Toward climate neutral data centers: Greenhouse gas inventory, scenarios, and strategies

Summary

The climate impact of data centers is expected to increase due to rising demand for information and communication technology services. At the same time, the European Union aims for climate neutral data centers by 2030. To map potential developments of emissions associated with data centers to the year 2030, we develop a generic data center greenhouse gas (GHG) inventory in accordance with the GHG protocol. Our results show that climate neutrality cannot be achieved by reduction measures only. We predict an increase in energy use for our illustrative Germany-based data center by 20%, and an increase in the overall carbon footprint by 13% between 2020 and 2030. Although operational measures and increased use of renewable electricity can decrease GHG emissions over the 2020 level by up to 70%, achieving net zero inevitably requires mechanisms to mitigate and compensate emissions across the value chain.

Graphical abstract

Highlights

• •We expect a 20% increase in energy use and a 13% rise in GHG emissions by 2030
• •High utilization rates and efficient infrastructure are highly effective measures
• •Renewable energy continues to be fundamental for achieving net-zero data centers
• •Refining GHG reporting metrics is essential for transparency for data centers

Information systems; Energy sustainability; Energy systems; Energy Modelling

Introduction

About 1.4% of global greenhouse gas (GHG) emissions in 2020 originate from the information and communication technology (ICT) sector, of which 16% are attributable to data centers.¹ Between 2010 and 2018, global data center energy use increased by 6% due to efficiency improvements in hardware and software, despite a much higher increase in demand.² However, particularly as a consequence of anticipated developments, such as demand growth induced by new emerging technologies,³ the slowing of Moore law,^4,5,6,7,8,9 and increased use of technologies with more embodied emissions, in particular Solid-State Drives (SSDs),¹⁰ a steeper increase in energy use is anticipated for the upcoming years.^11,12 Hence, efforts to enhance the efficiency and minimize the climate impact of ICT and specifically cloud data centers are increasingly becoming a priority for industry and politics.

Numerous ICT companies, such as Google and Cisco,^13,14 have initiated emission reduction efforts by setting net-zero targets. Many of these organizations define and monitor their targets in adherence with the GHG Protocol Corporate Accounting and Reporting Standard¹⁵; however, climate neutrality targets of different organizations vary in their extent. Some focus solely on mitigating Scope 1 and 2 emissions, whereas others aim for net zero throughout the entire value chain, including Scope 3 emissions. Besides industry, the European Commission has made commitment to achieving climate neutral data center operations—referring to net-zero Scope 2 emissions—as part of its digital strategy by 2030.¹⁶ However, from a scientific perspective, the principle of life cycle completeness demands to include all GHG emissions when making carbon neutrality claims, i.e., partial footprints of products are not a proper basis for carbon neutrality claims and for organizations. This implies the need to include their Scope 3 emissions.¹⁷

Following the European Union (EU) targets, various data center operators became part of the self-regulatory initiative Climate Neutral Data Center Pact. The targets of the initiative consider aspects such as energy and infrastructural efficiency, clean energy, water consumption, and a circular economy.^18,19

Net-zero data centers are pursued despite the anticipated increase in data center energy use. This raises questions as to whether and under what circumstances fulfilling these targets is possible.²⁰

Current forecasts of data center electricity consumption and GHG emissions vary in their assumptions, system boundaries, accounting methods, and reference locations. Thus, they are not consensual or fully comparable on future developments.^{2,4,6,21,22,23,24} Moreover, numerous studies and neutrality targets do not consider emissions upstream and downstream of data center operation, Scope 3. However, these are a non-negligible.^5,21,25,26

Recent research addresses the challenges future projections are bound to. First, external circumstances, such as the development of demand, and rapid technological changes are subject to uncertainty.^4,5,6,27 Second, modeling data centers is highly complex, as they differ in several properties.^{6,21,26,28,29,30,31} Differences include operational and technical characteristics, such as equipment configuration, average utilization, as well as infrastructural efficiency, which is measured by the indicator power usage effectiveness (PUE).³² Additionally, the lack of transparency and data availability concerning equipment power consumption and operational management are major limitations to build representative models of data centers.²⁵

To address the aforementioned challenges and enhance the understanding of the dynamics behind data centers and their climate impact, we present a generic model for quantifying GHG emissions of an illustrative Germany-based data center by 2030. We pursue three main goals: (1) to explore potential future developments of data center energy use and associated GHG emissions, (2) to provide a more profound understanding of the driving forces of data centers’ climate impact, (3) and to evaluate options to achieve climate neutral data centers.

We map several scenarios, whereby we cover uncertainties in future developments and elaborate on the impact of individual parameters. Moreover, we examine the feasibility of net-zero targets under varying assumptions, boundaries, and accounting methods. Although our study focuses on operational and embodied emissions associated with the information technology (IT) equipment and power consumption, we also give insights on the role of other Scope 3-related aspects covered by other studies.

Recognizing the scarcity of publicly available models for estimating data center GHG emissions,² we provide a simple and generic approach that can be adopted and modified to optimize data center operations and can serve as a supportive tool in policymaking.

