The impact of computing infrastructure on carbon emissions: an empirical study based on china national supercomputing center

Digitalization is viewed as a potential solution to environmental sustainability issues. However, existing researches suggest that the environmental impact of digital technology is uncertain. This study focuses on the National Supercomputing Center (NSC) in China, a large-scale computing infrastructure, and expands the correlation between digital infrastructure and green development from a computing perspective. Based on the synthetic control method, we select non-supercomputing cities as the control group and assign appropriate weighting. Through the fitting of a synthetic control group (refer to as the synthetic city) with similar characteristics, the analysis is conducted to compare carbon emissions (CO2 emission) between NSC city and the synthetic city. The empirical results show that the NSC may worsen regional CO2 emissions, and this result still holds true after a series of robustness tests. Mechanism examinations show that the NSC does not exhibit significant composition effect (energy structure improvement) and technology effect (green technology innovation), while scale effect (increase in energy demand and consumption) dominate the NSC’s carbon emissions. Based on these findings, we consider that in addition to improving the energy efficiency of supercomputing centers, the adoption of cleaner renewable energy and the promotion of knowledge spillovers are crucial for achieving a green transformation for computing infrastructure.


Introduction
With the control of greenhouse gas emissions becoming a global consensus, major countries around the world have announced carbon neutrality goals and put forward strict emission reduction requirements.The scale and impact of international green initiatives are increasing (Arbolino et al 2018, Qiu et al 2021, IEA 2023).At the same time, digital transformation is continuously expanding worldwide, with digital technology like 5G, big data, and artificial intelligence deeply integrated with the real economy (Fan et al 2022, Wang et al 2022a).Due to the continuous growth of data volume, computing infrastructure has gradually become an important foundation for scientific development, social production, and governance (Yu et al 2016, Allen 2022).Exploring the use of digital tools to help achieve low-carbon goals has become an important issue for future social development under the goal of the Paris Agreement.CO2 Emissions in 2022 released by the International Energy Agency shows that global CO 2 emissions related to energy increased by less than 1%, which is below the level of concern.However, the report also highlights that the electricity and heating sectors are responsible for the largest rise in emissions, contributing to 261 million tons of carbon emissions (IEA 2023).The digital industry itself is a field with high energy consumption and carbon footprint.According to Galperova and Mazurova's (2019) research, data centers and networks accounted for 3% and 1%, respectively, of global electricity consumption in 2015.Joppa and Herweijer (2018) estimated that the greenhouse gas emissions arising from digital technology would rise from 2.5% in 2013 to 4% in 2020.Jones (2018) forecasts that the Information and Communication Technology (ICT) industry will comprise 8%-21% of the world's electricity demand by 2030.Given the on the local power supply system and generates a significant amount of heat, resulting in related environmental issues.As such, there is a pressing need to prioritize the implementation of environmentally-friendly strategies in the development and operation of the computing infrastructure.Given the government's intervention in the construction of NSC, it can be viewed as a quasi-natural experiment, providing a prerequisite for examining the impact of large-scale computing infrastructure construction on the local environment within a causal inference framework.
Prior to the investigation, some questions necessitate elucidation, encompassing whether NSC will increase CO 2 emissions, what specific pathways contribute to the impact of NSC on regional carbon emissions, and which mechanisms play a predominant role in this regard.The resolution of these inquiries will facilitate this study to make contributions in the following aspects within the existing body of knowledge: 1. Existing research on the environmental impact of digitalization either approaches it from the perspectives of finance, trade, innovation, etc (Perrons 2021, Bianchini et  .However, there is a notable research gap when it comes to investigating the environmental consequences of constructing computational infrastructure.This study provides empirical evidence on the carbon emissions impact of developing countries' supercomputing center construction, as a supplement to the literature on the impact of digital infrastructure on green development. 2. The Environmental Kuznets Curve (EKC) theory, which divides the scale effect, structure effect, and technology effect, has been applied in studies evaluating the environmental impact of different socioeconomic activities due to its high interpretability (Grossman and Krueger 1991, Grether et al 2009, Zhang 2012, Ahmad et al 2020, Wang et al 2022a).To clarify the environmental effect of NSC, this study converts the three effects into the energy context, emphasizing the dominant role of scale effect in increasing carbon emissions.
3. Regarding the impact of large-scale research infrastructure (LSRI) on socio-economics, the RI-PATHS workshops designated Energy and Waste issues as specific dimensions for assessing the influence on environmental sustainability.(Griniece et al 2019, Kroll et al 2019).Although there have been some discussions on the green sustainability of LSRIs (Fu et al 2010, Beck andCharitos 2021), there is a lack of systematic evaluation of LSRIs' impact on regional environment.This study places existing discussions on LSRIs' economic benefits in the context of a low-carbon economy and focuses on NSC's impact on the carbon emissions in cities.

