The environmental footprint of data centers in the United States

Information availability on data center location and size varies depending on its type and owner. Ganeshalingam et al [4] reports likely locations of in-house small and midsize data centers, which house approximately 40% of US servers. Detailed information on colocation and hyperscale data centers is derived from commercial compilations [19–21] that get direct support and input from data center service providers.

We classified data centers based on the International Data Corporation classification system (summarized in table S1) and estimated the electricity use based on data center floor space. We used IT load intensity values (IT_s in watt/ft²) for different data center types (s) from Shehabi et al [22] to estimate the total energy requirements ($DC\_E_{\textrm{total}}$; in MWh) of colocation and hyperscale data centers as follows:

$$\begin{equation} DC\_E_{\textrm{total}} = IT_{s}\times PUE_{s}\times A \end{equation} \tag{ 1 }$$

where PUE_s is the power usage effectiveness of space type s, and A is the floor area of data center in ft². We account for potential overstatement of data center capacity [4], a lack of distinction between gross and raised floor area, and unfilled rack capacity by scaling our server counts to match the 2018 estimate of servers by data center type [3], as shown in table 1 and figure S2. Scaled server estimates are then spatially distributed in proportion to the current spatial distribution of installed server bases. The number of servers by state is shown in figure S2.

Power usage effectiveness (PUE) is a key metric of data center energy efficiency [23]. A value of 1.0 is ideal as it indicates all energy consumed by a data center is used to power computing devices. Energy used for non-computing components, such as lighting and cooling, increases the PUE above 1.0 (see equation (2)). Generally, a data center's PUE is inversely proportionate to its size since larger data centers are better able to optimize their energy usage. Average PUE values and energy use by data center type were taken from Masanet et al [3] and shown in table 1 and table S1.

$$\begin{equation} PUE = \frac{\textrm{Total power supplied to the data center} }{\textrm{Power consumed by the IT equipment}}. \end{equation} \tag{ 2 }$$

Power plant-specific electricity generation and water consumption data come from the US Energy Information Administration (EIA) [24]. Of the approximately 9000 US power plants, the EIA requires nearly all power plants report electricity generation. However, only power plants with generation capacity greater than 100 MW (representing three-fourths of total generation) must report water consumption. We assigned national average values of water consumption per unit of electricity generation by fuel type (i.e. water intensity; m³ MW h⁻¹) to all power plants with unspecified water consumption. Operational water footprints of solar and wind power were taken from Macknick et al [25]. Following Grubert [26], we assign all reservoir evaporation to the dam's primary purpose (e.g. hydropower). We connected hydroelectric dams with their respective power plants using data from Grubert [27]. Reservoir specific evaporation comes from Reitz et al [28].

The U.S. Environmental Protection Agency's eGRID database [29] provided GHG emissions associated with each power plant. GHG emissions are converted to an equivalent amount of carbon dioxide (CO₂)-eq with the same global warming potential so to derive a single carbon footprint metric [30]. Direct GHG emission during the operation of data centers are negligible [18] and therefore not considered in this study.

Data centers, water suppliers, and wastewater treatment plants typically utilize electricity generated from a mix of power plants connected to the electrical grid. Within the electrical grid, electricity supply matches electricity demand by balancing electricity generation within and transferred into/out of a power control area (PCA). Though it is infeasible to trace an electron generated by a particular power plant to the final electricity consumer, there are several approaches to relate electricity generation to electricity consumption (Siddik et al [31] summarizes the most common approaches).

Here, we primarily rely on the approach used by Colett et al [32] and Chini et al [33] to identify the generative source of electricity supplied to any given data center. This approach assesses electricity generation and distribution at the PCA level where it is primarily managed. PCA boundaries are derived from the Homeland Infrastructure Foundation level data [34] and crosschecked against Form EIA-861 [35], which identifies the PCAs operating in each state. Annual inter-PCA electricity transfers reported by the Federal Energy Regulatory Commission [36] are also represented within this approach. A data center (as well as water and wastewater utilities) draws on electricity produced within its PCA, unless the total demand of all energy consumers within the PCA exceeds local generation, in which case electricity imports from other PCAs are utilized. If a PCA's electricity production equals or exceeds the PCA's electricity demand, it is assumed all electricity imports pass through the PCA and are re-exported for utilization in other PCAs. Siddik et al [31] notes that water and carbon footprints are sensitive to the attribution method used to connect power plants to energy consumers. Therefore, we conduct a sensitivity analysis (see the supporting information for additional details) to test the degree to which our electricity attribution method affects our results. Additionally, we also test different assumptions regarding the water footprint of hydropower generation, as this too is a key source of uncertainty.