Method overview and key assumptions

Our methodological approach consists of four steps. At first, we defined a GHG inventory for data centers that is based on the GHG protocol. In the second step, we conducted literature research on existing models quantifying the electricity consumption and GHG emissions of data centers. Based on these steps, we developed a model that enables the determination of Scope 1, Scope 2, and Scope 3 emissions, and makes future projections considering anticipated developments. Last, we performed a scenario analysis considering anticipated developments and operational measures. Tables 1 and 2 summarize all parameters, values, and corresponding literature sources for our baseline scenarios and sensitivity assumptions. All other method details are described in the STAR Methods section.

Table 1.

Parameters and assumptions for the Baseline scenarios 2020 and 2030

	2020 Baseline	2030 Baseline
Electricity specific emission factor (location based)a	0.377 kg CO2e/kWh33	0.3 kg CO2e/kWh34
Server power	187 W35
Storage power	450 W36
Equipment lifetimea	4 years (Also assumed by Lin et al.37)35,36
PUEa	1.6338	1.35 (Extrapolation of a 20% 10-year improvement in the PUE between 2010 and 2020, based on Hintemann’s38 empirical evidence, also consistent with Lin et al. 37 and Stobbe et al.39)
Utilizationa	25%40
Efficiency gainsa	15% per annum (p.a. ) (assuming that computational efficiency gains will persist until 2030, based on Masanet et al.2)41
Workload demand	Compound annual growth rate (CAGR) 19%42
Storage demand	CAGR 16% (assuming a constant CAGR from 2020 to 2030, extrapolated based on the projections between 2020 and 2026)43
Manufacturing emissions servera	1,243 kg CO2e35
Transportation emissions server	25 kg CO2e35
Manufacturing emissions storagea	1,296 kg CO2e36
Transportation emissions storage	486 kg CO2e36
Electricity upstream emission factor	0.024 kg CO2e/kWh44,45	0.026 kg CO2e/kWh44,46

^aDetails on alternative sensitivity assumptions for the 2030 Baseline can be found in Table 2.

Table 2.

Overview of alternative sensitivity assumptions

Electricity-specific emission factor	German climate protection goals: 0.268 kg CO2e/kWh (location based)47
	Worst case: 0.377 kg CO2e/kWh (location based)33
	0 kg CO2e/kWh (green tariff/100% renewable PPA, market based),48 Upstream: 0.028 kg CO2e/kWh44,49
Utilization	30% (Own assumptions based on forecasts and values of utilization)21,50,51
	50% (Assuming a constant CAGR from 2020 to 2030, extrapolated based on the projections between 2020 and 2026)21,50,51
PUE	1.15 (Assuming increased efficiency, based on data indicated the sources)52,53
	1.63 (Assuming no improvement when compared with 2030)38
Efficiency gains	10% p.a. (Assuming flattening efficiency gains, based on Koot and Wijnhoven,4 Shahidi,7 and The Economist54)41
	20% p.a. (Assuming improvements in efficiency gains due to Moore law or other technological breakthroughs23)
Equipment manufacturing emissions	−75% (Own assumption, assuming a notably better grid mix in the producer region)
	+75% (Own assumption, covering emission-intensive technologies/underestimation of the manufacturing emissions)
Equipment lifetime	3 years55
	6 years55

Supplement to Table 1 and 2030 Baseline.

Results

Our results are structured in three sections. The first section gives an overview of our 2030 Baseline projection of data center energy use and associated GHG emissions when compared with the reference year 2020 (location based). The second section presents our findings from analyzing the outcome of sensitivity assumptions (both market and location based). The last section reflects on climate neutrality for different boundaries and accounting methods (market vs. location based).

Baseline projections

Our baseline results indicate a significant growth in the number of installed IT equipment from 2020 to 2030, which leads to an increase in both data center electricity consumption and GHG emissions. According to the expected workload and storage demand increase, the number of actively installed servers increases by 56% and storage devices by 22% (see Figure 1).

Figure 1.

Evolvement of the number of installed servers and storage systems

Underlying data for Figure 1 is available in Supporting Information S1.

We found a 20% increase in power consumption from 2020 to 2030 (Figure 2A). Electricity-related Scope 2 emissions decrease by 4% due to the anticipated renewable energy expansion by lowering the electricity-specific grid emission factor, leading to fewer GHG emissions for higher power consumption levels. Electricity-related Scope 3 emissions increase by 30% (Figure 2B).

Figure 2.

Comparison of 2020 and 2030 Baseline scenarios

(A and B) (A) Data center electricity consumption and (B) electricity-related GHG emissions (location based). Underlying data for Figure 2 is available in Supporting Information S2.

Total location-based emissions are provided in Figure 3. These are anticipated to increase by 13%, from 13,714 t CO₂e to 15,522 t CO₂e. In 2020, 83% of these emissions are Scope 2, which we anticipate will correspond to 70% of total emissions by 2030. Server power consumption dominates electricity-related GHG emissions (Figure 2B).

Figure 3.

Results of the 2030 Baseline scenario

(A–C) (A) Absolute GHG emissions, (B) Scope 2 GHG emissions breakdown, and (C) Scope 3 GHG emissions breakdown. Underlying data for Figure 3 is available in Supporting Information S3.

We forecast an increase of Scope 3 emissions by 101%. Manufacturing emissions of IT equipment are the main drivers for this growth (Figure 3C). Overall, we found that Scope 3 emissions will gain relevance by increasing from 17% in 2020 to about 30% in 2030. Still, Scope 2 emissions continue to account for the greater share (as shown in Figure 3A).