Theoretical basis and mechanism analysis
2.1.Scale, composition, and technology effect As global environmental quality continues to deteriorate, people's concern about environmental issues is increasing.Since the 1970s, discussions on sustainable development have been on the rise.In 1991, Grossman and Krueger investigated the relationship between income and environmental quality indicators, including SO 2 and smoke.Their study revealed an inverse U-shaped correlation between pollutant emissions and income (Grossman and Krueger 1991).This results give confidence to those who were trying to reconcile the relationship between environmental protection and economic development.Because it is similar to the Kuznets curve, it is also known as the Environmental Kuznets Curve.The EKC phenomenon is seen as a result of the natural evolution of economic scale and structure, and can be divided into three different paths: scale effect, technology effect and composition effect (Grether et al 2009, Cole and Elliott 2003, Zhang 2012, Hua et al 2018, Ahmad et al 2020).Specifically, in the early stages of development, the economic structure shifts from agriculture to energy-intensive heavy industry, which increases pollutant emissions.During this period, economic development had a negative impact on environmental quality mainly in two ways: on one hand, economic growth necessitates augmented input, which increases the use of resources.On the other hand, more output increased the emission of harmful substances, leading to environmental degradation.This is also seen as the scale effect of economic growth on the environment.With the improvement of development level, the economic structure would change, shifting from energy-intensive heavy industry to low-pollution service and knowledge-intensive industries, and the level of pollutant emissions per unit output will decrease, resulting in an improvement in environmental quality.This effect is also known as the composition effect.In addition, during the process of economic transformation, the increase in research and development (R&D) spending has promoted technological progress, which has improved resource utilization efficiency, increased output per unit of input, and weakened the impact of production activities on the natural environment.This is also known as the technology effect, leading to a reduction in the negative impact on the environment caused by economic development through technological progress.
Studying the driving factors of carbon emissions is of great significance in clarifying the path for industrial emissions reduction and providing targeted measures.The classification of the above three effects has high explanatory power and has been applied in research on the environmental impact of various socio-economic activities, covering economic development and international trade.For instance, Zhang (2012) employed a structural decomposition analysis method to evaluate the scale, composition, and technology effects of China's mainland trade carbon emissions from 1987 to 2007.Ahmad et al (2020) constructed a theoretical framework based on three channels of action to explore the impact of natural resources, technological innovation, and economic growth on ecological footprints in emerging economies.This framework also involves public sector spending and industrial production, as Grether (2009) decomposed the changes in manufacturing emissions of 62 countries into scale, composition, and technology effects.Hua et al (2018) examined whether education and R&D spending could reduce air pollution by promoting human capital accumulation (composition effect) and adopting clean technology (technology effect), using data from 284 Chinese cities. Wang et al (2022a), whose research is closely related to the theme of our study, explored the impact of the digital economy on CO 2 emissions based on three intermediate paths of scale effect, composition effect, and technology effect using provincial data from China.Based on the mature research framework, we will examine the impact of NSC construction on carbon emissions through these three mechanisms.

Scale effect-Increase in energy consumption scale
The scale effect can be understood as an increase in production scale without changing technology and economic structure.Previous research has mainly used indicators related to economic development to measure the scale effect.Cole (2003) considered the applicability of GDP/km2 and per capita GDP as indicators for the scale effect.Similarly, Ahmad et al (2020) used per capita GDP to reflect economic growth under the scale effect.Liobikienė and Butkus (2019) attributed GDP, trade, foreign direct investment, urbanization, and other factors to the scale effect.Since the aim of this study is to examine the impact of NSC on local environment, we extend the representation of the scale effect to energy demand and consumption, using the increase in electricity consumption to reflect the increase in NSC operating scale.
The NSC enhances the city's energy consumption scale through two main aspects.
(1) Direct impact.The operation of the supercomputing center, which requires millions of processors for large-scale parallel computing, along with high-speed network connections and abundant storage systems for data transmission and storage, consumes a significant amount of electricity.The substantial electricity consumption is not only attributed to large-scale computations but also to maintaining optimal environmental conditions, such as suitable temperature and humidity, to ensure the normal functioning of the computers.As a considerable amount of electrical energy is eventually converted into heat, the cooling equipment in the supercomputing center accounts for a significant proportion of the total power consumption, and the operation of the cooling systems also relies on energy support.Additionally, other equipment in the data center, including fire protection, monitoring, and emergency lighting, contributes to a certain amount of energy consumption.(2) Environmental rebound effect.The rebound effect refers to the phenomenon where environmental protection measures or reduced utilization of natural resources unintentionally lead to the occurrence or exacerbation of other environmental problems (Hertwich 2005).The development of digital technology is believed to enhance productivity.However, the improvement in resource (energy) efficiency may stimulate further expansion of energy demand (Pohl et al 2019, Wu et al 2023, Bieser et al 2023).The use of supercomputers shortens the research and development cycle, improves the efficiency of knowledge production and technological innovation, which may further increase the computational requirements in various fields, including science and industry.This, in turn, intensifies the operation of supercomputing centers and contributes to the overall energy consumption in the region.If the improvement in the city's energy structure, is limited, the increase in energy consumption will ultimately lead to intensified carbon emissions.Based on this analysis, the following hypothesis is proposed: H1: The NSC exacerbates urban CO 2 emissions by increasing energy consumption scale.

