Introduction

The Internet of Things, Cloud Computing, and Blockchain technologies have emerged as pivotal components within the information technology landscape over the past few decades (Meijer, 2010). Data centers serve as critical physical infrastructures that provide essential computing, storage, and networking resources to support these technologies (Ding et al., 2024). With the rapid advancement of artificial intelligence (AI), particularly in the realms of generative AI and large language models, there has been a significant increase in demand for computility. This surge in demand has consequently led to the densification of data centers (Edelenbosch et al., 2024; Abderrahim et al., 2024). From 2018 to 2021, the global server population within data centers expanded at an annual rate surpassing 6000 megawatts (MW), demanding considerable computility to sustain diverse applications and consequently contributing significantly to energy consumption (Uddin et al., 2012; Al Kez et al., 2022). According to the International Energy Agency (IEA), data centers are expected to become the largest consumers of energy worldwide. Global data center energy consumption is projected to double within just 4 years, escalating from 460 terawatt hours (TWh) in 2022 to 1050 TWh by 2026 (Cao et al., 2022). This trend is anticipated to continue, with global data center energy consumption estimated at ~158 billion kilowatt-hours (kWh) by 2034 (Abderrahim et al., 2024). Moreover, it is projected that the data center sector will generate around 2.5 billion tons of carbon emissions, accounting for a significant 8% of global carbon emissions by 2030 (Xue et al., 2022). By 2040, computility-related emissions are projected to surpass 14% of global emissions (Xu et al., 2024). Data centers have become the contemporary equivalent of “industrial factories”. Despite the extensive utilization of renewable energy sources, reversing the substantial carbon footprint left by data centers remains challenging, underscoring the urgent need for sustainable solutions in this rapidly expanding industry.

China’s cross-regional computility transfer strategy represents an effective practice for mitigating this issue. The carbon emissions from China’s data centers significantly contribute to global data center carbon emissions, primarily due to the rapid development of the digital economy, which has driven a steady increase in both energy consumption and carbon emissions from these facilities. Data indicates that the carbon emissions intensity of China’s data centers is approximately twice the global average (Greenpeace and MIIT, 2021). In 2020 alone, electricity consumption reached 270 billion kWh, equivalent to 2.5 times the cumulative power generation of the Three Gorges Power Station during the same period (China Intelligent Computing Industry Alliance, 2022). Furthermore, in the first 2 months of 2024, the growth in data center electricity consumption surged by 40%, and this trend continues to escalate. It is projected that by 2035, electricity consumption will be ~2.5–3 times higher than in 2020, with carbon emissions from these facilities expected to reach 310 MT (Cyberspace Administration of China, 2023; Liu et al., 2023). To address the rising trend of high carbon emissions in data centers, the EDWC Project was officially launched in February 2022 as a strategic response. This project aims to harness the abundant renewable energy resources in western regions, such as wind and solar power, to construct a network of data centers (Zhang et al., 2018; Zheng et al., 2020). By systematically redirecting high-intensity computility demands from the east to the west, the EDWC Project seeks to reduce carbon emissions. When the utilization rate of renewable energy sources reaches about 80%, the reduction in carbon emissions by mid-century could be equivalent to that of a megacity by 2025 (Bertram et al., 2024). For instance, the Chengdu data center is projected to save 40 million kWh of electricity and reduce carbon emissions by 19,000 tons over a decade (Safi et al., 2024). This case effectively demonstrates the core essence of national data centers for cross-regional computility transfer and highlights their potential impact on emissions reduction.

