 1 L/kWh for Google’s annualized global number [21] and 9 L/kWh for a large
commercial data center during the summer in Arizona [44]).
Outside air cooling with water assistance. When the climate condition is appropriate, data centers may
use “free” outside air to directly cool down the servers without cooling towers. A large amount of outside
air is blown through the servers and then exhausted to the outside. For outside air cooling, however, water
evaporation is still needed when the outside air is too hot (e.g., higher than 81 or 89 degrees Fahrenheit);
additionally, water is needed for humidity control when the outside air is too dry [38, 45, 46]. The added
water is considered withdrawal, out of which about 70% is evaporated based on Meta’s report [25]. Typ-
ically, the average water efficiency of outside air cooling is better than that of cooling towers (e.g., water
consumption of 0.2 L/kWh averaged over Meta’s global data centers) [25]. Nonetheless, when the outside
air temperature is high, outside air cooling evaporates a significant amount of water, thus resulting in a low
4
average but high peak water withdrawal. This may be especially problematic when the demand for water
is also higher for other users on certain hot summer days [47]. Additionally, the application of outside air
cooling may have significant challenges in hot regions and/or for many colocation facilities that are located
in business districts.
Some data centers may also use a hybrid design that dynamically switches between cooling tower and
outside air cooling [48]. Note also that AI servers often have high power densities due to specialized designs,
including multiple GPUs and/or purpose-built hardware, to speed up AI model training and inference. As
such, on-chip liquid cooling may be employed: closed-loop circulating liquid directly moves the heat from
the servers to the data center facility (e.g., the facility’s cooling water loop). Then, the heat moved to the
facility will be rejected by cooling towers or simply outside air [48].
2.2.2 Scope-2 Water Usage
Despite the increasing adoption of solar and wind energy, 73% of the utility-scale electricity generated in the
U.S. came from thermoelectric power plants in 2021 [49]. Thus, electricity generation does not only emit car-
bon, but also uses a huge amount of water [20,41]. Just like scope-2 carbon emissions, data centers, including
AI workloads, are also responsible for off-site scope-2 water usage due to electricity generation [6,19, 44].
Different thermoelectric power plants (e.g., coal and natural gas) use different amounts of water for each
kWh generation. The amount of water withdrawal also depend on the cooling techniques [49]. Typically,
water withdrawal due to hydropower generation is excluded from the calculation of water withdrawal, but
water consumption due to expedited water evaporation from hydropower generation is often included [6].
In many countries, thermoelectric power is among the top sectors (including agriculture) in terms of water
withdrawal and water consumption [41]. For electricity generation, the U.S. national average water with-
drawal and consumption are estimated at about 43.8 L/kWh [49] and 3.1 L/kWh [20], respectively. While
Google and Microsoft did not disclose their scope-2 water usage, Meta’s self-reported average water con-
sumption for its global data center fleet was 3.58 L/kWh (i.e., 41,172,356 cubic meters divided by 11,508,131
MWh) in 2022 [25].
2.2.3 Scope-3 Water Usage
AI chip and server manufacturing uses a huge amount of water [50, 51]. For example, ultrapure water is
needed for wafer fabrication, and clean water is also needed for keeping semiconductor plants cool. Overall,
a large semiconductor plant may withdraw several million liters of water each day [52]. Importantly, the
discharged water can contain toxic chemicals and/or hazardous wastes, which need additional processing
before reused for other purposes. While water recycling at semiconductor plants can effectively reduce water
withdrawal, the water recycling rate in many cases can still remain low, e.g., the average recycling rate for
wafer plants and semiconductor plants in Singapore are 45% and 23%, respectively [51]. Unlike scope-1 and
scope-2 water usage, the data for scope-3 water usage (including withdrawal and consumption) remains
largely obscure.
3 Estimating Water Footprint of AI Models
While an AI model’s water footprint depends in part on its energy consumption, such dependency varies
spatially and temporally. As a result, simply multiplying the AI model’s energy consumption by a constant
and fixed WUE may not yield an accurate estimate of AI models’ water footprint. Next, by accounting for
time-varying WUE, we present a methodology for a fine-grained estimate of an AI model’s water footprint.
Here, we focus on water consumption footprint to describe our water footprint modeling methodology. To
obtain the water withdrawal footprint, we simply replace the WUE with a new coefficient that represents
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. To model
the operational water footprint, we need to know the on-site WUE and off-site WUE.
•On-site WUE. We denote the on-site scope-1 WUE at time tby 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. Concretely, ρs1,t increases significantly for cooling towers when the outside wet bulb
temperature increases [34,53], and increases for outside air cooling when the outside dry bulb temperature is
5
too hot or the humidity is too low [38,46]. For example, the monthly average on-site WUE for a commercial
data center reaches as high as about 9 L/kWh in the summer and becomes about 4 L/kWh in the winter
[44]. Thus, ρs1,t can be modeled as a function in terms of the outside weather condition based on empirical
operational data (see [34] for an example of on-site WUE based on cooling towers).