Technology effect-the development of green technology
The term technology effect generally refers to the increase in productivity resulting from technological advances and the development of clean technology.The measurement of technology effect typically centers around these two aspects.Sugiawan, Managi (2016) used total factor productivity to characterize the adoption of new technology.Liobikienė and Butkus (2019) measured technology effect using energy efficiency and found that a 1% increase in the latter results in a 0.135% decrease in greenhouse gas emissions.In other studies, scholars have measured technology effect through patent.For instance, Ahmad et al (2020) characterized technological progress under this framework using the number of patent applications, while Wang et al (2022a) measured green technological innovation by selecting the quantity of green patent application.Given that the increase in productivity may lead to rebound effect and thus have a more ambiguous impact on the environment (Pohl et al 2019, Wu et al 2023).Therefore, this study characterizes technology effect using green innovation.
In the context of this study, the technology effect can be understood as the impact of NSC on carbon emissions through the promotion effect of technology, especially green technology.This is mainly reflected in the following ways: (1) Increase in green demand.Due to the high energy consumption characteristics (scale effect) of the supercomputing center, green energy-saving technologies need to be adopted to improve energy efficiency.With the continuous growth of data storage and processing demands, it is necessary to develop more efficient and energy-saving technologies and algorithms to reduce the pressure on the energy system caused by increasing demand.In the background of the dual carbon targets (carbon peaking and carbon neutrality), green development has become an important indicator for local government performance assessment in China (Su et al 2022).Local governments have sufficient motivation to accelerate the green transformation of supercomputing centers.(2) Direct technical support.The supercomputing center can play the role of an innovation laboratory and testing base in the development of green technology.Researchers and engineers can validate and optimize green technologies in the environment of the supercomputing center, exploring new energy-saving technologies and sustainable solutions.Moreover, the supercomputing center actively participates in cooperation and exchanges in green technology innovation, collaborating with academic institutions, industries, and government departments, among other partners, in R&D project.Taking the Tianjin NSC as an example, it relies on supercomputing, cloud computing, big data centers, and artificial intelligence open-source platforms to collaborate with universities and research institutions in areas like lowcarbon and smart cities, conducting scientific research cooperation to help companies implement the national strategy of carbon peaking and carbon neutrality through technological innovation, process optimization, and improved energy efficiency.(3) Indirect knowledge spillover.As a center for scientific research and technological innovation, NSC not only generates abundant knowledge and technology in the research field but also disseminates this knowledge and technology to related industries through various means, such as knowledge spillover formed through innovation clusters or talent mobility.The formation and development of such an innovation ecosystem contribute to attracting more high-tech industries, thereby reducing the proportion of polluting industries in the economic sector and promoting the aggregation and development of green and lowcarbon industries.Based on the above analysis, the hypothese is proposed as follows: H2: The NSC reduces urban CO 2 emissions by promoting the development of green technology.

Composition effect-optimization of energy structure
Existing literatures characterized the composition effect from two perspectives.One approach emphasizes the impact of the shift in industry from labor-intensive to capital-intensive on the environment, based on the intensity of factor inputs.The economic structure is often measured by the capital-labor ratio (Cole and Elliott 2003).Another approach places more emphasis on the industrial structure, typically reflected by the proportion of industry in GDP (Hua et al 2018, Luo et al 2022, Wu et al 2023), or changes in the proportion of clean and polluting industries within the industrial sector (Grether et al 2009, Shahzad et al 2022).This study focuses on energy structure, as it significantly affects the impact of energy consumption on the environment.
The composition effect originally referred to the structural changes in the economic sector during the development, which would affect the role of growth in the environment change.We extend the concept of composition effect to the energy structure because the environmental costs of supercomputing centers largely depend on the attributes of the energy they use (Allen 2022).Whether it is the supercomputing facility SURF in the Netherlands or the Max Planck Institute for Astronomy in Heidelberg, Germany, they have less carbon footprint due to the adoption of more clean energy sources (Jahnke et al 2020, Van der Tak et al 2021).It is important to emphasize that this study is not concerned with static energy structures but focuses on the dynamic impact of NSC on urban energy structure.Firstly, through advanced computing capabilities and simulation techniques, NSC can conduct large-scale energy system modeling and optimization analysis, providing scientific basis for urban energy planning and decision-making.By simulating and predicting the performance of different energy structures and their environmental impacts, supercomputing centers can evaluate the potential of various renewable and clean energy sources, helping decision-makers choose appropriate energy structures.Secondly, due to the technology effect, NSC can contribute to energy technology research and development.Supercomputing centers can simulate and optimize the performance of new energy technologies, providing solutions for renewable energy generation, energy storage, smart grids, and other aspects, promoting the application and popularization of new energy, thereby increasing the proportion of clean energy in the energy structure.Existing research has revealed the impact of digital infrastructure construction on energy structure transformation (Fan et al 2022), which can optimize energy allocation through direct effect and have multiple indirect impact paths, including the green production level and green finance level.In summary, there is reason to propose the following hypothesis: H3: The NSC reduces urban CO 2 emissions through the optimization of energy structure.
The research mechanism diagram (figure 1) illustrating the three hypotheses is presented below:

Institutional background
The National Supercomputing Centers are data computing organizations approved by the Chinese Ministry of Science and Technology.Since 2009, NSCs have been established in various locations such as Tianjin, Shenzhen, Changsha, Guangzhou, Jinan, Wuxi, and Zhengzhou with the support of the Ministry of Science and Technology.Currently, NSCs have been widely applied in various fields such as climate, energy, medicine, materials, aviation and aerospace, artificial intelligence, and smart cities.Among these NSC cities, Tianjin and Wuxi are the most representative cities.The former was the first city in China to establish NSC and also owns the first petaflop supercomputer in China -'Tianhe-1'.This supercomputer was ranked first in the TOP500 list in 2010.The latter is currently the city where the bestperforming supercomputer in China, 'Sunway TaihuLight', is located.This supercomputer ranked first in the TOP500 list in 2016 and 2017.As of the latest TOP500 list in November 2022 (https://www.top500.org/lists/top500/2022/11/), 'Sunway TaihuLight' still ranks seventh with a performance of 93.01 PFlop/s, maintaining its position as the top supercomputer in China.It is worth mentioning that the ranking of these two supercomputers in the Green500 list (https://www.top500.org/lists/green500/2022/11/) is not as impressive as in the TOP500 list.'Sunway TaihuLight' is only ranked 72nd, while 'Tianhe-1' is ranked 278th.Figure 2 further illustrates the spatial distribution of NSC cities and the carbon emissions in 2020.It can be seen that the CO 2 emissions in Tianjin and Wuxi are at relatively high levels, indicating the necessity of examining NSCs' environmental impact on the city.