Sensitivity assumptions

Analyzing the outcome of multiple sensitivity assumptions, we identified a wide range of total emission trajectories ranging from a 66% decrease to a 74% increase in overall GHG emissions when compared with Baseline 2020. Compared with Baseline 2030, alternative assumptions induced difference between 70% less and 54% more GHG emissions. The range of relative differences of the sensitivity assumptions compared with Baseline 2030, grouped into the parameter sets, is presented in Figure 4.

Figure 4.

Relative difference ranges from the sensitivity assumptions to the 2030 Baseline, grouped by parameters

Underlying data for Figure 4 is available in Supporting Information S4 and S5.

Efficiency gains

Analyzing two alternative assumptions for efficiency gains, e.g., by hardware miniaturization, we found that a decrease of efficiency gains from annually 15% to 10% increases energy use up to 83% (54,863 MWh), and total emissions by 74% over the 2020 level. Compared with Baseline 2030, these values correspond to 51% and 54%, respectively. Higher efficiency gains of 20% per annum induce a decrease in total emissions by 33% over Baseline 2030. These results imply that efficiency gains overall are decisive in the evolvement of overall GHG emissions and that the slowing of Moore law⁹ can notably increase data-center-associated GHG emissions.⁴ On the other hand, technological developments that increase equipment performance and enhance efficiency can create mitigation opportunities.

Share of renewable energy

Alternating our baseline assumption from 0.3 kg CO₂e/kWh to 0.268 kg CO₂e/kWh (climate protection scenario) results in a decrease of 11% in Scope 2 emissions. Conversely, assuming no improvement in the grid mix, 0.377 kg CO₂e/kWh (conservative scenario), leads to a 26% increase in Scope 2 emissions. The market-based accounting, 0 kg CO₂e/kWh (green tariff/100% renewable power purchase agreement [PPA]), enables complete abatement of Scope 2 emissions (see also Figure 3A). Overall, changes in Scope 2 emissions are proportional to improvements or deterioration in the electricity-specific emission factor.

Nonetheless, it should be noted that the certification of 100% renewable energy does not abate all electricity-related emissions, as upstream chains of electricity generation are considered Scope 3. These cannot be fully avoided by the use of 100% renewable energy, as renewable energies also cause emissions in their upstream chains, e.g., through the construction of power plants.

Equipment utilization

One effective operational measure is increasing average equipment utilization, which mitigates Scope 2 and Scope 3 emissions. Higher equipment utilization improves workload processing and storage capacities of IT equipment, enabling the initial workload and storage demands to be met with fewer devices. This way not only manufacturing emissions (Scope 3) but also electricity-related emissions due to idling losses are avoided (Scope 2 and Scope 3). Increasing average utilization from 25% to 30% led to an overall reduction of total emissions by 26%. Assuming average utilization of 50%, we yielded 41% fewer total emissions when compared with Baseline 2030. Particularly, virtualization technologies and modernization of hardware and software that enable higher utilization rates are measures for high operational efficiency.^26,56

Power Usage Effectiveness

When decreasing the PUE value from 1.35 to 1.15, which can, e.g., be achieved by switching to liquid cooling, we observed a 15% reduction of Scope 2 and 3% of Scope 3 emissions. The reduction in total GHG emissions corresponds to 11%. When assuming no infrastructural improvements when compared with 2020, a PUE of 1.63, overall GHG emissions increase by about 16%.

Manufacturing emissions of IT equipment

Manufacturing emissions of IT equipment make up a notable portion of Scope 3 emissions (Figure 2C). By decreasing the initial assumption by 75%, which could result from an increased use of renewable electricity in manufacturing processes, we observed that Scope 3 emissions nearly halve. Conversely, increasing manufacturing emissions by 75% led to a 55% increase of Scope 3 emissions. These findings highlight the importance of decarbonizing manufacturing processes to achieve reductions in Scope 3.

Net-zero data centers

Taking into consideration Scope 1 and 2 emissions, as many climate neutrality targets, our location-based results anticipate an increase in power consumption and related GHG emissions caused by data center operations in 2030. This indicates a trend of deviating further from net zero. Thus, large-scale reduction efforts in GHG emissions are central. These can be achieved by enhancing efficiency through operational measures, decarbonization of the electricity mix, and expansion of on-site renewables. When examining an individual data center (market based), Scope 1 and Scope 2 emissions can theoretically become nearly zero upon acquiring 100% renewable energy through PPAs or other contractual instruments, such as energy attribute certificates (EACs), or by meeting energy demand by on-site renewables. Nevertheless, we underline the necessity of encompassing Scope 3 emissions, as climate neutrality assertions counting only Scope 1 and Scope 2 emissions are incomplete and subject to debate.

For proper net zero and climate neutrality claims, it is required to include Scope 3 emissions regardless of whether Scope 2 emissions are accounted market based or location based. Our results support that manufacturing of IT equipment and upstream chains of electricity generation are non-negligible sources of emissions. Although mitigating embodied emissions of IT equipment can be partly achieved through acquiring devices with fewer manufacturing and transportation emissions, emissions upstream of electricity generation cannot be fully avoided, as even renewable energy sources are not net zero in their value chain.⁴⁴ Consequently, efforts to reduce Scope 3 emissions through working with suppliers and setting minimum requirements across the value chain and compensating emissions are inevitable to become climate neutral.