•Off-site WUE. We denote the off-site scope-2 WUE at time tas ρs2,t , which is defined as the ratio of off-
site water consumption for each kWh of electricty generation and measures the electricity water intensity
factor (EWIF). In practice, ρs2,t depends on the energy fuel mixes (e.g., coal, nuclear, hydro) as well as
cooling techniques used by power plants [20, 54, 55]. Since electricity produced by different energy fuels
becomes non-differentiated once entering the grid, we consider the average method to calculate ρs2,t, which
can be estimated as ρs2,t =Pkbk,t×E W IFk
Pkbk,t where bk,t denotes the amount of electricity generated from fuel
type kat time tfor the grid serving the data center under consideration, and E W IFkis the EWIF for fuel
type k[56,57]. As a result, variations in energy fuel mixes of electricity generation (to meet various demand
levels) result in temporal variations of the off-site WUE. Moreover, the off-site WUE also varies by location,
because each fuel type has its own distinct WUE and the energy fuel mix is typically different between states
as some states may use less water-efficient energy generation than others [18,20, 34, 55].
•Operational water footprint. The on-site scope-1 water consumption can be obtained by multiplying
AI’s energy consumption with the on-site WUE, while the off-site scope-2 water consumption depends on
the electricity usage as well as the local off-site WUE. Consider a time-slotted model t= 1,2,· · · , T , where
each time slot can be 10 minutes to an hour depending on how frequently we want to assess the operational
water footprint, and Tis the total length of interest (e.g., training stage, total inference stage, or a combi-
nation of both). At time t, suppose that an AI model uses energy et(which can be measured using power
meters and/or servers’ built-in tools), the on-site WUE is ρs1,t, the off-site WUE is ρs2,t , and the data center
hosting the AI model has a power usage effectiveness (PUE) of θtthat accounts for the non-IT energy such
as cooling systems and power distribution losses. Then, the total water footprint Wof the AI model can be
written as
W aterOper ational =
T
X
t=1
et·[ρs1,t +θt·ρs2,t].(1)
If we are interested in the operational water withdrawal footprint, we simply replace ρs1,t and ρs2,t with
new values that represent the on-site and off-site water withdrawal efficiencies, respectively.
3.2 Embodied Water Footprint
The embodied water footprint is primarily due to server manufacturing. Like accounting for embodied
carbon footprint [16,58], the total water for manufacturing is amortized over the lifespan of a server. Specif-
ically, suppose that it uses Wamount of water for manufacturing the AI servers in total and the servers are
expected to last an effective period of T0time (i.e., the total lifespan multiplied by the average utilization
rate). Then, for a total length Tof interest (e.g., training stage, total inference stage, or a combination of
both), the embodied water footprint for AI servers is obtained by adding up the amortized per-time manu-
facturing water over the total time T, i.e.,
W aterE mbodied =T·W
T0
.(2)
By adding up the operational and embodied water footprints, we obtain the total water footprint as
W aterT otal =
T
X
t=1
et·[ρs1,t +θt·ρs2,t] + T·W
T0
.(3)
Our methodology for estimating AI models’ water footprint is general and applies to data centers with
any cooling systems. We simply plug in the values ρs1,t,θtand ρs2,t to obtain a fine-grained estimate of
the operational water footprint. Alternatively, to obtain a rough estimate, we can use the (annualized)
average values for these parameters and the estimated AI server energy consumption (e.g., by multiplying
the average GPU power consumption with the total training or inference time) [7].
6
Table 1: Estimate of GPT-3’s average operational water consumption footprint. “*” denotes data centers under
construction as of July 2023, and the PUE and WUE values for these data centers are based on Microsoft’s projection.
Location PUE WUE
(L/kWh)
Electricity Water
Intensity
(L/kWh)
Water for Training (million L) Water for Each Inference (mL) # of Inferences
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
Wyoming 1.125 0.230 2.574 0.296 3.727 4.023 0.920 11.583 12.503 40.0
Iowa 1.160 0.190 3.104 0.245 4.634 4.879 0.760 14.403 15.163 33.0
Arizona 1.223 2.240 4.959 2.883 7.805 10.688 8.960 24.259 33.219 15.1
Washington 1.156 1.090 9.501 1.403 14.136 15.539 4.360 43.934 48.294 10.4
Virginia 1.144 0.170 2.385 0.219 3.511 3.730 0.680 10.913 11.593 43.1
Texas 1.307 1.820 1.287 2.342 2.165 4.507 7.280 6.729 14.009 35.7
Singapore 1.358 2.060 1.199 2.651 2.096 4.747 8.240 6.513 14.753 33.9
Ireland 1.197 0.030 1.476 0.039 2.274 2.313 0.120 7.069 7.189 69.6
Netherlands 1.158 0.080 3.445 0.103 5.134 5.237 0.320 15.956 16.276 30.7
Sweden 1.172 0.160 6.019 0.206 9.079 9.284 0.640 28.216 28.856 17.3
Mexico* 1.120 0.056 5.300 0.072 7.639 7.711 0.224 23.742 23.966 20.9
Georgia* 1.120 0.060 2.309 0.077 3.328 3.406 0.240 10.345 10.585 47.2
Taiwan* 1.200 1.000 2.177 1.287 3.362 4.649 4.000 10.448 14.448 34.6
Australia* 1.120 0.012 4.259 0.015 6.138 6.154 0.048 19.078 19.126 26.1
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
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
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 built based on subsequent versions
of GPT-3. Due to the limited public data available for GPT-4, we now present a case study to estimate GPT-
3’s operational water consumption footprint. In particular, we consider the full GPT-3 model with 175 billion
parameters, which is also the one considered for estimating its carbon footprint [7]. We exclude embodied
water footprint due to the lack of public data for scope-3 water usage in GPT-3’s supply chain. We choose
GPT-3 as Microsoft publishes its location-wise WUE and PUE [46], whereas such information is often limited
for other companies. The results are shown in Table 1. Note that the newer GPT-4 model currently used by
ChatGPT reportedly has a significantly larger model size [39] and hence likely consumes more training and
inference energy than GPT-3 on average.
