Direct and Indirect Effects of Carbon Emissions from Digital Technology

The distinction between the direct and indirect effects of digital technology on the environment can be traced back to the research by Berkhout and Hertin. They categorized the environmental impact of ICT into three types: direct effect, indirect effect, behavioral and structural effect, which goes beyond the traditional binary understanding [5]. Specifically, the negative impacts resulting from the production, use, and disposal of ICT devices are categorized as direct effect, while indirect effect include the productivity improvements brought about by ICT applications and the wider dissemination of related products (dematerialization). Structural and behavioral effect refer to the series of impacts ICT have on economic structure and demand behavior. Subsequently, the aforementioned categorization about environmental impacts of digital technology has also been termed as first-order effect (direct effect, the demand for materials and energy throughout the lifecycle of digital technology) and higher-order effect (indirect effect, the structural and behavioral effect) [17]. Similarly, this logical classification can be applied in the evaluation of carbon emissions from digital technology [7, 26].

It is worth noting that despite significant progress in understanding the complex relationship between digital technology and environmental sustainability, the categorization of direct and indirect effects varies considerably across different literature. In the study of Yi et al., they classify aspects including digital technology, industrial digitization, digital governance, etc. as direct impacts, while considering changes in energy structure as indirect impacts [27]. Wang et al. distinguish indirect effect into scale effect (added value of the tertiary industry), structural effect (energy structure), and technological effect (green patent), which aligns with the understanding of Zhang et al. and Dong et al. [6, 10]. Tang and Yang elucidate the direct effect of digital infrastructure on carbon emissions from the perspectives of energy consumption and environmental governance, while analyzing potential indirect effect from aspects like residential consumption behavior and energy factor allocation [11].

Among these studies, the works of Luan and He et al. stand out as the most typical examples. The former, contrary to Berkhout and Hertin, describes direct effect as the positive impacts robots directly bring to the environment, while portraying indirect effect as the negative consequences they generate [28]. The latter, similar to Berkhout and Hertin, characterizes direct effect as the increased electricity consumption due to digital transformation, while including a series of positive impacts as part of the indirect effect, including productivity enhancement, easing of enterprise financing constraints, and energy conservation [4].

Mechanism Framework

Based on the above analysis, two important premises can be summarized.

Firstly, the core factors through which digital technology affects greenhouse gas emissions are energy consumption and energy structure. Secondly, there is notable variation in the interpretation of direct and indirect effects in existing research. Hence, before constructing the mechanism framework in this study, it’s essential to clarify their meanings in the context of NSC’s impact on carbon emissions. this paper will explore the driving factors (mechanisms) behind the carbon emissions effect of supercomputing centers at the city level, focusing on the aspects of energy consumption and energy structure.

NSCs demand substantial electricity for data handling and high-performance computing tasks [29]. They also require power to maintain equipment and stable environments. In this study, we define the electricity used during NSC operations as direct carbon emissions. As for the indirect effect. Their supercomputers accelerate research and calculations, saving energy through optimized algorithms and parallel computing. NSCs offer digital services to local institutions, enhancing resource allocation, including energy. This transforms the regional industry towards high-tech, efficient, and cleaner production. Precise prediction and simulation improve energy allocation, reducing waste. However, better computing efficiency may boost demand in research and industry, potentially increasing energy use.

In terms of energy structure, regional carbon emissions linked to NSC construction depend on the local energy mix [30]. In simple terms, if coal and other non-clean energy sources dominate electricity generation during and after NSC construction, the supercomputing center’s power-intensive operation will increase CO₂ emissions (direct effect). The indirect effect focuses on shifts in energy structure resulting from NSC. As mentioned earlier, NSCs support R&D, including clean energy technology development, like the joint CO₂ electrocatalytic reduction project by Chengdu’s NSC and Chongqing University, aiming to convert CO₂ into new energy products. Additionally, the NSC’s higher energy use can drive the region towards greener digital infrastructure. For example, NSC in Jinan established an “energy pool” with clean and renewable energy sources, such as solar and geothermal, to reduce the NSC’s environmental impact.

Based on the above, we have visualized the research mechanism of this study as shown in Fig. 1.