Discussion

Our findings are consistent with other studies that predict an increase in data center energy and an increase in GHG emissions. However, the proportion of increase in existing studies varies. For Germany, Hintemann et al.⁵⁷ suggest for data centers an increase of energy use of approximately 65% from 2020 to 2030, and Stobbe et al.³⁹ suggest an increase of 76% from 2023 to 2033. Both projections significantly exceed our baseline estimate, indicating a 20% increase. When we factor in our sensitivity assumption of 10% annual efficiency gains, our predicted 83% increase aligns with these studies. Therefore, this discrepancy may stem from our optimistic baseline assumption of 15% per annum for efficiency gains, as well as other modeling differences. Notably, both studies analyze the entire data center landscape in Germany, incorporating comprehensive factors beyond top-down assumptions regarding efficiency gains and demand growth. In contrast, Stobbe et al.⁵⁷ base their data center model on actual sales data and market forecasts for IT equipment, rather than solely relying on top-down assumptions. Compared on the EU level, our findings are in line with Montevecchi et al.’s¹² Western Europe projection, when extrapolating their results between 2020 and 2025. Compared with global studies, our results are more optimistic than suggested by The Shift Project⁵⁸ and Andrae’s⁵⁹ “Data Centers Expected” projection of a 225% increase between 2020 and 2030. Similar as with the comparison with the two German studies, our 10% efficiency gains sensitivity assumption is in line with the “Data Centers Best” prediction by Andrae,⁵⁹ which corresponds to an 86% increase. The variability in results based on different efficiency gain assumption supports the findings of Masanet et al.² demonstrating that sustained efficiency gains can further mitigate the growth of energy use in data centers.

Further assumptions and developments

As aforementioned, external developments may substantially affect total emissions. This includes technological advancements and breakthroughs, such as advanced cooling systems, AI, machine learning, and 5G.^60,61 Considering the anticipated shifts to cloud services^51,62 and increased use of virtualization technologies enabling higher equipment utilization and further efficiency gains through software modernization, the development of emissions may substantially differ than suggested. Higher utilization requires additional cooling to ensure reliable operations. As cooling energy demand increases with power draw increase, there is a trade-off between higher utilization and extra cooling energy consumption. In turn, infrastructure energy consumption may increase, and PUE improvements may flatten. Nevertheless, the exact effects of these reciprocal interactions were not subject in our study and require more extensive research. The model allows adaptation of these parameters, and the scenarios chosen exemplify different parameter sets.

An even more central aspect is the development of the national electricity mix. We assumed 0.3 kg CO₂e/kWh for 2030 in our baseline prediction. It should be noted that the value is aligned to the German climate protection goals. Contrary to these goals, the actual value may be up to 0.4 kg CO₂e/kWh⁶³ in the case of slowdowns in the energy transition, for instance, due to increased demand for electricity leading to delays in coal phaseout. This assumption might differ for other countries or regions. For example, the CO₂ intensity of the grid in Norway, which is a common location for data centers, is notably lower, the production mix corresponding to only 0.07 kg CO₂e/kWh already by 2022.⁶⁴ Scope 2 emissions would be nearly zero for an emission factor of this value, also for location-based considerations.

In addition to the presented sensitivity assumption, we analyzed the effects of equipment lifetime, for which we assumed alternative of 3 and 6 years of service life for servers and storage devices. For both servers and storage equipment, a longer service life leads to a greater number of actively installed devices in 2030, as devices with lower processing capacity remain in use for an extended period. Conversely, the count of newly acquired devices in 2030 decreases with a longer service life. As a result, we observed an increasing effect on Scope 2 emissions, and a decreasing effect on Scope 3 emissions. Our study indicates an optimum lifetime of 4 years, although the discrepancies are notably marginal when juxtaposed with findings from other scenarios. The influence of equipment lifetime is intricately tied to other parameters’ values, including annual efficiency gains and demand growth.

Our findings are generally consistent with the real-life practices of several technology companies, such as Google and Meta. Before 2022, these companies have announced plans to extend the service life of their IT equipment up to six years by 2022 to achieve cost savings, and as a result of slowing performance improvements. As we did not consider financial aspects in our study, a direct comparison with these real-world practices is not feasible. However, an assumption of lower annual efficiency gains would lead to a longer optimal lifetime, making the results consistent with real-world practices. Given the complexity and variety of dependencies, it is unwarranted to make a general statement about the broader impact of equipment lifetime on total emissions for a given reporting year. Furthermore, it should be noted that our study focused only on the impact on a specific reporting year, without considering the long-term impact on total emissions.⁵⁵

As a third aspect of future uncertainty, we assumed identical values for manufacturing emissions of IT devices for the years 2020 and 2030, since there are no robust public models for how these emissions will evolve. This is particularly important for flash memory or SSDs, which are employed in great number and capacity to speed up specialized computations. Decreasing costs for SSDs are likely to increase the future use of this technology further.¹⁰ In light of the anticipated growth in the increased use of SSDs, Scope 3 emissions may accelerate more than expected. On the contrary, shifting manufacturing processes from Asia to Europe, where the GHG intensities of national energy mixes are substantially lower,^65,66 may oppose this trend. Such shifts are expected due to political and economic developments induced by the EU Chips Act, including the anticipated settlement of major global chip manufacturers, such as Intel and TSMC.^67,68,69 However, according to current plans, these semiconductor fabrication plants would not include the manufacturing of flash memory devices.