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 [36, 59]. While Microsoft recently disclosed that OpenAI had used its Iowa data center for
training some models such as GPT-4 [47], the specific data center location for training GPT-3 has yet to
be public. Thus, we estimate the water consumption footprint for training GPT-3 by considering different
data center locations, including non-U.S. data centers for references. In line with the practice of estimating
the carbon footprint [7], we use the annualized average on-site power usage effectiveness (PUE) and water
usage effectiveness (WUE) for each data center location. The PUE and WUE for each data center location
are based on Microsoft’s most recent disclosure [46], while the average U.S. data center PUE and WUE are
based on [60]. For data centers under construction, we use the PUE and WUE data projected by Microsoft.
To estimate the off-site scope-2 water consumption, we use the regional electricity water intensity pro-
vided by [20] wherever applicable to ensure maximum data consistency. Specifically, for each U.S. data
center location, we use the eGRID-level average electricity water intensity [20]. For Microsoft’s Taiwan data
center, we use the average electricity water intensity of 2.177L/kWh provided by [61]. For Microsoft’s Sin-
gapore data center, we calculate the electricity water intensity by considering a mixture of 96% natural gas
and 4% renewables (with zero water operational water consumption) as Singapore’s energy sources for
electricity generation [62] and using Malaysia’s water intensity for natural gas-based electricity generation
provided by [20] as a substitute. For all the other non-U.S. data center locations, we use their country-level
average electricity water intensities from [20].
7
0.1 0.3 0.5 0.7 0.9
Carbon (kg/kWh)
0
4
8
12
16
Water (L/kWh)
AKMS
HIOA
MROE
NWPP
NYUP
AZNM
(a) eGRID-level carbon/water efficiency
MON
TUE
WED
THU
FRI
1.5
1.8
2.1
2.4
2.7
3.0
Water (L/kWh)
0.30
0.32
0.34
0.36
0.38
Carbon (kg/kWh)
(b) Hourly carbon/water efficiency
MON
TUE
WED
THU
FRI
0
20
40
60
Percentage
Coal
Natural Gas
Nuclear
Oil
Wind
Solar
Hydro
Other
(c) Hourly energy fuel mixes
Figure 2: (a) The U.S. eGRID-level scope-2 water consumption intensity factor vs. carbon emission rate [20, 63].
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 5-day snapshot of scope-2 carbon emission rate and water consumption
intensity for electricity generation serving Virginia, starting from April 4, 2022. The values are calculated based on the
energy source mixes, carbon emission rate and water consumption intensity for each energy fuel type [20,63, 64]. The
scope-2 carbon and water efficiencies only have a Pearson correlation coefficient of 0.06, showing a weak correlation.
(c) A 5-day snapshot of energy fuel mixes serving Virginia, starting from April 4, 2022 [64].
3.3.2 Inference
AI models are often deployed globally for inference to minimize the latency and/or comply with privacy
regulations. We estimate the water consumption footprint for GPT-3 inference in different data centers.
The PUE, WUE and electricity water intensity are the same as those used for estimating the training water
footprint. The official estimate shows that GPT-3 consumes on the order of 0.4kWh electricity to generate
100 pages of content (e.g., 0.004kWh per page) [36]. Meanwhile, the average inference energy for BLOOM
(a language model with a slightly larger size of 176 billion parameters than GPT-3) is about 0.00396kWh
per request (914kWh for 230,768 requests), including both dynamic energy and amortized idle energy [16].
BLOOM was deployed on Google cloud, which has a similar energy efficiency as Microsoft’s Azure cloud.
Thus, we consider 0.004kWh as the inference energy consumption per request.
Remarks. As Microsoft only reports its annualized PUE and on-site scope-1 WUE [60], the actual PUE
and WUE at certain times of the year can be different from the annualized numbers. Moreover, if GPT-3 is
deployed in third-party colocation data centers other than Microsoft’s own state-of-the-art data centers, the
water footprint for inference may also be higher due to the often worse PUE and WUE in colocation data
centers. Our electricity water intensity for the U.S. (i.e., 3.14L/kWh on average) is lower than 7.6L/kWh used
by Lawrence Berkeley National Laboratory to estimate the offsite scope-2 water consumption [6]. Therefore,
our estimated water footprint for GPT-3 can absorb some potential discrepancies in the estimated inference
energy of 0.004kWh, while noting that GPT-4 currently used by ChatGPT may use different, and likely more,
energy than GPT-3 on average due to the reportedly larger model size [39].
4 Our Findings and Recommendations
We provide our findings and recommendations to address the water footprint of AI models, making future
AI more socially responsible and environmentally sustainable.