We focus on the annual temporal resolution and assume an average electricity mix proportional to the relative annual generation of each contributing power plant. Though the electricity mix within a PCA can fluctuate hourly depending on balancing measures, these intra-annual variations will not significantly impact our annual-level results. While it is infeasible to determine the precise amount of electricity each power plant provides to each data center, water utility, and wastewater treatment plant, our approach will enable us to estimate where each facility is most likely to draw its electricity. The dependency of a data center on local and imported electricity from other PCAs was calculated using equations (3) and (4).

$$\begin{equation} DC\_E_{p,l} = DC\_E_{p}\times\Big(1-\sum_{i} r_{i}\Big). \end{equation} \tag{ 3 }$$

$$\begin{equation} DC\_E_{p,im} = DC\_E_{p}\times\sum_{i} r_{i} \end{equation} \tag{ 4 }$$

where ${DC\_E_{p,l}}$ and $DC\_E_{p,i}$ are the local (l) and imported (im) electricity (MWh) to a data center from PCA p, respectively. $DC\_E_p$ is the total electricity consumption of the data center, whereas r_i represents the electricity contribution of each PCA i to PCA p as follows:

$$\begin{equation*} r_i = \begin{cases} \displaystyle\frac{Import_{con}}{Generation_p + \sum Import_p - \sum Export_p},& \mathrm{if\ PCA}\ p \ \textrm{is net importer}\\ 0, & \mathrm{if\ PCA}\ p \ \mathrm{is\ net\ exporter} \end{cases} \end{equation*}$$

where Import_con is defined as the electricity from a linked PCA i that was consumed within PCA p. Any imported electricity not consumed with PCA p is re-exported.

Adjusted electricity consumption from the PCAs were assigned to the power plants using equation (5).

$$\begin{equation} DC\_E_{p,k} = DC\_E_{p,adj} \times \frac{{PP_k}} {\sum_{k = 1}^{n} PP_k} \end{equation} \tag{ 5 }$$

where $DC\_E_{p,k}$ is the total energy directly consumed [MWh/y] by data centers from power plant k that is attributed to PCA p, $DC\_E_{p,adj}$ is the total electricity consumption of the data center from PCA p after adjusting for the inter-PCA electricity transfers, PP_k is the net generation by a specific power plant in MWh/y, and n is the number of power plants within PCA p. A similar approach was taken to connect power plants to water and wastewater utilities, with their electricity usage (and associated environmental footprints) then linked to the data center they service. Boiler feed pumps require an insignificant amount of electricity to provide water to power plants. Therefore, we truncate our analysis at this point.

The indirect water and carbon footprint of each data center consists of water consumption or GHG emissions associated with the generation of (i) electricity utilized during data center operation, (ii) electricity used by water treatment plants for treatment and supply of cooling water to data centers, and (iii) electricity used by wastewater treatment plants to treat the wastewater generated by a data center. The GHG emissions or water consumption of a power plant supplying electricity to a data center is attributed to the data center as follows:

$$\begin{equation} DC\_IF_{k} = DC\_E_{k} \times F_k \end{equation} \tag{ 6 }$$

where $DC\_IF_k$ is the indirect footprint (water or carbon) associated with electricity used during the operation of a data center from power plant k and $DC\_E_k$ is the total energy used [MWh/y] by a data center from power plant k (from equation (5)). When calculating the indirect water footprint, F_k is the water consumption per unit of electricity generated by power plant k in m³ MWh⁻¹. When calculating the indirect carbon footprint, F_k is the GHG emitted per unit of generated electricity by power plant k (tons CO₂-eq MWh⁻¹).

Although the IPCC does not consider water treatment a notable emitter of GHGs [37], wastewater treatment plants are a major source of GHG emission [38, 39]. In 2017, total GHG gas emission from wastewater treatment plants was estimated to be 20 million metric tons, with a direct emission rate of 0.3 kg CO₂-eq/y per m³ of wastewater treated [38, 39]. In absence of facility specific emission data, we have used the average emission rate for treating wastewater for all wastewater generated from data center operation [39]. No direct GHG emissions are assumed to be associated with data center operation at the facility [18].

The EPA Safe Drinking Water Information System contains information on the location, system type, and source of water for each public water and wastewater utility [40, 41]. We assumed the nearest non-transient water treatment plant and wastewater treatment plant services a data center's water demand and wastewater management, respectively. After calculating the water supply requirement of a data center (discussed later in this section), the electricity needed for treatment and distribution of cooling water can be calculated using the data from Pabi et al [13] (see table S2). Water and wastewater treatment plants were linked to power plants (as described previously) to estimate the indirect water footprint associated with electricity required to distribute and treat water and wastewater used by a data center. We then sum the water consumed by each power plant to directly or indirectly service a data center to determine the total indirect water footprint of that data center. The indirect water footprint associated with each power plant was also aggregated within watershed boundaries to determine which water sources each data center was reliant upon.