Challenges with GHG reporting

The trustworthiness of net-zero claims and the overall informative value of GHG emission reporting is debatable, which is mainly attributable to a couple of issues.

First, emissions from capital goods are recorded in the specific year when equipment is replaced or a building is constructed. Although emissions related to data center buildings constitute a significant portion of overall emissions, the GHG protocol only accounts for these emissions in the year of construction. As a result, construction-related impacts do not appear in the Scope 3 emissions for subsequent years when using the GHG protocol for reporting. In this context, considering the cumulative carbon footprint may provide a more comprehensive assessment of emissions.³⁷ Although it is mandatory to report market-based and location-based emissions, the absolute numbers do not sufficiently reflect efficiency-related aspects. As energy efficiency plays an essential long-term role in overall decarbonization and enables significant cost savings for operators, striving for efficiency is a crucial strategy to reduce Scope 2 emissions.⁷⁰ For this reason, initiatives for efficient data centers and ICT and best practices can serve as supplementary tools to improve current reporting standards by sector-specific key performance indicators and efficiency metrics by increasing comprehensiveness and transparency.^71,72,73

The validity of contractual instruments used to certify green energy is controversial as the environmental benefit of these instruments is only guaranteed through certain criteria, such as additionality.^74,75 Yet, the GHG protocol does not require that EACs fulfill additionality criteria, making climate neutrality claims in market-based reporting of Scope 2 emissions debatable. Particularly for data centers this is of importance, as the vast majority of emissions are Scope 2.

A similar issue is also observed for carbon offsets that are acquired to compensate emissions.⁷⁶ Consequently, acquired EACs and carbon offsets may not reflect actual positive changes in the climate impact. Policies and reporting standards must ensure that contractual instruments and carbon offsets are subject to stricter requirements, e.g., fulfilling a minimum of additionality criteria, to ensure trustworthiness and reliability. Such improvements are beneficial not only for the ICT sector but also for other industries. Furthermore, double counting challenges regarding the parallel application of market-based and location-based accounting are not solved.⁷⁷ This becomes even more relevant when considering Scope 3 emissions, which are essential for a holistic assessment.

Reporting Scope 3 emissions can be challenging due to low transparency across value chains, which is mainly attributable to the involvement of several actors. Particularly, the lack of available data and discrepancies across information sources on upstream emissions of IT equipment portrays a major challenge to identifying reduction potentials and accurately determining Scope 3 emissions.^10,78,79 This is also a hurdle for defining the amount of necessary carbon offsets, and a cost factor for reporting entities. Regulations and requirements for manufacturers to precisely determine and indicate product carbon footprints (PCFs) are particularly important and can serve as an incentive mechanism for establishing more climate-friendly products on the market.

We further highlight the necessity for further research by explaining our omissions within investigating the role of Scope 3 emissions. Our study focuses on a specific use case and does not consider the increasingly important factor of leased assets, such as data center space provided by colocation providers. This omission could result in an incomplete picture of the overall emissions impact, especially given the rising reliance on colocation services within the data center industry.

Furthermore, we excluded numerous categories of emissions, particularly in Scope 3, from our consideration. Including more categories, such as the construction of buildings, leads to a higher share of Scope 3 emissions. As in our study, the focus has historically been on operational emissions in data centers, overlooking the significant impact of construction-related emissions and other categories. Historically, studies, including ours, have primarily focused on operational emissions in data centers, often neglecting the significant impact of construction-related emissions and other categories. Although we account for electricity and IT-related Scope 3 emissions, we do not include construction-related emissions or other categories such as employee commuting. However, examining cumulative emissions provides additional valuable insights. Lin et al.³⁷ note that, in the first year of a data center’s operation, construction-related GHG emissions are roughly equivalent to those of IT equipment. By the end of 15 years, construction accounts for about 18% of cumulative embodied carbon. However, IT equipment and electricity-related Scope 3 emissions are the most significant contributors.

For a data center in the United States, cumulative Scope 3 emissions over a 15-year lifetime represent approximately 40% of total GHG emissions. Notably, over 84% of these Scope 3 emissions stem from capital goods and fuel and energy-related activities. Categories like employee commuting and business travel contribute a relatively minor share. Both the study by Lin et al.³⁷ and our study highlight that as renewable energy becomes more prevalent, the conversation is shifting toward embodied emissions. This emphasizes the importance of considering emissions from building data centers. Hence, with the increasing efficiency of IT hardware and the transition to renewable energy grids, solely relying on operational efficiencies for sustainability gains will become insufficient in achieving net-zero goals.⁸⁰

Conclusion

In summary, our findings and discussions lead to several recommendations for policymakers, which will increase transparency and sustainability in the data center industry and make progress toward achieving climate goals: (1) expand renewable energies and encourage the adoption and integration of renewable energy sources in data center operations, (2) support and invest in the development of efficient data center infrastructures and enhance overall energy efficiency within these facilities, (3) refine GHG reporting metrics tailored to data centers to improve the informative value of reporting structures and increase transparency, and (4) foster national and international collaborations between industry and research to improve data accessibility and transparency and strengthen supply chain accountability.