Research Object

Since 2009, China has established nearly a dozen national-level NSCs in various cities, considering high-performance computing as vital for technological advancement. These NSCs, serving as hubs for innovation, utilize supercomputers for research in fields like physical chemistry, astronomy, climate meteorology, biomedical sciences, etc. While driving technological progress, they also contribute to local socio-economic development. Supported by national science and technology initiatives such as the “863” program, China has made significant strides in supercomputing over the past two decades. Considering the potential lag effect of NSC influence, This study focuses on five major NSCs established between 2009 and 2011 in Tianjin, Shenzhen, Changsha, Guangzhou, and Jinan. These NSCs employ supercomputers from the “Tianhe,” “Sunway,” and “Dawning” series, which have held leading global positions (https://www.top500.org/lists/ top500/). It should be noted that China’s supercomputers do not perform as remarkably in the Green500 rankings as they do in the Top500. The latest list in June 2023 showed that although China’s supercomputers occupied two seats in the Top500’s top ten, none of them appeared in the top 60 of the Green500 list. This indicates that there is still significant room for improvement in the energy efficiency of China’s supercomputers, which may lead to higher environmental costs.

Fig. 1.

Mechanism diagram.

As shown in Fig. 2 below, cities hosting national-level NSCs tend to have higher carbon emissions.

Methodology

In the evaluation of the environmental effects of digital technology, some studies use instrumental variable methods for identification [9, 11, 27], but obtaining accurate and unbiased estimation results places high demands on the selection of instrumental variables. To examine the impact of NSC construction on urban carbon emissions, this study treats NSC construction as a quasi-natural experiment and conducts the analysis within a causal inference framework. Currently, for the evaluation of exogenous shocks or policy effects, the most common approach is the difference-in-differences method [3, 10, 11, 13, 20-22, 24]. This method compares the sample that is affected by the shock or policy with the sample that is not affected to assess the impact of the shock or policy. However, a critical issue is how to choose an appropriate comparison sample. To overcome this problem, this study adopts the Synthetic Control Method proposed by Abadie et al. [31, 32]. This data-driven method fits a synthetic control group that closely resembles the treatment group based on multiple control units, reducing selection bias and endogeneity problems.

In the context of this study, we have data on the carbon emissions of K+1 cities over time t∈[1, T]. Among them, S_it^N represents the carbon emissions of the city i at time point t without NSC shock, and S_it^I represents the carbon emissions of the city i at time point t after the construction of NSC. If T₀ is the year of NSC construction, when t<T₀, S_it^N = S_it^I; and when NSC is constructed, i.e., t≥T₀, a_it = S_it^I – S_it^N represents the carbon emissions effect of NSC. For a specific NSC city, we can only observe the carbon emissions value S_it^I after the construction of NSC, and we cannot observe the carbon emissions value S_it^N when NSC is not constructed during the same period. This study uses the factor model proposed by Abadie et al. for estimation [31]:

In Equation (1), δ_t represents the time fixed effects that affect the sample cities, Z_t denotes the observable predictor variables, λ_t is the unobservable (1×F) vector of common factors, μ_i is the (F×1) vector of unobservable city fixed effects, and ε_it is the short-term shock (unpredictable) with a mean of 0 at the city level.

Considering a city, i = 1, implementing NSC construction, while the remaining K cities have not conducted NSC construction. In pursuit of the outcome variables for region i unaffected by the influence of NSC, we designate K other cities as control units and employ a non-parametric weighting method to model the characteristics of the experimental group. Thus, Let W be a (K×1) weight vector defined as W = (w₂, w₃, ..., w_K+1), where W_K≥₀, w₂ + ... w_K+1 = 1, and each W represents a virtual synthetic control combination, i.e., weights for the K cities. For NSC cities, W represents the synthetic control contribution rate of each city in the control group to the NSC city, weighting the outcome variables for each control city.

Fig. 2.

Carbon emission and NSC distribution.

Assuming the existence of a weight vector W^* = (w₂^*, w₃^*, ..., w^*_K+1) such that for t<T₀, ∑k=2K+1wk*Skt=S11, ⋯, ∑k=2K+1wk*SkT0=S1T0, and ∑k=2K+1wk*Zk=Z1. According to Abadie’s proof, if ∑t=1T0λt′λt is a non-singular matrix, and there is an adequate time period for pre-NSC shock matching, the formula (3) approaching zero could be approved:

Finally, the estimated value of the NSC carbon emissions effect a_it is obtained as follows:

In the above equation, t∈[T₀, T]. To overcome the limitations of the traditional synthetic control method regarding the number of treated units, this study employs the synth_runner program developed by Quistorff and Galiani in Stata. This program allows for the presence of multiple treated units and provides p-values for statistical inference [25].