4.1 “When” and “Where” Matter
Judiciously deciding “when” and “where” to train a large AI model can significantly affect the water foot-
print. As shown in Figures 2(a) and 2(b), the water efficiency has spatial-temporal diversity — on-site 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(c)) [44, 65]. In
fact, water efficiency varies at a much faster timescale than monthly or seasonably. Therefore, by exploiting
spatial-temporal diversity of water efficiency, we can dynamically schedule AI model training and inference
to cut the water footprint. For example, if we train a small AI model, we can schedule the training task at
midnight and/or in a data center location with better water efficiency. Likewise, some water-conscious users
may prefer to use the inference services of AI models during water-efficient hours and/or in water-efficient
8
data centers, which can contribute to the reduction of AI’s water footprint.
4.2 More Transparency is Needed
To exploit the spatial-temporal diversity of water efficiency, it is crucial to have better visibility of the run-
time water efficiency and increase transparency by keeping the AI model developers as well as end-users
informed. Nonetheless, such data is often lacking. For example, even scope-1 water usage (either with-
drawal or consumption) is not included in today’s AI model cards (e.g., [32]), not to mention the scope-2
water usage. Additionally, there is very limited data available for embodied water usage by chip making,
which adds challenges to a holistic lifecycle view of AI’s water footprint.
We recommend AI model developers and data center operators be more transparent. For example, what
are the runtime (say, hourly) on-site scope-1 WUE and off-site scope-2 WUE? What about the water footprint
of AI models trained and/or deployed in third-party colocation data centers? Such information will be of
great value to the research community and the general public. As the first step, we recommend that the
scope-1 and scope-2 water usage information be included in AI’s model cards.
4.3 “Follow the Sun” or “Unfollow the Sun”
To cut the carbon footprint, it is preferable to “follow the sun” when solar energy is more abundant. Nonethe-
less, 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. This conflict can be shown in Figure 2(a), where we see that the scope-2
water consumption intensity factor and carbon emission rate are not well aligned: minimizing one footprint
might increase the other footprint. Thus, to judiciously achieve a balance between “follow the sun” for car-
bon efficiency and “unfollow the sun” for water efficiency, we need to reconcile the potential water-carbon
conflicts by using new and holistic approaches. In other words, only focusing on AI models’ carbon footprint
alone may be insufficient to enable truly sustainable AI.
5 Related Works
The growing resource and energy consumption of AI models have placed an increasing emphasis on sus-
tainable AI [7–10, 36, 59, 66]. Specifically, a variety of approaches can be leveraged to make AI more sus-
tainable, including novel GPU and accelerator designs [7, 67, 68], efficient AI model training and infer-
ence [69, 70], carbon-aware AI model scheduling [10,15], green data center designs [71–74]. These studies
have mostly focused on scope-2 carbon footprint, neglecting water footprint which is another environmen-
tal footprint [75–77]. Crucially, despite the correlation between water footprint and carbon footprint, the
existing techniques that optimize for carbon efficiency do not necessarily equate to, and may even worsen,
water efficiency [34].
Data centers have increasingly adopted climate-conscious cooling system designs (e.g., air-side econo-
mizers and purifying non-potable water) [21, 38]. These water-saving approaches can be viewed as sup-
ply-side solutions — saving water while supplying enough cooling to meet the given demand. But, the
demand-side management — cooling demands are affected by “when” and “where” AI models are trained
and used — is not addressed. Additionally, these approaches only focus on the on-site scope-1 water usage,
whereas the off-site scope-2 water usage with time-varying off-site WUE due to variations in energy fuel
mixes is not addressed.
Finally, it is worth mentioning that some data center operators have also begun to build water restoration
projects to indirectly compensate for their water footprint [21,25, 37]. While this is certainly encouraging, it
is is an offsetting method (like building renewables projects to offset location-based carbon emissions) and
hence orthogonal to reducing water withdrawal or consumption in the first place.
6 Conclusion
In this paper, we recognize the enormous water usage as a critical concern for socially responsible and
environmentally sustainable AI. Our key contribution is to make the first-of-its-kind efforts to uncover the
secret water footprint of AI models. Specifically, we present a principled methodology to estimate AI’s water
footprint. Then, using GPT-3 as an example, we show that a large AI model can consume a stunning amount
of water in the order of millions of liters for training. 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
9
model can significantly cut the water footprint. In addition, we point out the need for increased transparency
of AI models’ water footprint, and highlight the necessity of holistically addressing water footprint along
with carbon footprint to enable truly sustainable AI.
AI models’ water footprint can no longer stay under the radar — water footprint must be addressed as a priority as
part of the collective efforts to combat global water challenges.
References
[1] Facebook. Sustainability report, 2020, https://sustainability.fb.com/wp-content/uploads/
2021/06/2020_FB_Sustainability-Report.pdf.
[2] Google. Water commitments, 2023, https://sustainability.google/commitments/water/.
[3] UN Water Conference. How ’aquapreneurs’ are innovating to solve the water crisis, 2023, https:
//www.weforum.org/agenda/2023/03/un-water-conference-aquapreneurs- innovation.
[4] U.S. White House. White House action plan on global water security, 2022, https://www.whitehouse.
gov/wp-content/uploads/2022/06/water-action-plan_final_formatted.pdf.