Model
This study focuses on the impact of NSC construction on CO 2 emissions in Tianjin and Wuxi.Due to the limited number of treatment groups and the potential influence of insufficient sample size on estimation accuracy, the synthetic control method (SCM) proposed by Abadie et al is used in this study to evaluate the environmental The core idea of this method is the 'counterfactual framework', which first assumes that a certain city is not affected by NSC, and then constructs a 'counterfactual' reference group by weighted averaging data from other cities.This reference group is then compared with the actual data that has been intervened by the NSC to estimate the actual impact of the shock by testing the difference between the two, which is the 'treatment effect'.In addition, the synthetic control method is a non-parametric method that determines the weight of the control group based on data, which helps to reduce subjective judgments.
This study assumes that the CO 2 emissions of K+1 cities during the period t ä [1, T] are collected.The city i (treatment group) starts to construct and operate the NSC from T0 (1 T0 T), and the other K cities do not carry out NSC construction (control group).
In equation (2), * t w represents the synthetic control contribution rate of the control group to the treatment group, and k is the serial number of the control group individual.

Mechanism variable
According to the research framework diagram in section 2.2, we have identified three types of mechanism variables: (1) Energy consumption.Given that electricity consumption is the largest energy consumer in supercomputer centers, this study selects city industrial electricity consumption as a proxy variable for energy demand and consumption.
(2) Energy structure.Since energy balance sheets are not publicly available at the prefectural level, this study adopts the approach of Ren et al (2020), using data on natural gas, liquefied petroleum gas, and total social electricity consumption published in the China Urban Statistics Yearbook, and converts them to coal consumption using standard coal conversion coefficients (natural gas: 13.3 tons of standard coal/10,000 cubic meters; liquefied petroleum gas: 1.7143 tons of standard coal/ton; electricity consumption: 1.229 tons of standard coal/10,000 kilowatt-hours), and uses the ratio of total social electricity consumption (converted) to the sum of natural gas (converted), liquefied petroleum gas (converted), and total social electricity consumption (converted) as the proportion of coal consumption in urban energy consumption.

Sample and data processing
(1) Sample period.The using of the SCM requires a sufficiently large sample period prior to the event occurrence to better fit the developmental characteristics of the treated cities and enhance the reliability and explanatory power of the study.As the construction of NSC in Tianjin (2009) and Wuxi (2016) occurred at different times, the sample periods also differ.For the Wuxi case, this study selects a time span from 2003 to 2020, while for Tianjin, the time span is set from 2003 to 2013 to ensure that the period before the event occurrence is longer than the period after it.
(2) Control group.To enhance comparability between the treatment and control groups, 35 large and medium-sized cities are chosen as the control group (Chen 2015), including 4 municipalities directly under the central government, 26 provincial capital cities, and 5 single-planned cities.To eliminate interference from other NSCs construction, supercomputing center cities, except for the treatment group itself, are further removed from the 35 cities.For example, when fitting Wuxi, cities including Tianjin, Guangzhou, Shenzhen, Changsha, Jinan, and Zhengzhou are removed from the control group.
(3) The logarithmic transformation is conducted on indicators including per capita GDP and total population at the end of the year.
A small number of missing values for some control variables are handled by interpolation or replaced by the mean.Based on the above, the descriptive statistics of the variables are shown in table 1.