Data Indicators and Research Process

Baseline Test

In this study, we first estimate the overall NSC carbon emission effect. As shown in Fig. 4, the solid line on the left graph represents the carbon emissions of the entire treatment group, including cities like Tianjin, Shenzhen, and Changsha. The dotted line reflects the synthetic control group’s CO₂ emissions, which is weighted from the control group. The vertical dotted line represents the year when the computational infrastructure construction began. The left side of the dotted line indicates the carbon emissions trend of the treatment and synthetic control groups before the NSC impact. It can be observed that before the construction of NSC, the two curves nearly overlapped, indicating that the control group cities provided a good fit for the carbon emissions trend of the treatment group before NSC. However, starting from the second year of NSC construction, the carbon emission values of the treatment group cities show a significant increase compared to the synthetic control group cities. The difference between the treatment and control group cities increases over the years, reaching a peak difference of approximately 10 million tons. These results suggest that the treatment group cities experienced a significant increase in carbon emissions during the sample period, which we attribute to a strong association with the construction of NSC.

Fig. 3.

Research Procedure.

To further validate this conjecture, we conduct a placebo test following the approach of Abadie et al. to assess the robustness of the evaluation results [32]. In this test, we assume that cities in the control group also “implemented” NSC construction during the same period and constructed synthetic control groups for each city in the control group using the same method. We then calculate the carbon emission difference between the actual cities and the synthetic cities, considering it as the “effect” of “implementing” NSC construction for each city. Finally, we compare these placebo effects with the actual carbon emission effect of the real treatment group (cities like Tianjin, Shenzhen, Changsha, etc.). It is worth noting that the good fit between the treatment and control groups before the NSC impact is crucial for evaluating the treatment effect. Therefore, in this study, we use ten times the RMSPE value before NSC construction as the threshold to exclude estimates that exceed this threshold. Using the program developed by Quistorff and Galiani [25], we directly provide the P-values for statistical inference after conducting the placebo test, as shown on the right side of Fig. 4. It can be observed that the P-value for the treatment effect decreases to below 0.1 in the last period, indicating that to some extent, the NSC carbon emission effect is statistically significant (though only at the 10% level).

Heterogeneity Analysis

The results of the baseline test show that, the carbon emission effect of NSC construction is marginally significant, answering the question in the flowchart “Does NSC lead to an increase in urban carbon emissions?” At the same time, we also observe that the carbon emission effect is only statistically significant at the 10% level in the final period. Therefore, we move on to the second question in the study, “Does this impact vary across different cities?”. This aims to explore whether the marginally significant results are driven by the heterogeneous effects of NSC construction. In this study, we evaluate the carbon emissions effect of NSC construction in different cities, including Tianjin, Shenzhen, Changsha, Guangzhou, and Jinan. Based on the results, we divide them into two groups: “No carbon emission effect” and “Existence of carbon emission effect,” as shown in the top and second rows of Fig. 5, respectively.

Looking closely, cities like Guangzhou, Jinan, and Shenzhen do not experience a significant increase in carbon emissions due to NSC construction. For Guangzhou, although the difference in actual carbon emissions from the synthetic control group increases after NSC, the carbon emission effect is not statistically significant (as shown in the P-value distribution on the right side of Fig. 5). Similar results are observed for Jinan and Shenzhen. In contrast, both Tianjin and Changsha show significant carbon emission effects from NSC construction. As shown in the bottom part of Fig. 5 (second row), before NSC impact, the carbon emissions of the synthetic control group closely mimic those of the treatment group, indicating a good fit for the carbon emissions characteristics of the control group before NSC construction. After the NSC, the carbon emission difference between the treatment group and the synthetic control group rapidly widens in one to two periods and receives statistical support. The carbon emission effect of Tianjin’s NSC construction remains statistically significant at the 1% level, while Changsha’s carbon emission effect becomes statistically significant at the 5% level after the mid-term.

Fig. 4.

Baseline test results (overall).

Fig. 5.

Heterogeneity test of NSC carbon emission effects.