[5] David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, An-
drew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Alexandra Sasha
Luccioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla P. Gomes,
Andrew Y. Ng, Demis Hassabis, John C. Platt, Felix Creutzig, Jennifer Chayes, and Yoshua Bengio.
Tackling climate change with machine learning. ACM Comput. Surv., 55(2), feb 2022.
[6] A. Shehabi, S. J. Smith, N. Horner, I. Azevedo, R. Brown, J. Koomey, E. Masanet, D. Sartor, M. Herrlin,
and W. Lintner. United States data center energy usage report. Lawrence Berkeley National Laboratory,
Berkeley, California. LBNL-1005775, 2016.
[7] David Patterson, Joseph Gonzalez, Urs H¨
olzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel
Rothchild, David R. So, Maud Texier, and Jeff Dean. The carbon footprint of machine learning training
will plateau, then shrink. Computer, 55(7):18–28, 2022.
[8] Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. Green AI. Commun. ACM, 63(12):54–63,
nov 2020.
[9] Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep
learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics,
pages 3645–3650, Florence, Italy, July 2019. Association for Computational Linguistics.
[10] Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau. Towards
the systematic reporting of the energy and carbon footprints of machine learning. J. Mach. Learn. Res.,
21(1), jan 2020.
[11] Bill Tomlinson, Rebecca W. Black, Donald J. Patterson, and Andrew W. Torrance. The carbon emissions
of writing and illustrating are lower for ai than for humans. arXiv 2303.06219, 2023.
[12] The water resources group: Background, impact and the way forward, http://www3.weforum.org/
docs/WEF/WRG_Background_Impact_and_Way_Forward.pdf.
[13] United Nations. Water scarcity, https://www.unwater.org/water-facts/water-scarcity.
[14] United Nations Children’s Fund (UNICEF). Water scarcity — addressing the growing lack of available
water to meet children’s needs., https://www.unicef.org/wash/water-scarcity.
[15] Carole-Jean Wu, Ramya Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani, Kiwan Maeng, Glo-
ria Chang, Fiona Aga, Jinshi Huang, Charles Bai, et al. Sustainable AI: Environmental implications,
challenges and opportunities. In Proceedings of Machine Learning and Systems, volume 4, pages 795–813,
2022.
10
[16] Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. Estimating the carbon footprint
of BLOOM, a 176B parameter language model. Journal of Machine Learning Research, 24(253):1–15, 2023.
[17] Steven Gonzalez Monserrate. The Cloud Is Material: On the Environmental Impacts of Computation
and Data Storage. MIT Case Studies in Social and Ethical Responsibilities of Computing, Winter, January
2022. https://mit-serc.pubpub.org/pub/the-cloud-is-material.
[18] Mohammad. A. Islam, Shaolei Ren, Gang Quan, Muhammad Z. Shakir, and Athanasios V. Vasilakos.
Water-constrained geographic load balancing in data centers. IEEE Trans. Cloud Computing, 2015.
[19] Md Abu Bakar Siddik, Arman Shehabi, and Landon Marston. The environmental footprint of data
centers in the United States. Environmental Research Letters, 16(6):064017, 2021.
[20] Paul Reig, Tianyi Luo, Eric Christensen, and Julie Sinistore. Guidance for calculating water use em-
bedded in purchased electricity. World Resources Institute, 2020.
[21] Google. Environmental report, 2023, https://sustainability.google/reports/.
[22] Urs H¨
olzle. Our commitment to climate-conscious data center cool-
ing, 2022, https://blog.google/outreach-initiatives/sustainability/
our-commitment-to-climate- conscious-data-center- cooling/.
[23] Microsoft. Environmental sustainability report, 2022, https://www.microsoft.com/en-us/
corporate-responsibility/sustainability/report.
[24] U.S. Central Intelligence Agency. The world fact book — total water withdrawal, 2020, https://www.
cia.gov/the-world-factbook/field/total-water- withdrawal/.
[25] Meta. Sustainability report, 2023, https://sustainability.fb.com/
2023-sustainability-report/.
[26] Hannah Ritchie and Max Roser. Water use and stress. Our World in Data, 2017, https://
ourworldindata.org/water-use-stress.
[27] The Guardian. ‘It’s pillage’: Thirsty Uruguayans decry Google’s plan to exploit water supply, July 11
2023.
[28] Myeongjae Jeon, Shivaram Venkataraman, Amar Phanishayee, unjie Qian, Wencong Xiao, and Fan
Yang. Analysis of large-scale multi-tenant gpu clusters for dnn training workloads. In Proceedings
of the 2019 USENIX Conference on Usenix Annual Technical Conference, USENIX ATC ’19, page 947–960,
USA, 2019. USENIX Association.
[29] Kim Hazelwood, Sarah Bird, David Brooks, Soumith Chintala, Utku Diril, Dmytro Dzhulgakov, Mo-
hamed Fawzy, Bill Jia, Yangqing Jia, Aditya Kalro, James Law, Kevin Lee, Jason Lu, Pieter Noordhuis,
Misha Smelyanskiy, Liang Xiong, and Xiaodong Wang. Applied machine learning at facebook: A dat-
acenter infrastructure perspective. In 2018 IEEE International Symposium on High Performance Computer
Architecture (HPCA), pages 620–629, 2018.
[30] Alex de Vries. The growing energy footprint of artificial intelligence. Joule, October 2023.