Limitations of the study

Our model is a highly simplified approach to estimate GHG emissions. Therefore, its overall representativeness is limited, as we do not consider different data center categories or applications, which would lead to multiple differences regarding technical and operational properties. Also, our initial assumption on the number of devices (6,000 servers) and the selection of only one server and storage system model does not reflect the reality, as the number of devices has a notable impact and systems highly differ in their average processing capacities, power consumption, and manufacturing emissions. Hence, the selection of different reference systems may change the results notably and will be a matter of further research.

Further limitations occur due to omission of numerous Scope 3 categories, high uncertainty of assumptions, the lack of available and transparent data for the IT equipment in data center use and operational aspects, and inconsistency across data sources.

An example for data inconsistency can be observed looking at PCF datasheets for servers and storage systems,^35,78 which differ strongly for different manufacturers. Additionally, the PCF sheets generated with PAIA (Product Attributes to Impact Algorithm)⁸¹ point to a potentially high standard deviation from the indicated PCF. Comparing the PCF datasheet provided by Dell⁷⁹ to a life cycle assessment (LCA) of the same server model¹⁰ reveals a notable gap between the indicated manufacturing PCFs. Although an LCA is considered more credible, the Dell PCF sheets are deliberately adopted to ensure data consistency for the selected servers and storage systems.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Tuğana Aslan (tugana.aslan@mikroelektronik.fraunhofer.de).

Materials availability

No materials were generated in this study.

Data and code availability

• •Data: This paper analyzes existing, publicly available data. All data sources all are indicated in the manuscript and key resources table.
• •Code: This paper does not report original code.
• •Additional Information: Any additional information required is available from the lead contact upon request.

Acknowledgments

This work was supported by Fraunhofer TALENTA, funded by Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e. V. Furthermore, we would like to thank Monica Ruge, Research Fab Microelectronics Germany (FMD), for her valuable contribution in designing the graphical abstract for this study.

Author contributions

Conceptualization, T.A., P.H., L.S., and A.G.; methodology, T.A., P.H., and L.S.; formal analysis, T.A.; writing – original draft, T.A.; writing – review & editing, T.A., P.H., L.S., A.G., N.F.N., and M.F.; supervision, P.H., L.S., N.F.N., and M.F.

Declaration of interests

The authors declare no conflict of interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the author(s) used ChatGPT/OpenAI exclusively to make linguistic improvements. After using this tool or service, the author(s) reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Electricity-specific Emission Factor, Baseline 2020	Icha and Lauf33	N/A
Electricity-specific Emission factor, Baseline 2030	European Commission34	N/A
Electricity-specific emission factor, 2030 climate protection goal	Fritsche and Greβ47	N/A
Electricity-specific emission factor, worst case scenario	Icha and Lauf33	N/A
Electricity-specific emission factor, market-based scenario	This paper	N/A
Electricity upstream emission factors	This paper	N/A
Average storage power	Dell Technologies36	N/A
Average server power	Dell Technologies35	N/A
PUE Baseline 2020	Hintemann38	N/A
PUE Baseline 2030	This paper	N/A
PUE best case scenario	This paper	N/A
PUE worst case scenario	This paper	N/A
Workload demand	Cisco42	N/A
Storage demand	Mellor,43 This paper	N/A
Average server manufacturing and transportation emissions	Dell Technolologies35	N/A
Average storage manufacturing and transportation emissions	Dell Technologies36	N/A
Utilization baseline	Bio by Deloitte and Fraunhofer IZM40	N/A
Electricity upstream emissions factors	This paper	N/A
Utilization scenarios	This paper	N/A
Equipment manufacturing scenarios	This paper	N/A
Efficiency gains best case scenario	Andrae and Edler,41 This paper	N/A
Efficiency gains worst case scenario	This paper	N/A
Equipment lifetime baseline	Dell Technologies35,36	N/A
Equipment lifetime scenarios	Moss55	N/A
Software and algorithms
Excel	Microsoft	–

Method details

This section describes the steps of our method in more detail.

Definition of the GHG inventory

In the first step we examined studies specifying data center value chains and existing models assessing data center electricity consumption and GHG emissions to identify the main emission drivers across data center value chains.^{2,6,21,23,25,39,82} As the GHG protocol does not provide a predefined GHG inventory for data centers, we assigned the identified aspects to the categories Scope 1, Scope 2, and Scope 3.

The most significant proportion of data center GHG emissions are attributable to electricity consumption and the manufacturing of IT hardware. Following the GHG Protocol we assigned direct emissions from on-site power generation to Scope 1. However, on-site electricity generation is not covered in this study, as in case of assuming 100% renewables the results are identical to the 100% renewable market-based consideration. Due to the minor role in comparison to CO₂e emissions from grid electricity consumption, we do not consider Scope 1 emissions from fossil based on-site emergency power generators.⁸³ We consider indirect emissions from the generation of acquired electricity in Scope 2, and indirect emissions upstream to electricity generation as well as upstream chains of IT equipment (manufacturing and transportation) in Scope 3. Due to the minor role in overall emissions, we do not consider aspects such as employee commuting, use of paper, and the EoL treatment of IT devices.^10,84 Also, we did not consider upstream (Scope 3) emissions related to data center building construction, as our main focus lies in data center operation related emissions. As our study examines the level of single data centers, we excluded many categories of corporate emissions in Scope 1 and Scope 3. Other studies differentiate between operational and embodies emissions.¹ These categories, are equivalent to Scope 2 and Scope 3 emissions in our study, respectively.