Carbon Emission Driving Factors Analysis

Energy Structure

Although the results of the analysis in section 4.3.1 suggest that the insignificant carbon emission effects of NSC construction in some cities may be due to the lack of significant increase in local energy consumption (as observed in Jinan), the estimation results for cities like Shenzhen indicate that the energy consumption scale is not a sufficient condition for significant NSC carbon emissions. In this subsection, we focus on the perspective of energy structure. In fact, some studies have emphasized the crucial role of energy composition in the cost of supercomputing environments [29, 30, 37, 38]. Even if a supercomputer center is more efficient, its green advantage will gradually diminish if it relies on a considerable share of non-renewable energy. Based on the analysis in section 2.2, we explore the role of energy structure from both direct and indirect effects. Let us now proceed with specific examples and data illustration.

(1) Direct effect

If a region has a higher proportion of clean energy in its energy structure, even if the operation of NSC leads to an increase in regional electricity consumption, the negative impact on the environment is expected to be smaller, provided that most of this electricity is produced from clean energy sources. Conversely, if the proportion of non-clean energy, such as coal, is high in the energy structure, the impact would be the opposite. To verify this hypothesis, we construct two major indicators: the proportion of coal and the low-carbonization index (as described in section 3.3.2), and match them with the corresponding cities.

Using data from the China Energy Statistical Yearbook and converting various energy types into standard coal based on energy conversion coefficients, we obtain the proportion of various energy sources. As shown in Fig. 6, we divide the provinces into two groups based on the significance of NSC carbon emission effect: G1 (significant, consisting of Tianjin and Hunan) and G2 (insignificant, consisting of Guangdong and Shandong). It is evident that both before and after the NSC shock, the average coal proportion in G2 provinces is significantly lower than that in G1 provinces (with a peak difference of 15 percentage points). Similarly, the low-carbonization index in G2 provinces also has a significantly higher average compared to G1 provinces (with a peak difference of 0.5). The above results demonstrate that, despite NSC construction in Tianjin, Changsha, Shenzhen, and Guangzhou all leading to an increase in regional energy consumption, the carbon emission effects of Tianjin and Changsha supercomputing centers are significantly higher than those of the latter two cities. This is largely attributed to their higher proportion of coal in the energy structure.

In the lower part of Fig. 6, we can observe the specific evolution trend of energy structure in different provinces. It is evident that the coal proportion in Guangdong Province (covering Shenzhen and Guangzhou) is significantly lower than that of Hunan and Tianjin (conversely, its low-carbonization index is higher than the latter two). In fact, both Tianjin and Changsha are typical industrial cities in China. Tianjin is one of the birthplaces of modern industry in China, and industries such as automobile manufacturing, petrochemicals, metallurgy, and steel are dominant, leading to a high dependence on high-carbon energy sources. On the other hand, Changsha faces severe energy resource shortages, with over 80% of its energy needs being supplied from other regions, and its coal consumption proportion is particularly prominent. In specific data, in the year 2005, the coal proportion in Changsha’s energy structure reached 68%. Moreover, the average value of the sample period for coal proportion ranks first among the supercomputing cities.

Fig. 6.

Evolution trend of energy structure in supercomputing cities.

(2) Indirect effect

The impact of energy structure on NSC carbon emissions is not only determined by the existing energy composition but also reflected in its indirect effects. In other words, NSC construction can also influence the energy structure, thereby affecting its own carbon emissions. Therefore, this study takes the low-carbonization index as the outcome variable and conducts another set of synthetic control estimations for each city. The results are shown in Table 3. Among the cities with “no carbon emission effects,” only the treatment effect of Jinan is statistically significant. On the other hand, for cities with “carbon emission effects” (Tianjin, Changsha), NSC construction also contributes to improving the urban energy structure. However, the magnitude of this effect does not reach the level of altering the fact of higher coal proportion in cities like Tianjin and Changsha. Therefore, it does not significantly reduce the environmental costs incurred by NSC construction.

Table 3.

The impact of NSC on urban energy structure.