Impact of NSC on carbon emissions
The comparison between the fitted values of the control variables and the actual urban data is presented in table 2. Panel A demonstrates that the synthetic Wuxi is primarily composed of four cities, namely Hohhot (0.460), Hangzhou (0.311), Chengdu (0.131), and Urumqi (0.099), distributed across China's inland and coastal areas, as well as its northern and southern regions.The Carbon emission indicates that the difference in carbon emissions between actual and synthetic Wuxi is relatively small before the establishment of NSC.Although actual Wuxi has a higher proportion of secondary industries and a larger population, which may result in higher carbon emissions, its economic scale is smaller than synthetic Wuxi, leading to less carbon emissions.Thus, the small carbon emission gap between actual and synthetic Wuxi is reasonable.Panel B shows that synthetic Tianjin is mainly composed of four cities, namely Chengdu (0.475), Shanghai (0.215), Chongqing (0.170), and Hohhot (0.140), with diverse geographical distributions.Furthermore, while actual Tianjin has a higher proportion of secondary industries and a larger population, which may result in higher carbon emissions, its population is smaller, which to some extent led to a small difference in carbon emissions between actual and synthetic Tianjin prior to the establishment of NSC.As illustrated in figure 3, In the graph where Carbon emission (Wuxi) serves as the vertical axis, the solid line represents the actual CO 2 emission of Wuxi, while the dotted line denotes the CO 2 emission of the synthetic city.The vertical dotted line indicates Wuxi's NSC establishment in 2016, and the trend of CO 2 emission prior to the shock is presented on the left side of the dotted line.It can be observed that before 2016, the CO 2 emission levels of the synthetic and actual Wuxi are comparable.However, in the second year of NSC construction, the carbon emission of actual Wuxi increases significantly compared to the synthetic Wuxi.
Similarly, in the graph with Carbon emission (Tianjin) as the vertical axis, the synthetic Tianjin's CO 2 emission followes the pattern of the actual Tianjin before 2009.After the intervention, actual Tianjin and synthetic Tianjin show a divergence, and the former's increase is significantly higher than the latter's.The treatment effect of NSC is further illustrated in the lower part of figure 3.For Wuxi, the carbon emissions of actual city and synthetic city begin to diverge in 2017.The treatment effect of NSC becomes increasingly apparent, and the trend of the curve flattens in 2020.For Tianjin, before the intervention, the difference between the actual Tianjin and synthetic Tianjin's carbon emissions fluctuates around zero.However, after 2009, this difference expand significantly.The results preliminarily indicate that the establishment of NSC in both Wuxi and Tianjin has to some extent exacerbated the increase of urban carbon emissions.
Table 3 presents the specific values of carbon emissions for the treatment group and the control group, with the absolute effect being the difference between the two groups and the relative effect being the ratio of the difference to the average (%).From the data characteristics, it can be observed that both the absolute effect and the relative effect of the NSC construction in Tianjin are higher than those in Wuxi.The absolute effect in Wuxi reaches its peak of 6.517 in the fourth year after the shock, while in Tianjin, this indicator reaches 25.132 in the fifth year (2013) and does not show any signs of slowing down, indicating that the negative environmental impact of the NSC construction is more severe in Tianjin.This may be due to the fact that the construction of the Wuxi National Supercomputing Center occurred later than Tianjin, and in the year preceding its construction  (2015), the revised Environmental Protection Law of China was implemented, and China joined the Paris Agreement the following year.Compared with the first decade of the 21th century, China's actions in sustainable development have been more frequent.This may have led to the construction of the Wuxi supercomputer center being greener and more energy-efficient, thereby relatively reducing the carbon emissions generated by its construction and operation.

Placebo test
This study employes the method proposed by Abadie et al (2010) to verify the robustness of the results.This method assumes that all cities in the control group have constructed NSCs in the same year as Wuxi or Tianjin, and uses the synthetic control method to construct synthetic objects for each city.The difference between the synthetic objects and the actual carbon emissions of the cities is calculated as the effect of implementing 'NSC'.
Then, the actual treatment effect of Wuxi and Tianjin are compared with the 'NSC' effect under the assumption of the control group.If the former is significantly higher than the latter, it indicates that the NSC construction in Wuxi and Tianjin effectively increase the local carbon emissions, and the results previously are robust.It should be noted that this placebo test requires a good fit between the control group and its synthetic objects before the shock.To achieve this, we set the 10 times of the root-mean-square prediction error (RMSPE) value before the NSC construction in Wuxi and Tianjin as the threshold and eliminate control group cities that exceed this threshold.Then, ranking tests are conducted.
As depicted in figure 4, the solid line represents the treatment effects of NSC construction in Wuxi and Tianjin, while the dashed lines represent the 'treatment effects' of NSC construction in other cities.It is evident that the treatment effects in Wuxi and Tianjin are significantly higher than other cities.Specifically, under the constraint of the RMSPE threshold, for the placebo test in Wuxi, we exclude Shanghai.The results of the placebo test indicate that the mean treatment effect of Wuxi ranks second among the remaining cities for the control group, with an occurrence probability of approximately 6.89% (2/29) under the randomization of control group cities.For Tianjin, we exclude Shanghai, Chongqing, and Yinchuan.Under this premise, the mean treatment effect of Tianjin is ranked first among the remaining cities in the control group, with a probability of approximately 3.7% (1/27).In other words, the treatment effect of Tianjin is statistically significant at the 5% level.The above results indicate that the findings in section 5.1 are highly robust.