Existing models on data center electricity consumption and GHG emissions

We examined various models and studies on data center electricity consumption and GHG emissions.^{1,2,4,6,21,25,26,59,85,86} During our research, we identified three approaches: top-down, bottom-up, and extrapolation. While top-down models are mostly based on simplified general assumptions for a whole system, bottom-up approaches consider data on component level first and then aggregate. Extrapolation uses baseline values from top-down or bottom-up approaches and apply growth rates for future estimations.⁸⁶ Due to the level of detail, bottom-up approaches are considered more reliable.²⁵

Bottom-up models break total power consumption down into the sum of server, storage devices, network, and infrastructure power consumption.^{4,21,26,85,87} While some models consider infrastructure one category,^4,85 others distinguish among three subcategories: uninterruptible power supply (UPS), cooling, and other infrastructure.^21,26,87

To calculate infrastructure power consumption following the bottom-up approach, specific data of all infrastructural components is aggregated. However, in this case, a common approach is to simplify the calculation with the help of the PUE factor that describes the ratio of IT power consumption to data center total electricity consumption³² (The lowest possible value for PUE equals 1, meaning that the infrastructure power consumption equals zero. Currently, highly efficient data centers can have PUE around 1.09⁵³). IT power consumption corresponds to the sum of server, storage and network power consumption. Typically, the power consumption of each category is calculated by the number of devices and power consumption.

After the total electricity consumption of the data center is determined, GHG emissions are calculated by the intermediary variable, electricity-specific emission factor, indicating the GHG intensity per kWh electricity for a certain electricity mix.⁴⁸

Almost all of existing studies solely focus on data center operations and do not consider embodied emissions of IT equipment. Nevertheless, upstream emissions of IT equipment are provided in product data sheets of several manufacturers.^35,78,88 Also, data on emissions upstream electricity generation is available.⁴⁴

Scenario modelling for data center GHG emissions

To determine total power consumption and GHG emissions, we combine a bottom-up approach, similar to Stobbe et al.,³⁹ with top-down principles and extrapolation for various assumptions, as seen in the works of Andrae.^23,41 Our aim is to maintain flexibility in the model to accommodate variations in assumptions regarding demand growth and efficiency gains. Overall, while our study draws on methodological insights from all the referenced studies, it does not adhere strictly to any single approach. We assume a data center, in which all devices acquired at the beginning of the year 2019 (t=0). Initially, the number of servers and storage systems are 6,000 and 1,214 respectively. The number of servers corresponds to the average server number in a large German data center.³¹ Using our reference devices, we calculated the initial number of storage systems by the respective share of storage systems in a data center total electricity consumption scheme.³⁸

Existing studies suggest network power consumption accounts for 4-10% of the total IT power consumption.^21,89 Thus, we presumed that network power consumption equals 10% of the total power consumption of servers and storage devices. Due to the variety of materials across networks and the lack of data, we neglected upstream emissions of network equipment.

We adopted only a single model for servers and storage system from Dell Technologies to keep the model as simple and generic as possible. Our selection of Dell Technologies is due to their market leadership⁹⁰ and available PCF datasheets for the respective products. Based on the screening of technical details regarding processing capacities and device power consumption, we selected average models. We retrieved all data on upstream emissions for servers and storage devices from datasheets provided by Dell Technologies.^35,36,91 As for the Baseline projection 2030, we did not assume a change in the manufacturing and transportation emissions of servers and storage because of lack of data on how these emissions will evolve in the future. However, we considered sensitivity assumptions of 75% increase and 75% decrease of manufacturing emissions to see the potential impacts.

Unless specified otherwise, we presumed the data center’s electricity demand is either met fully by a green tariff or a 100% renewable PPA.⁴⁸ For location-based calculations, we assume an average electricity grid mix for Germany, as our reference data center is Germany-based.

The model is generally designed for a fixed device lifetime of 4 years, which is also indicated in the product datasheets. Therefore, we do not consider device lifetime a variable parameter. Still, we considered alternative lifetimes of 3 and 6 years (see also Discussion). Similarly, server and storage power consumption are based on a fixed usage profile with an overall average utilization rate. For this, we defined three utilization conditions for servers and storage devices. We assume that higher utilization increases processing capacity as well as power consumption of devices.

Our approach comprises two steps: (1) determining the number of devices, and (2) calculating electricity consumption and GHG emissions.

The number of devices can be either directly specified or calculated by predefining workload/storage demand and devices’ capacity. In both cases, it is necessary for future projections to introduce two additional aspects: demand growth and performance improvement. These allow us to consider improvements/deterioration that occur in the future.

The approach considers the acquisition of new equipment with increasing demand and the replacement of equipment at the end of the device lifetime.