[31] The U.K. Department for Environment, Food and Rural Affairs. Drought explained, August 2023,
https://deframedia.blog.gov.uk/2023/08/09/drought-explained/.
[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay
Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Can-
ton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu,
Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini,
Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Ko-
renev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai
11
Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew
Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael
Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang
Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan
Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open
foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.
[33] Kai Fang, Reinout Heijungs, and Geert R. de Snoo. Theoretical exploration for the combination of the
ecological, energy, carbon, and water footprints: Overview of a footprint family. Ecological Indicators,
36:508–518, 2014.
[34] Mohammad A. Islam, Kishwar Ahmed, Hong Xu, Nguyen H. Tran, Gang Quan, and Shaolei Ren.
Exploiting spatio-temporal diversity for water saving in geo-distributed data centers. IEEE Transactions
on Cloud Computing, 6(3):734–746, 2018.
[35] Andrew A Chien, Liuzixuan Lin, Hai Nguyen, Varsha Rao, Tristan Sharma, and Rajini Wijayawardana.
Reducing the carbon impact of generative ai inference (today and in 2035). In Proceedings of the 2nd
Workshop on Sustainable Computer Systems, HotCarbon ’23, New York, NY, USA, 2023. Association for
Computing Machinery.
[36] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss,
Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens
Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark,
Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language
models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors,
Advances in Neural Information Processing Systems, volume 33, pages 1877–1901. Curran Associates, Inc.,
2020.
[37] Melanie Nakagawa. The journey to water positive, 2023, https://blogs.microsoft.com/
on-the-issues/2023/03/22/water-positive- climate-resilience-open- call/.
[38] Meta. Sustainability — water, https://sustainability.fb.com/water/.
[39] Harsha Nori, Nicholas King, Scott Mayer McKinney, Dean Carignan, and Eric Horvitz. Capabilities of
GPT-4 on medical challenge problems. arXiv preprint arXiv:2303.13375, 2023.
[40] Paul Reig. What’s the difference between water use and water consump-
tion? World Resources Institute Commentary, 2013, https://www.wri.org/insights/
whats-difference-between-water- use-and-water- consumption.
[41] Jordan Macknick, Robin Newmark, Garvin Heath, and KC Hallett. A review of operational water con-
sumption and withdrawal factors for electricity generating technologies. NREL Tech. Report: NREL/TP-
6A20-50900, 2011.
[42] OECD Data. Water withdrawals, https://data.oecd.org/water/water-withdrawals.htm.
[43] Elliot Cohen and Anu Ramaswami. The water withdrawal footprint of energy supply to cities. Journal
of Industrial Ecology, 18(1):26–39, 2014.
[44] Leila Karimi, Leeann Yacuel, Joseph Degraft-Johnson, Jamie Ashby, Michael Green, Matt Renner, Aryn
Bergman, Robert Norwood, and Kerri L. Hickenbottom. Water-energy tradeoffs in data centers: A case
study in hot-arid climates. Resources, Conservation and Recycling, 181:106194, 2022.
[45] D. Lee. Water efficiency at Facebook’s Prineville data center, https://www.opencompute.org/blog/
water-efficiency-at-facebooks- prineville-data-center.
[46] Microsoft. Microsoft in your community, https://local.microsoft.com/.
12
[47] Microsoft. How a small city in iowa became an epicenter for advancing AI, 2023, https://news.
microsoft.com/source/features/ai/west-des-moines-iowa- ai-supercomputer/.
[48] David Sickinger, Otto Van Geet, Suzanne Belmont, Thomas Carter, and David Martinez. Ther-
mosyphon cooler hybrid system for water savings in an energy-efficient hpc data center. NREL Technical
Report NREL/TP-2C00-72196, 2018.
[49] U.S. Energy Information Administration. U.S. electric power sector continues water efficiency gains,
2023, https://www.eia.gov/todayinenergy/detail.php?id=56820.
[50] Kali Frost and Inez Hua. Quantifying spatiotemporal impacts of the interaction of water scarcity
and water use by the global semiconductor manufacturing industry. Water Resources and Industry,
22:100115, 2019.
[51] Singapore Public Utilities Board. Wafer fabrication and semiconductor plants benchmarks. https:
//www.pub.gov.sg/Documents/WaterEfficiencyBenchmark_WaferFab.pdf.
[52] Dexter Johnson. Scarcity drives fabs to wastewater recycling. IEEE Spectrum, January 2022.
[53] SPX Cooling. Water usage calculator, http://spxcooling.com/green/leed/
water-usage-calculator/.
[54] U.S. Dept. of Energy. Energy demands on water resources, Dec. 2006.
[55] The Green Grid. Water usage effectiveness (WUE): A green grid data center sustainability metric.
Whitepaper, 2011.
[56] Kishwar Ahmed, Mohammad A. Islam, Shaolei Ren, and Gang Quan. Exploiting temporal diversity
of water efficiency to make data center less ‘thirsty’. In ICAC, 2014.
[57] Peter Xiang Gao, Andrew R. Curtis, Bernard Wong, and Srinivasan Keshav. It’s not easy being green.
SIGCOMM Comput. Commun. Rev., 2012.