Other ways
Apart from the placebo test, this study also conducts the following tests to ensure the robustness and reliability of the research results: (1) Change control group.The size of the treatment effect is heavily influenced by the carbon emissions characteristics of the control group cities.We further expand the list of cities from 35 to 70 large and medium-sized cities, which is publicly released by the National Bureau of Statistics of China (http://www.stats.gov.cn/sj/).Based on this expansion, tests are conducted again, as shown in Plan II of table 4. The control group of Wuxi is composed of Baotou (0.333), Pingdingshan (0.224), Tangshan (0.186), Quanzhou (0.141), Qingdao (0.074), Beijing (0.025), and Shanghai (0.017).The results of the absolute effect and relative effect indicate that the treatment effect of NSC still exists, and similar results could be found in the test of Tianjin.
(2) Alternative policy interference.In addition to the impact of NSC construction, carbon emissions at the city level may also be affected by other policies.Among these, China's national low-carbon city pilot policy is most relevant to our study.Since 2010, China has successively launched three batches of low-carbon city pilots (Qiu et al 2021).This study considers the possible impact of this policy on the research results.Firstly, since Tianjin itself is a low-carbon city, even if the synthetic Tianjin is composed of areas that do not belong to the low-carbon pilot cities, it will not negate the existence of the treatment effect in Tianjin.Wuxi, on the other hand, is not a low-carbon pilot city itself, thus the identified treatment effect of NSC in Wuxi may be due to the fact that Wuxi does not carry out low-carbon city pilot construction.To mitigate this potential impact, we conduct another test after removing all samples belonging to low-carbon pilot cities from the control group.As shown in Plan III in table 4, the absolute effect can still reach above 5, indicating that the estimation of the treatment effect is not disturbed by the low-carbon city policy.
(3) Change research dimension.This study further elevates the dimension of the control group from the city level to the provincial level, that is, the control group consists of 29 provinces, municipalities, and autonomous regions (excluding Tibet due to data deficiency), and retests Tianjin (as one of the four municipalities directly under the central government, it appears in both city-level and provincial-level statistics along with Beijing, Shanghai, and Chongqing).Synthetic Tianjin is composed of Beijing (0.544), Chongqing (0.279), Qinghai (0.084), Shaanxi (0.049), and Inner Mongolia (0.044).The results, shown in Plan IV in table 4, indicate that the absolute effect and relative effect reach 25.198% and 18.9%, respectively, further ensuring the robustness of the results earlier.
(4) Change treated group.To address the potential influence of sample selection on the negative environmental impacts of NSC, namely whether other NSC cities also exhibit similar treatment effects, this study conducts additional tests by replacing the treatment group with Jinan and Changsha, two cities with established NSCs (with the former commencing its construction in 2011 and the latter in 2010).The results, as presented in Plan V in table 4, indicate that the treatment effects still exist.In comparison to Jinan, the negative impact of NSC construction on the environment is even stronger in Changsha.The reason behind this phenomenon may stem from the fact that the NSC in Jinan has established an ' energy pool ' , which is formed by the five types of clean and renewable energy sources (computer waste heat, solar energy, geothermal energy, air energy, and natural gas).Thus, despite the center's high energy consumption, its environmental impact, specifically in terms of carbon emissions, remains relatively minimal.
In the preceding text, we estimate the treatment effects of NSCs in each city separately.Here, we use the 'synth_runner' command in stata to reevaluate the treatment effects, which allows for multiple treatment objects to be impacted at different times and could be able to directly provide statistical inference P-values to compare the effectiveness of the placebo tests (Galiani and Quistorff 2017).Specifically, to ensure the fitting effect before the shock while keeping the control variables unchanged, we excluded cities with later construction times such as Wuxi and Zhengzhou, and evaluated the treatment effect of NSC including Tianjin, Changsha, Jinan, Shenzhen, and Guangzhou as a whole.As shown on the left side of figure 5, it can be observed that prior to the NSC shock (vertical line), the carbon emission values of the treatment group and its synthetic object closely coincide.Following the shock, the difference between the two gradually increases, and the policy treatment effect begins to emerge.The right side of figure 5 shows the change in statistical inference P-value (the probability of an accidental occurrence), indicating that apart from the first two periods, the P-values for other years are all below 0.05, further confirming the existence of the treatment effect.