We used the following variables to forecast the number of servers/storage devices:

• (1)Initial demand ($d$) (Bytes), denoting the total initial workload/storage demand
• (2)Initial processing capacity ($c$) (Bytes), denoting the initial workload/storage capacity per device
• (3)Annual demand growth rate ($g$) (%), denoting the growth of total workload/storage demand
• (4)Annual efficiency gains ($m$) (%), denoting the increase in workload/storage capacity per device

The initial number of devices is determined by ($d/c$). These are acquired at t₀ (01/01/2019). Each year, the demand increases by the growth rate ($g$). At t₁, the beginning of the following year, the total demand equals $d\left(1+g\right)$. Since the initial demand ($d$) is already covered by devices acquired in 2019, only the newly added demand in the year t₁, $d\left(1+g\right)-d$, is considered. To meet this newly added demand, new devices are acquired. Compared to 2019, these devices now have a higher workload processing capacity equal $c\left(1+m\right)\text{.}$ Accordingly,

$$\left(d\left(1+g\right)-d\right)/c\left(1+m\right)=numberofdevicespurchasedin{t}_1\text{.}$$

The same logic applies for t₂−in which

$$\frac{d\left({\left(1+g\right)}^2-\left(1+g\right)\right)}{c\ast {\left(1+m\right)}^2}=numberofdevciespurchasedin{t}_2\text{.}$$

devices are required– and t₃.

By t₄, the devices purchased in t₀ have completed their service life of 4 years. Accordingly, (d/c) devices are decommissioned from the installed base, and the demand deficit caused by their withdrawal (d) is considered in addition to the newly added demand in that year. Hence,

$$d\left(({1+g)}^4-{\left(1+g\right)}^3\right)+d=totaldemanddeficitin{t}_4\text{.}$$

This demand is met by devices with a processing capacity of $c{\left(1+m\right)}^4$. This logic continues until t₁₁ (2030).

The number of installed devices in a year is determined by aggregating the number of acquired devices per year over the service life. Hence, for a 4-year lifetime,

$$t_8+t_9+t_{10}+t_{11}=numberofinstalleddevicesin2030\left(t_{11}\right)\text{.}$$

After defining the number of servers and storage devices, we calculated data center electricity consumption and respective GHG emissions. Here, we first multiplied the number and power consumption of servers and storage devices, respectively; which we later aggregated.

$$servertotal\left(kWh/a\right)=\left(\left(\#installedsevers\right)\times \left(serverpower\left(W\right)\right)\times 8760\left(h/a\right)\right))/1000$$

$$storagetotal\left(kWh/a\right)=\left(\left(\#installedstorage\right)\times \left(storagepower\left(W\right)\right)\times 8760\left(h/a\right)\right))/1000$$

In the next step, we quantified total data center electricity consumption by adding network and infrastructure power consumption. Using the electricity-specific emissions factor, we lastly defined total GHG emissions from data center operation (Scope 2).

$$totaldatacenterenergyuse\left(kWh/a\right)=\left[\left(servertotal+storagetotal\right)\times \left(1+10\%\right)\times \left(PUE\right)\right]$$

$$Scope2emissions\left(tCO2e\right)=\left(totaldatacenterenergyuse\left(kWh/a\right)\right)\times specificemissionfactors\left(kg\frac{CO2e}{kWh}\right)/1000$$

For the electricity-specific emission factor, we either used average electricity specific emission factors (location-based), or an emission factor of 0 kg CO₂e/kWh as a green tariff/100% renewable PPA (market-based).

For Scope 3, we added up upstream emissions of IT equipment and upstream emissions of electricity generation. We adopted manufacturing and transportation emissions of servers and storage from the datasheets of our reference devices. For electricity upstream emissions, we defined emission factors for the three different electricity mixes we used – location-based 2020, location-based 2030, and green-tariff.^45,46,49 For these we considered different compositions of energy sources and determined the respective upstream emission factors.

$$Scope3emissions\left(tCO2e\right)=\left[\left(totaldatacenter\right)\times \left(emissionfactorelectricityupstream\right)\right]+\left[\left(\#acquiredservers\right)\times \left(servermanufacturing+transportation\right)\right]+\left[\left(\#acquiredstorage\right)\times \left(storagemanufacturing+transportation\right)\right]/1000$$

All our calculations were conducted through implementing a model in Microsoft Excel, which simulates outputs of electricity consumption of different categories, Scope 2 and Scope 3 emissions depending on a set of adjustable input parameters.

Based on literature review, we identified feasible values for the aforementioned parameters. To ensure high data quality, we sourced data either from official sources, such as the German Environmental Agency (UBA), or from assumptions used in carious peer-reviewed papers. For our baseline scenario referring to the reference year 2020, we identified representative values for Germany. For 2030, we considered multiple anticipated developments and options concerning operational management. Using these values, we established one Baseline scenario for 2020 and 2030, respectively. These pool the assumptions that are determined to be most likely or most reliable.

These explore the effects of alternative developments and the impact of single parameters, we further defined sensitivity assumptions for the following variables: (1) electricity-specific emission factor, (2) utilization, (3) PUE, (4) efficiency gains, and (5) manufacturing emissions of IT equipment. All calculations are modeled with Microsoft Excel.

Quantification and statistical analysis

Microsoft Excel was used to calculate the relative differences of the sensitivity scenarios to the Baseline 2030 for Figure 4. Underlying data for Figure 4 is available in the Supporting Information S4 and S5. Apart from this calculation this paper does not include statistical analysis or quantification.

Published: December 21, 2024