Direct water consumption of a data center can be estimated from the heat generation capacity of a data center [42], which is related to the amount of electricity used [43]. Estimates of data center specific electricity demand were multiplied by the typical water cooling requirement [1]—1.8 m³ MWh⁻¹—to estimate the direct water footprint of each data center. The direct water consumption is assigned to the watershed where the water utility supplying the data center withdraws its water.

Data center wastewater is largely comprised of blowdown; that is, the portion of cooling water removed from circulation and replaced with freshwater to prevent excessive concentration of undesirable components [44]. We assume all data centers utilize potable water supplies and cycle this water until the concentration of dissolved solids is roughly five times the supplied water [44]. We calculate blowdown from data center cooling towers using the following commonly employed approach [45]:

$$\begin{equation} R_{\textrm{Blowdown}} = \frac{1 } {C-1} \times {R_{\textrm{Evaporation}}} \end{equation} \tag{ 7 }$$

where $R_{\textrm{Blowdown}}$ is the blowdown rate required for a cooling tower (m³ MWh⁻¹), C is the cycle of concentration for dissolved solids (assumed here as 5), and $R_{\textrm{Evaporation}}$ is the rate of evaporation (m³ MWh⁻¹).

The water scarcity footprint (WSF; as defined by ISO 14046 and Boulay et al [46]) indicates the pressure exerted by consumptive water use on available freshwater within a river basin and determines the potential to deprive other societal and environmental water users from meeting their water demands. We quantified the WSF of data centers using the AWARE method set forth by Boulay et al [46] (see the Supportive Information for more details). Other societal and environmental water use data, as well as data on natural water availability within each US watershed, come from [47–49].

The total annual operational water footprint of US data centers in 2018 is estimated at 5.13 × 10⁸ m³. Data center water consumption is comprised of three components: (i) water consumed directly by the data center for cooling and other purposes (figure 2(A)), (ii) water consumed indirectly through electricity generation (figure 2(B)), and (iii) water consumed indirectly via the water embedded with the electricity consumption of water and wastewater utilities servicing the data center (figure 2(C)). The data center industry directly or indirectly draws water from 90% of US watersheds, as shown in figure 3(A).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Roughly three-fourths of US data centers' operational water footprint is from indirect water dependencies. The indirect water footprint of data centers in 2018 due to their electricity demands is 3.83 × 10⁸ m³, while the indirect water footprint attributed to water and wastewater utilities serving data centers is several orders of magnitude smaller (4.50 × 10⁵ m³). Nationally, we estimate that 1 MWh of energy consumption by a data center requires 7.1 m³ of water. However, this national average masks the large spatial variation (range 1.8–105.9 m³) in water demand associated with a data center's energy consumption. Data centers are indirectly dependent on water from every state in the contiguous US, much of which is sourced from power plants drawing water from subbasins in the eastern and western coastal states. Less than one-fifth of the industry's total electricity demand is from data centers in the West and Southwest US (regions as defined by NOAA [50]; see outlined areas in figures 2–5, and figure S4 for region identification), yet nearly one-third of the industry's indirect water footprint is attributed to data centers in these regions. Indirect water consumption associated with energy production in Southwest subbasins is particularly high, despite relatively low electricity supplied from this region, due to the disproportionate amount of electricity from water-intensive hydroelectricity facilities and the high evaporative potential in this arid region. Conversely, the Southeastern region consumes one-quarter of the electricity used by the industry but only one-fifth of the indirect water since data centers in this region source their electricity from less water-intensive sources.

On-site, direct water consumption of US data centers in 2018 is estimated at 1.30 × 10⁸ m³. Collectively, data centers are among the top-ten water consuming industrial or commercial industries in the US [47]. Approximately 1.70 × 10⁷ m³ of water directly consumed by data centers are sourced from a different subbasin than the location of the installed servers. Large direct water consumption in the Northeast, Southeast, and Southwest regions indicate clustering of servers in these regions. Combined direct and indirect water and carbon intensities are broken down by data center type in table 1.

The WSF of data centers in 2018 is 1.29 × 10⁹ m³ of US equivalent water consumption, which is more than twice that of the volumetric water footprint reported in the previous section. The WSF (including both direct and indirect water requirements) per unit of energy consumption is 17.9 m³ US-eq water MWh⁻¹, more than double the nationally averaged water intensity (7.1 m³ MWh⁻¹) that does not account for water scarcity. WSFs that are larger than volumetric water footprints suggest that data centers disproportionately utilize water resources from watersheds experiencing greater water scarcity than average.