Mechanism test
To investigate the mechanisms by which NSC affects CO 2 emissions, this study conducts the mechanism test from three aspects: scale effect, technology effect, and composition effect, based on the theoretical analysis in section 2.2.As shown in figure 6, in the graph with 'Gap in electricity consumption' as the y-axis, the solid line represents the treatment effects of NSC in Wuxi and Tianjin (the impact on energy consumption), and the dotted line represents the effect of 'NSC' in other cities.It can be seen that the real treatment effect is significantly larger than that of 'NSC' in other cities, indicating that NSCs lead to increased total energy consumption in Wuxi and Tianjin, resulting in higher regional carbon emissions.The software and hardware equipment of the computing infrastructure needs to operate continuously to ensure timely and effective supply of services like data processing, storage, and transmission.Furthermore, compared to general computing centers, NSCs have greatly shortened the R&D cycle and improved knowledge production efficiency due to their high-performance supercomputers, which may further trigger environmental rebound effects of digital technology, leading to a rapid increase in demand for computing services and thus expanding energy (electricity) consumption.Based on the above results, hypothesis H1 is confirmed.
In the graph with 'Gap in energy structure' as the y-axis, although the trend of the solid line suggests that NSCs have an impact on the energy structure in Wuxi and Tianjin, compared with the 'treatment effects' in other cities, this effect is not relatively prominent, and a considerable number of dotted lines show greater changes than the solid line.Therefore, the potential composition effect of NSC is not supported statistically, rejecting the H2 hypothesis proposed earlier.
One possible reason why NSCs do not have a significant impact on carbon emissions through composition effect is that the influence of computing infrastructure on urban energy supply is relatively limited, and its knowledge spillover and radiation effect are insignificant.Moreover, since the NSC did not significantly improve the energy structure, it can be inferred that the computing infrastructure do not substantially leverage its advantages in green and energy-saving technology innovation.In the graph with 'Gap in Patent_green' as the yaxis, it can be observed that both Wuxi and Tianjin show a downward trend after NSC construction.Although it cannot be concluded that NSC hinder green innovation, it can be certain that the effect of NSCs on reducing carbon emissions through technology effect is also limited.Thus far, the examination of the technology effect mechanism has refuted hypothesis H3.
Overall, based on the above mechanism tests, it can be inferred that, compared to technology effect and composition effect, Scale effect still dominates the mechanism by which computing infrastructure affects carbon emissions.
6. Discussion, conclusion, and policy implications  ecosystem consisting of a series of different, complementary, and interrelated knowledge and technology components, and these heterogeneous digital technologies may have different impacts on the environment.
In recent years, major economies around the world have increasingly recognized the importance of supercomputing capabilities.The European Union has undertaken the construction of eight major supercomputing centers involving a budget of 840 million euros and 28 member countries.Likewise, the United States has introduced a series of acts and reports, including the National Artificial Intelligence Initiative Act and the Pioneering the Future Advanced Computing Ecosystem: A Strategic Plan.In 2023, the Chinese government also issued the Overall Plan for Digital China Construction and initiated the National Supercomputing Internet project in April of the same year.Throughout this process, it is crucial to give due attention to the energy consumption and carbon emissions of supercomputing centers, which affects the role of computing infrastructure in sustainable development.Given that there are few studies examining the relationship between digital technology and green development from the perspective of computing infrastructure, this study focuses on China's National Supercomputing Center, a large-scale computing infrastructure, to investigate its impact on carbon emissions.Our research results show that NSC construction could have a negative impact on the environment to some extent, that is, it expands the city's carbon emissions.Although the degree of impact varies among different cities, the result of carbon increase has remained robust.
Through the investigation of specific mechanisms at play, it has been determined that the increase in energy demand and consumption is the main driver behind CO 2 emissions in the process of NSC operations.However, two important prerequisites must be considered.Firstly, the local energy structure still exhibits a considerable dependence on coal.Given that the NSC is characterized by a significant demand for electricity (JoppaL and Herweijer 2018, Jones 2018, Galperova and Mazurova 2019, Allen 2022, Open Data Center Committee 2022), its construction and operation will substantially increase local power consumption.As more than half of China's total power generation comes from thermal power, which is predominantly coal-based, the rise in energy consumption resulting from NSC construction will inevitably lead to increased greenhouse gas emissions.Secondly, the level of green innovation has not yet reached a point where it can fully mitigate the carbon footprint associated with NSC development.While the development of carbon capture and storage technology, biological carbon fixation technology, hydrogen, solar energy, wind energy, and other renewable energy sources can assist in reducing CO2 emissions from different aspects (Fan et al 2022, Wang et al 2022a, Jin et al 2022, Bianchini et al 2022), it is apparent that the development of green technology has not eased the negative impacts of rising energy demands (including improvements to the energy structure).The results of the investigation into energy structure and green innovation mechanisms also indicate that the NSC project has not demonstrated significant composition effect (energy structure improvement) and technology effect (green technology innovation).Although the existence of both cannot be denied, it is certain that these factors have only limited impact on reducing the overall negative environmental impact of the NSC project.
It should be noted that our research still has certain limitations.On one hand, our study focuses on the computing infrastructure, which includes not only supercomputing centers but also large data centers (both collectively referred to as computational and data infrastructure).Evaluating the carbon emissions impact of large data centers and comparing it with the carbon emissions impact of supercomputing centers is one of the future directions to consider.On the other hand, the assessment of supercomputing centers in our study is based on NSCs.In reality, there are still considerable-sized supercomputing centers owned by local governments, universities, or private enterprises.If this aspect is considered collectively in the future, it will contribute to obtaining more accurate results.Nevertheless, overall, this study expands the evaluation of carbon emissions impact of digital infrastructure from the perspective of national computing infrastructure and reveals the primary mechanisms leading to carbon emissions increase.
Based on the above, we can summarize the following conclusions: • The construction of National Supercomputing Centers in both Tianjin and Wuxi has increased local CO 2 emissions.
• The NSC primarily increases CO 2 emissions through the rise in local energy demand and consumption (electricity usage), indicating that scale effect (including rebound effect) play a dominant role.
• The identification of composition effect and technology effect is relatively unclear, meaning that the NSC's effect on carbon emissions through improving energy structure and promoting green innovation is limited.
• Compared to Tianjin, the negative impact of NSC construction on the environment in Wuxi is smaller, largely due to the development of technology and the improvement of the institution (later construction time).This conjecture is also supported by the results of the tests conducted in Jinan and Changsha.

Policy implication
The research findings of this study possess certain policy implications.Given that the construction of the NSC has significantly increased the local (urban) carbon emissions, both the public sector and operating units should pay sufficient attention to it.It is necessary to conduct a rigorous examination in various aspects of the NSC, including electricity utilization efficiency as well as power supply and air conditioning systems, to ensure compliance with green and low-carbon requirements.Updating and replacing inefficient equipment, implementing more efficient coding, and considering targeted data storage are all feasible strategies to reduce energy consumption and carbon emissions in the NSC.Moreover, controlling power usage alone may be not enough for NSC.Reducing energy consumption from non-computing activities such as cooling can be achieved through the development of more advanced liquid cooling method.Additionally, a innovative decarbonization mechanism that is driven by low-carbon computing consumption could be considered.In the past, innovation organizations including universities and research institutions in China have relied on government-funded budgets to acquire computing power services.However, under this new mechanism, the low-carbon development of supercomputing centers can be promoted from the demand side by purchasing computing power services through the payment of 'carbon coin'.
As Allen (2022) has pointed out, a more efficient computing center may quickly lose its green advantage if it uses less renewable energy compared to a less efficient center.The conclusions drawn from this study reveal the role of energy structure in the carbon emissions increase at NSC.Therefore, we suggest an increase in the supply of renewable energy, including photovoltaic and wind power.Such measures have been successfully implemented at the Dutch national supercomputing facility SURF and the German Max Planck Institute, where the use of renewable energy like wind and solar power has resulted in lower carbon emissions compared to those using fossil fuels (Jahnke et al 2020, Van der Tak et al 2021).The NSC in Jinan has already constructed an 'energy pool' consisting of five clean and renewable energy sources, which provides support for carbon reduction.This 'energy pool' covers an area of 630,000 square meters and is currently the largest distributed energy center in Jinan, equipped with a highly integrated system for heat and cold output.
Taking solar energy as an example, the rooftop of the supercomputing center's building accommodates photovoltaic panels with a total installed capacity of 2.5 MW, generating an annual electricity output of 2.9 million kWh.Currently, all the generated electricity is directed into the 'energy pool.' Inside the supercomputing center, the continuous operation of computers generates a substantial amount of heat, which is extracted and combined with geothermal energy to form a centralized ' energy pool ' for cooling and heating purposes.Natural gas serves as a supplementary energy source for peak load regulation.This multi-energy complementation model effectively compensates for the limitations of a single energy source, resulting in an annual savings of 8,000 metric tons of standard coal for the project.In this study, the absolute effect and relative effect of Jinan NSC in treatment effect are only 1.859 and 3.1%, respectively.Hence, for the NSC in China, it is crucial to accelerate the adjustment of energy supply and increase the proportion of renewable energy to achieve carbon reduction targets.
Finally, our research results indicate that the composition effect and technology effect of NSC construction are not significant.Therefore, it is necessary to strengthen the role of NSC construction in promoting the upgrading of the local industrial structure and green innovation in the future.By building a running environment that integrates supercomputers with big data and artificial intelligence applications, necessary resources and platforms can be provided for the output of digital products and services.Deepening cooperation with new energy, new materials, advanced manufacturing, biomedicine, and other industries can help companies improve production and R&D efficiency, leading to accelerate digital and green transformation.In general, the construction of NSC should not only support the local innovation-driven strategy and serve the development of local industries, but also leverage the spillover effect of knowledge, driving the city towards intelligent, green and sustainable development.