[58] Udit Gupta, Mariam Elgamal, Gage Hills, Gu-Yeon Wei, Hsien-Hsin S. Lee, David Brooks, and Carole-
Jean Wu. Act: Designing sustainable computer systems with an architectural carbon modeling tool. In
Proceedings of the 49th Annual International Symposium on Computer Architecture, ISCA ’22, page 784–799,
New York, NY, USA, 2022. Association for Computing Machinery.
[59] David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild,
David So, Maud Texier, and Jeff Dean. Carbon emissions and large neural network training, 2021.
[60] Noelle Walsh. How Microsoft measures datacenter water and energy use to improve Azure Cloud
sustainability. Microsoft Azure Blog, April 2022.
[61] Jahau Lewis Chen, Yen-Bou Chen, and Hua-Chih Huang. Quantifying the life cycle water consump-
tion of a machine tool. Procedia CIRP, 29:498–501, 2015. The 22nd CIRP Conference on Life Cycle
Engineering.
[62] U.S. Energy Information Administration. Analysis of Singapore data, https://www.eia.gov/
international/analysis/country/SGP.
[63] U.S. Environmental Protection Agency. eGRID data explorer, https://www.epa.gov/egrid/
data-explorer.
[64] U.S. Energy Information Administration. Open data, https://www.eia.gov/opendata/.
[65] Uisung Lee, Jeongwoo Han, Amgad Elgowainy, and Michael Wang. Regional water consumption for
hydro and thermal electricity generation in the United States. Applied Energy, 210:661–672, 2018.
13
[66] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze
Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng,
Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, De-
hao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou,
Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kath-
leen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny
Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Ale-
jandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew
Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-
Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. LaMDA: Language models for dialog applica-
tions, 2022.
[67] Pengfei Xu, Xiaofan Zhang, Cong Hao, Yang Zhao, Yongan Zhang, Yue Wang, Chaojian Li, Zetong
Guan, Deming Chen, and Yingyan Lin. AutoDNNchip: An automated DNN chip predictor and builder
for both FPGAs and ASICs. In FPGA, 2020.
[68] Suyog Gupta and Berkin Akin. Accelerator-aware neural network design using AutoML. arXiv preprint
arXiv:2003.02838, 2020.
[69] Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ah-
mad Awan, Jeff Rasley, and Yuxiong He. Deepspeed-MOE: Advancing mixture-of-experts inference
and training to power next-generation AI scale. In ICML, 2022.
[70] Lingjiao Chen, Matei Zaharia, and James Zou. Frugalgpt: How to use large language models while
reducing cost and improving performance, 2023.
[71] Chao Li, Amer Qouneh, and Tao Li. iSwitch: Coordinating and optimizing renewable energy powered
server clusters. In ISCA, 2012.
[72] Anshul Gandhi, Kanad Ghose, Kartik Gopalan, Syed Rafiul Hussain, Dongyoon Lee, Yu David Liu,
Zhenhua Liu, Patrick McDaniel, Shuai Mu, and Erez Zadok. Metrics for sustainability in data centers.
In HotCarbon, 2022.
[73] Noman Bashir, Tian Guo, Mohammad Hajiesmaili, David Irwin, Prashant Shenoy, Ramesh Sitaraman,
Abel Souza, and Adam Wierman. Enabling sustainable clouds: The case for virtualizing the energy
system. In Proceedings of the ACM Symposium on Cloud Computing, SoCC ’21, page 350–358, New York,
NY, USA, 2021. Association for Computing Machinery.
[74] Bilge Acun, Benjamin Lee, Fiodar Kazhamiaka, Kiwan Maeng, Udit Gupta, Manoj Chakkaravarthy,
David Brooks, and Carole-Jean Wu. Carbon explorer: A holistic framework for designing carbon
aware datacenters. In Proceedings of the 28th ACM International Conference on Architectural Support for
Programming Languages and Operating Systems, Volume 2, ASPLOS 2023, page 118–132, New York, NY,
USA, 2023. Association for Computing Machinery.
[75] Google. Environmental report, 2022, https://sustainability.google/reports/.
[76] Meta. Sustainability report, 2021, https://sustainability.fb.com/.
[77] Shefy Manayil Kareem. Introducing critical new water data capabilities in Microsoft Cloud for Sustain-
ability, 2023, https://www.microsoft.com/en-us/industry/blog/sustainability/2023/03/22/
introducing-critical-new-water- data-capabilities-in- microsoft-cloud-for- sustainability/.
14
Appendix
A Operational Water for Google, Microsoft, and Meta in 2022
According to the sustainability reports [21, 23, 25], the electricity consumption by Google’s, Microsoft’s,
and Meta’s data centers in 2022 was about 21.7 TWh, 18.1 TWh, and 11.1 TWh, respectively. In particular,
their U.S. electricity consumption in 2022 was about 15.5 TWh, 10.9 TWh, and 8.6 TWh (excluding leased
colocation facilities), respectively. Microsoft and Google did not report country-level data for their electricity
consumption in North America, but they only have self-managed data centers in the U.S. as of 2023. Thus,
we attribute their electricity consumption in North America to the U.S. The reported electricity consumption
by Google and Microsoft may also include negligible usage by their offices and will be adjusted by rounding
down our calculations.
The total combined electricity consumption of Google, Microsoft, and Meta was about 50 TWh in 2022.