Only one-fourth of the volumetric water footprint of data centers resulted from onsite water use. Yet, more than 40% of the WSF is attributed to direct water consumption. This indicates that direct water consumption of data centers, which occurs close to where the data center is located, is skewed toward water stressed subbasins compared to its indirect water consumption, which is distributed more broadly geographically. We find that most of the watersheds that data centers draw from, particularly those in the Eastern US, face little to no water stress on average. In contrast, many of the watersheds in the Western US exhibit high levels of water stress, which is exacerbated by data centers direct and indirect water demands. Combined, the West and Southwestern watersheds supply only 20% of direct water and and 30% indirect water to data centers, while hosting approximately 20% of the nation's servers. Yet, 70% of the overall WSF occurs in these two regions (figure 3(B)), which indicates a disproportionate dependency on scarce waters in the western US.

Total GHG emissions attributed to data centers in 2018 was 3.15 × 10⁷ tons CO₂-eq, which is almost 0.5% of total GHG emissions in the US [10]. A little over half (52%) of the total emissions of data center operations are attributed to the Northeast, Southeast, and Central US, which have a high concentration of thermoelectric power plants, along with large number of data centers (figure 3(C)). Almost 30% of the data center industry's emissions occur within the Central US, which relies heavily on coal and natural gas to meet its electricity demand. Yet, only 10% the industry's energy demand comes from the Central US, and just 9% of the water consumption associated with data centers operation occurs in this region. Moreover, the Central region is a net exporter of electricity to other regions, providing electricity for data centers located in the Northeast and Southeast regions, which houses almost one-third of servers. Yet, the generation of less carbon intensive electricity in the Northeast (hydroelectricity) and Southeast (wind/solar) regions means that while their electricity consumption comprises 34% of data centers' national electricity demand, these regions only constitute 23% of the industry's GHG emissions. The GHG emissions from treating the wastewater generated from data centers is around 550 tons/y (0.002% of total GHG emissions associated with data centers).

Our results indicate significant variability of environmental impacts depending on where a data center is located. Here we explore how the geographic placement of a data center can lead to improved environmental outcomes. We find that the total water intensity of a data center can range from 1.8–106 m³ MWh⁻¹, the water scarcity intensity from 0.5 to 305 m³ US-eq MWh⁻¹, and the carbon intensity from 0.02 to 1 ton CO₂-eq MWh⁻¹ depending on where the data center is placed (figure 4). Data center placement decisions are complicated by the electricity grid, which displaces environmental impacts from the physical location of a data center.

Zoom In Zoom Out Reset image size

Figure 5 depicts subbasins in the top quartile of environmental performance as it relates to water footprint (5(A)), WSF (5(B)), and carbon footprint (5(C)) per MWh of electricity used by a hypothetical data center located within each subbasin. Less than 5% of subbasins are in the top quartile of environmental performance for both WSF and carbon footprint (hatched areas in figures 5(B) and (C), meaning that 40% of subbasins will require making a trade-off between reducing WSFs and carbon footprints. The remaining 55% of subbasins (white areas shared by figures 5(B) and (C) are not among the best locations to place a data center for either water or GHG reduction. Though the water footprint and WSF are related concepts, we show that nearly one-fifth of subbasins that were in the top quartile with respect to the water footprint are in the bottom quartile for WSF. In other words, a data center placed in these basins would use less water than 75% of potential sites, but it would draw that water from subbasins facing higher levels of water scarcity. In general, locating a data center within the Northeast, Northwest, and Southwest will reduce the facilities carbon footprint, while locating a data center in the Midwest and portions of the Southeast, Northeast, and Northwest will reduce its WSF.

Zoom In Zoom Out Reset image size

In the coming years, cloud and hyperscale data centers will replace many smaller data centers [3]. This shift will lower the environmental footprint in some instances but introduce new environmental stress in other areas. Assuming added servers employ similar technology as existing servers and are placed in cloud and hyperscale data centers in proportion to the current spatial distribution of data centers (i.e. business-as-usual scenario), these new data center servers will have a collective water footprint of 77.77 × 10⁶ m³ (15% of the current industry total), WSF of 170.56 × 10⁶ m³ US-eq (9%), and 4.36 × 10⁶ tons CO₂-eq (14%). However, if these new servers are strategically placed in areas identified to have a lower environmental footprint, their water and carbon burden could be significantly reduced.

The WSF and carbon footprint of new data centers can be reduced by 153.00 × 10⁶ m³ US-eq (90% less than business-as-usual expansion) and 2.34 × 10⁶ tons CO₂-eq (55%), respectively (figure 6(A)) if they are placed in areas with the lowest carbon and WSFs (hatched areas in figure 5). However, placing all new data centers within a small area may strain local energy and water infrastructure due to their collective water and energy demands. Data centers can be dispersed more broadly in areas that are favorable with respect to water footprint (figure 5(A)), WSF (figure 5(B)), or carbon footprint (figure 5(C)). However, only considering one environmental characteristic can lead to environmental trade-offs (figure 6).

Zoom In Zoom Out Reset image size