	Low carbonization level (structure)
	Shenzhen				Jinan
Year	Treated	Synthetic	Absolute Effect	Relative Effect	Treated	Synthetic	Absolute Effect	Relative Effect
Post_1	5.833	5.833	0.000	0.01%	5.348	5.318	0.030	0.57%
Post_2	5.913	5.903	0.010	0.17%	5.332	5.313	0.019	0.36%
Post_3	5.910	5.899	0.011	0.18%	5.463	5.387	0.076	1.40%
Post_4	5.957	5.947	0.009	0.16%	5.427	5.398	0.029	0.53%
Post_5	6.094	6.078	0.017	0.27%	5.681	5.410	0.272***	5.02%
Post_6	6.051	6.037	0.014	0.23%	5.867	5.441	0.426***	7.84%
Post_7	6.115	6.101	0.015	0.24%	5.906	5.544	0.362***	6.53%
Post_8	5.638	5.681	-0.043	-0.76%	6.174	5.646	0.527***	9.34%
Post_9	6.242	6.229	0.014	0.22%	6.178	5.678	0.500***	8.82%
Post_10	6.322	6.298	0.024	0.38%	6.185	5.707	0.478**	8.37%
Post_11	6.346	6.320	0.026	0.41%
Post_12	6.394	6.364	0.030	0.47%
	Tianjin				Changsha
	Treated	Synthetic	Absolute Effect	Relative Effect	Treated	Synthetic	Absolute Effect	Relative Effect
Post_1	5.406	5.521	-0.115*	-2.09%	5.433	5.123	0.310***	6.05%
Post_2	5.594	5.629	-0.035	-0.62%	5.488	5.175	0.313**	6.04%
Post_3	5.601	5.567	0.034	0.61%	5.472	5.198	0.274**	5.27%
Post_4	5.581	5.538	0.044	0.79%	5.504	5.271	0.233**	4.41%
Post_5	5.630	5.586	0.044	0.79%	5.498	5.257	0.241**	4.58%
Post_6	5.665	5.617	0.048	0.85%	5.287	5.266	0.022	0.41%
Post_7	5.749	5.607	0.142*	2.53%	5.261	5.324	-0.063	-1.18%
Post_8	5.826	5.597	0.229*	4.09%	5.189	5.400	-0.211	-3.91%
Post_9	5.942	5.709	0.234*	4.10%	5.337	5.513	-0.176	-3.19%
Post_10	5.992	5.854	0.138	2.35%	5.382	5.556	-0.174	-3.13%
Post_11	6.014	5.838	0.176	3.01%	5.438	5.586	-0.148	-2.65%
Post_12	6.014	5.989	0.024	0.40%

[i] The treatment effect of NSC in Guangzhou is not statistically significant in all periods after the shock.

Research Findings

As data and computing resources become deeply embedded in the real economy, it presents opportunities for high-quality socio-economic development, but also challenges in the fields of energy, environment, and climate. Existing literature has revealed the complex impact of digitization on the environment. On one hand, it can improve production and energy efficiency, promote innovation and industrial upgrading, thus reducing carbon emissions. On the other hand, the digital industry itself is energy-intensive, which may increase environmental burden and expand consumption demand through rebound effects, leading to an expansion of energy consumption. This study extends the assessment of carbon emissions from digital infrastructure to the computational level. Drawing on the research framework of direct and indirect effects based on energy consumption and energy structure by Berkhout and Hertin [5], we treat the construction of National Supercomputing Centers (NSCs) as a quasi-natural experiment. Utilizing the multiple synthetic control estimation proposed by Quistorff and Galiani [25], we examine the impact of NSCs on urban carbon emissions and explore feasible pathways for green transformation of computational infrastructure through the inter-city heterogeneity of NSC carbon emission effects. Specific research findings include:

(1) To a certain extent, NSCs significantly promote the increase in regional carbon emissions, but the carbon emission effects show inter-city heterogeneity. In some cities, NSCs do not significantly exacerbate local CO₂ emissions. Existing research has explored various dimensions, including enterprises [4], cities [3, 10, 11, 13], regions [6, 9, 19], etc., providing positive impacts of digitization on environmental sustainability from multiple perspectives including digital economy [3, 6, 9, 19, 27], specific digital technologies (e.g., the internet, big data, artificial intelligence) [12, 15, 14, 28], and digital infrastructure [10, 11, 13, 20-24]. Consistent with the research results of Rao et al., Tang and Yang, and Wu et al. [11, 13, 20], we find that China’s National Supercomputing Centers have significant carbon emission effects. This is also in line with the conclusion of Bianchini et al. (2022) [12], who used EU data and found that the development of local digital technologies increased greenhouse gas emissions, with big data and computational infrastructure having the highest environmental costs. This study extends the investigation about the environmental effects of network infrastructure and integrated infrastructure, and further explores the perspective of computational infrastructure. The inter-city heterogeneity of NSC carbon emission effects provides a basis for exploring its driving factors in this study.