Figure 2 .
Figure 2. The spatial distribution and carbon emissions (million tons) characteristics of NSCs.

( 3 )
Green innovation.This study measures the level of urban green innovation by the city's green patent acquisition situation(Wang et al 2022a, Bianchini et al 2022, Meng and Zhang 2022), with data sourced from the Green Patent Research Database (GPRD) in China Research Database Services (CRDS), which organizes patents from the State Intellectual Property Office and Google Patent according to the green patent classification number standard published by the World Intellectual Property Organization.4.2.3.Control variableBased on the EKC theory, this study further selects per capita GDP, total population at year-end, and the proportion of secondary industry as control variables.Per capita GDP and total population represent the economic scale(Hua et al 2018, Yang et al 2021), while the proportion of secondary industry represents the industrial sector structure(Yang et al 2021).Technology factor is considered by including the proportion of urban financial S&T expenditures to GDP as a control variable in the model(Hua et al 2018).In addition to the variables mentioned above, this study follows the research approach ofAbadie et al (2010) and also includes the value of the result variable (CO 2 emissions) in specific years before the event occurred as control variables.Specifically, for the estimation ofWuxi, years including 2003Wuxi, years including  , 2007Wuxi, years including  , 2011Wuxi, years including  , and 2015   are selected, while for the estimation ofTianjin, 2003Tianjin,  , 2005Tianjin,  , and 2007   are selected.

Figure 5 .
Figure 5.The overall treatment effect and the effectiveness of placebo tests.

6. 1 .
Discussion and conclusion Regarding the impact of digitalization on the environment, existing literatures examined it from the perspective of specific digital technology including the Internet and big data (Yang et al 2021, Wang et al 2022b, Wang et al 2022c), as well as regional policies that assess the effect of smart city and broadband network infrastructure on the environment (Zou and Pan 2022, Liu et al 2023, Wu et al 2023).Alternatively, discussions have been conducted on the impact of digital economy (Wang et al 2022a, Luo et al 2022), digital finance (Meng and Zhang 2022), digital trade (Fu et al 2022), and digital innovation (Perrons 2021, Bianchini et al 2022) on the environment.Among them, some studies have pointed out that the digital technology has not had a significant impact on environmental improvement, and may even have negative impacts under certain conditions (JoppaL and Herweijer 2018, Jones 2018, Galperova and Mazurova 2019, Pohl et al 2019, Jin et al 2022, Bianchini et al 2022, Jin et al 2022, Bianchini et al 2022, Bieser et al 2023).Digital technology is not a single entity, but a digital
al 2022, Fu et al 2022, Meng and Zhang 2022), or explores more specific digital technology (Yang et al 2021, Wang et al 2022b, Wang et al 2022c).There are also studies that examine the effect of digital infrastructure (Zou and Pan 2022, Liu et al 2023, Wu et al 2023), or construct indices for the digital economy using multiple indicators (Fan et al 2022, Wang et al 2022a, Luo et al 2022, Han et al 2022, Zhao et al 2023) (Shan et al 2022)O 2 emissions as the dependent variable, sourced from the China Emission Accounts and Datasets (CEADs).The database contains data including energy list, CO 2 emissions list, industrial process CO 2 emissions list, emission factors, and input-output tables.The particle swarm optimization-back propagation (PSO-BP) algorithm was employed to estimate the CO 2 emissions of 2735 Chinese counties, unifying the scales of Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS) and National Polar-Orbiting Partnership/Visible Infrared Imaging Radiometer Suite (NPP/VIIRS) satellite images(Chen et al 2020).The urban carbon emissions data used in this study are aggregated from county emissions and filled with missing values using the ARIMA model based on time series data of the same city.For the missing county data after 2017, the latest city data from CEADs (which considers 47 social and economic sectors and 17 types of fossil fuel emissions) is further referenced and supplemented according to corresponding proportions(Shan et al 2022).

Table 2 .
Fitting and comparison of control variables.

Table 3 .
Aasolute effect and relative effect.

Table 4 .
Other robustness tests.Change control group (Plan II) -70 major and medium-sized cities