While Microsoft did not disclose country-/location-wise electricity consumption, these three companies all
primarily operate in the U.S. Thus, we use the U.S. average electricity water withdrawal intensity factor 43.83
L/kWh [49], and electricity water consumption intensity factor 3.14 L/kWh [20] to calculate the scope-2
water withdrawal and water consumption, respectively. Note that the U.S. average electricity water with-
drawal/consumption intensity factor is lower than the global average [20]. Also, our value of 3.14 L/kWh
for the U.S. average water consumption factor is lower than 7.6 L/kWh used for estimating the U.S. data
center water footprint [6] as well as lower than Meta’s global electricity water consumption intensity factor
of 3.58 L/kWh (i.e., 41,172,356 cubic meters divided by 11,508,131 MWh) [25]. Therefore, 43.83 L/kWh and
3.14 L/kWh for electricity water withdrawal and consumption intensity factors are more on the conservative
side, and our estimate for Meta’s global scope-2 water consumption is less than its self-reported value [25].
Google and Microsoft did not disclose their scope-2 water withdrawal or consumption. By multiplying 43.83
L/kWh and 3.14 L/kWh by 50 TWh, we obtain the total scope-2 water withdrawal of 2.19 billion cubic meters
and water consumption of 0.157 billion cubic meters, respectively.
The on-site scope-1 water withdrawal for Google’s, Microsoft’s, and Meta’s data centers in 2022 were
about 25 billion liters, 8.5 billion liters (assuming 80% of Microsoft’s total water withdrawal based on Meta’s
and Google’s percentages), and 3.6 billion liters, respectively. Their scope-1 water consumption was about
19.7 billion liters, 6 billion liters (assuming the offices consumed 10% of their water withdrawal), and 2.5
billion liters, respectively. These result in a total scope-1 water withdrawal of about 37 billion liters and
water consumption of 28 billion liters in 2022, respectively.
By adding up the scope-1 and scope-2 water usage together, we get a total water withdrawal of 2.2 billion
cubic meters and total water consumption of 0.18 billion cubic meters, respectively, for Google, Microsoft,
and Meta combined together in 2022. Similarly, their 2022 total water withdrawal and water consumption
in the U.S. were about 1.5 billion cubic meters and 0.13 billion cubic meters, respectively.
According to the U.S. Central Intelligence Agency [24], the estimated annual water withdrawals of Den-
mark, Liberia and the U.S. in 2020 (the latest year available) were 0.98 billion cubic meters, 0.14 billion cubic
meters, and 444 billion cubic meters, respectively.
B Operational Water for Global AI in 2027
A recent study suggests that the global AI demand might consume 85 – 134 TWh of electricity in 2027 based
on the GPU shipment [30]. Based on this study, we estimate the potential water usage of global AI in 2027.
To account for the cooling energy overheads, we assume a power usage effectiveness (PUE) of 1.1, which
is a fairly low value even for state-of-the-art data center facilities [21]. We obtain an estimate of the total
electricity consumption of 93.5 – 147.4 TWh.
For electricity water withdrawal and consumption intensity factors, we use the U.S. average electricity
water withdrawal intensity factor 43.83 L/kWh [49], and electricity water consumption intensity factor 3.14
L/kWh [20], respectively. Note again that the U.S. average electricity water withdrawal/consumption in-
tensity factor is lower than the global average [20] and that the value of 3.14 L/kWh for the U.S. average
water consumption factor is also lower than Meta’s global electricity water consumption intensity factor of
3.58 L/kWh [25]. After multiplying 43.83 L/kWh and 3.14 L/kWh by 93.5 – 147.4 TWh, we obtain the total
scope-2 water withdrawal of 4.10 – 6.46 billion cubic meters and water consumption of 0.29 – 0.46 billion
15
cubic meters, respectively.
For on-site scope-1 water withdrawal, we assume 1.2 L/kWh (roughly the same as Google’s global scope-
1 water withdrawal efficiency [21]), which results in a total scope-1 water withdrawal of 0.11 – 0.16 billion
cubic meters. Similarly, assuming 1 L/kWh based on Google’s global scope-1 water consumption efficiency,
we obtain a total on-site scope-1 water consumption of 0.09 – 0.14 billion cubic meters.
By adding up scope-1 and scope-2 water usage together, the total water withdrawal and water consump-
tion of global AI may reach 4.2 – 6.6 billion cubic meters and 0.38 – 0.60 billion cubic meters, respectively.
According to the U.S. Central Intelligence Agency [24], the estimated U.S. annual water withdrawal in 2020
(the latest year available) was about 444 billion cubic meters. By using this number and assuming the U.S.
hosts half of the global AI workloads, the operation of AI workloads in the U.S. may take up roughly 0.5 –
0.7% of its total annual water withdrawal in 2027.
The future water withdrawal and consumption efficiency may both improve in 2027 compared to the cur-
rent values we use, especially for electricity generation. Also, the country-wide annual water withdrawal is
estimated by the U.S. Central Intelligence Agency [24] for the year of 2020, and may vary in 2027. Therefore,
like the prediction of global AI energy consumption in 2027 [30], the prediction of water withdrawal and
consumption for global AI in 2027 should only be interpreted as a rough first-order estimate subject to future
uncertainties.
16