(2) Despite NSCs primarily contributing to the increase in local carbon emissions through the channel of elevated energy consumption, specifically industrial electricity consumption, the estimation results for cities like Shenzhen indicate that energy consumption is not a sufficient condition for significant NSC carbon emissions. In other words, although NSCs lead to an increase in energy consumption in these areas, it does not significantly promote a rise in local carbon emissions. Energy consumption is considered a crucial variable in evaluating the environmental effects of digital technologies, yet there is still no unified conclusion on whether digitization can reduce energy consumption. As a typical comparison, Wang et al.’s research demonstrates that digital transformation can lower electricity consumption and intensity through technological optimization and industrial upgrading [22]. However, in the study by Tang and Yang, digital infrastructure increased urban CO₂ emissions by inducing a rise in per capita energy consumption, total energy consumption, and energy intensity [11]. The findings of this study complement this knowledge and confirm the significant electricity consumption caused by large-scale computational projects at the local level.

(3) The energy structure is a critical factor influencing the carbon emissions of NSC. This study finds that those supercomputing cities with significant carbon emission effects often have a higher proportion of coal in their energy structure (or lower low-carbonization index). Furthermore, we also discover that although NSC construction significantly promotes the improvement of low-carbonization levels in some supercomputing cities, this effect does not change the current situation of high coal proportion in cities like Tianjin and Changsha. As a result, it does not significantly reduce the environmental costs generated by NSC construction. The above findings further corroborate the assertion made by Allen that, concerning green performance, the energy attributes of supercomputing centers (supercomputers) may be more critical than their efficiency [30]. Research results by Van der Tak revealed that the carbon emissions of the Dutch supercomputing facility SURF are zero because it uses 100% green energy [37]. In contrast, estimates of the carbon footprint of the Max Planck Institute for Astronomy in Germany and Australian researchers captured the contribution of supercomputing to carbon emissions because both entities use non-clean energy to varying degrees [29, 38].

Practical Implications

Firstly, elevating energy efficiency and decreasing energy consumption. The research findings demonstrate that NSCs have markedly catalyzed an upsurge in carbon emissions. Mechanistic scrutiny further reveals that, excluding Jinan, the construction of NSCs in other cities has invariably led to a discernible augmentation in local energy consumption. Consequently, both the public sector and NSC operators should take steps to mitigate energy use. This includes optimizing hardware, choosing energy-efficient servers and processors, and implementing advanced energy management systems like Shenzhen’s “Upgrade and Replacement” project. Additionally, improving algorithms and task execution can reduce energy consumption in supercomputing centers.

Secondly, augmenting the proportion of clean energy and refining the energy structure. We find that, despite NSC leading to an escalation in urban energy consumption, it has not exacerbated CO₂ emissions, as observed in Shenzhen and Guangzhou. This is largely attributed to the reduced reliance on non-clean energy sources in these cities. Therefore, in the future, drawing inspiration from the practices of the Netherlands’ supercomputing facility, SURF, there is an opportunity to further enhance the utilization of renewable energy sources in NSCs. Presently, the energy pool established by NSC in Jinan, comprising various clean energy sources like waste heat recovery, solar, and air energy, has been reported to contribute to an annual reduction of 21,600 tons of CO₂ emissions.

Thirdly, expanding the spillover of knowledge and expediting regional emissions reduction progress. This study reveals that, whether in terms of energy consumption or energy structure, the contribution of NSCs to carbon reduction is relatively limited. Thus, improvements can be made from these two perspectives. On the one hand, NSC construction can foster urban sustainability through indirect effects on energy consumption channels, including enhancing computational efficiency, improving energy allocation capabilities, as well as promoting industrial structure upgrades. On the other hand, NSCs can collaborate with innovation hubs like universities and enterprises to provide computational support for green technology innovation, particularly in the realm of energy, thereby catalyzing the optimization of regional energy structures.