Although previous studies have investigated the carbon reduction potential of the EDWC Project in China (Xie et al., 2024), the impact of carbon emission inequality arising from computility clusters in the eastern and western regions remains unclear. As a significant initiative for energy conservation and emissions reduction recently proposed by China, cross-regional computility transfer has received limited attention in studies focusing on carbon emissions from data centers. For example, existing studies on data centers have predominantly focused on evaluating their internal aspects, such as energy efficiency (Uddin et al., 2012; Xie et al., 2024), carbon emissions (Ding et al., 2018; Wang et al., 2022), and strategies for reducing their carbon footprint (Li et al., 2023; Abderrahim et al., 2024). Some studies employ life cycle assessment (LCA) methodologies to analyze the environmental impacts associated with data centers (Isler-Kaya and Karaosmanoglu, 2023; Song et al., 2024). Additionally, other studies have demonstrated that thermal management techniques and carbon pricing mechanisms can effectively reduce energy consumption and carbon emissions (Habibi et al., 2017; Eric et al., 2020; Wang et al., 2022; Tareen et al., 2024). Furthermore, the synergy between industrial agglomeration and carbon trading frameworks can facilitate overall emissions reductions (Zheng et al., 2020; Chen et al., 2024). In general, there has been substantial academic research on the environmental impacts of data centers within the context of building the digital economy, both at national and global levels (Lagouvardou et al., 2023; Bertram et al., 2024). However, the specific carbon emissions related to data centers, particularly those linked to cross-regional computility transfer, have increasingly drawn attention in efforts to achieve significant emissions reductions. The trend of post-implementation regional inequality, especially regarding directional flows of carbon emissions from east to west corresponding to cross-regional computility transfer, akin to inter-provincial spatial transfer effects, has rarely been explored (Zhang et al., 2023). Moreover, whether cross-regional computility transfer will exacerbate regional inequality in carbon emissions remains a topic of debate. Given the current severe situation of data center carbon emissions, comprehensively addressing these issues is an urgent priority for further research.

Here, we focus on the impact of cross-regional computility transfer strategy on the inequality of regional carbon emissions flow in data centers. To this end, we have defined two scenarios: the pre-EDWC Project as the baseline scenario and the post-EDWC Project as the policy scenario. Through these two scenarios, we aim to quantify the energy consumption and carbon emissions of China’s eight major computility hubs from 2020 to 2060. This analysis will reveal the improvements in energy efficiency and carbon emissions reduction achieved by national data centers following the implementation of the cross-regional computility transfer strategy. We then calculated the carbon emissions gap between the east and west computility hubs and compared the contribution of each data center cluster. Furthermore, we constructed five varied learning rate scenarios based on the policy scenario and quantified their abatement potentials using the same methodology. Although these scenarios may deviate from realistic development trends, they help us understand the drivers behind differences in abatement effects. Finally, we examined whether the computility transfer leads to inequality in carbon emissions flow and compared the magnitude of carbon emissions flow under different scenarios. Our research confirms the following findings. The implementation of the cross-regional computility transfer strategy has significantly mitigated carbon emissions in data centers and contributed to narrowing the carbon emissions gap between the east and west regions. However, this strategy has inadvertently exacerbated the inequality in carbon emissions flow. Moreover, specific energy consumption metrics in data centers, particularly PUE and renewable energy utilization, play a substantial role in carbon emissions reduction efforts. Curtailing the transmission of clean energy from west to east can notably alleviate the inequality associated with this carbon emissions flow process. These findings underscore the importance of cross-regional computility transfer in promoting sustainable development. While highlighting the roles of various factors in reducing carbon emissions from data centers, they also reveal the significant potential of these factors to mitigate the inequality impact of cross-regional carbon emissions flow. The results of our analysis can provide policymakers with critical evidence for proposing measures to reduce carbon emissions flow while prioritizing carbon reduction in national data centers.

Results

Carbon emissions reduction under the EDWC Project

The model outcomes (Fig. 1) reveal that the policy scenario leads to a substantial reduction in carbon emissions from eight computility hubs in China. By 2060, the installed capacity of these computility hubs under the policy scenario is projected to reach 164.6 million (Supplementary Fig. 2, illustrating the installed capacity layout of these computility hubs). In the baseline scenario, energy consumption of data centers is projected to be 4.6 × 10−4 billion kWh by 2060 (Supplementary Fig. 3a). However, under the policy scenario, this figure increases to 6.1 × 10−4 billion kWh over the same period due to the expansion of data center construction (Supplementary Fig. 3b). Despite this increase, the effective cross-regional computility transfer strategy within the policy scenario can reduce data center carbon emissions to nearly meet the required levels by 2030. Thereafter, the carbon emissions of data centers are expected to continue rising until they reach a turning point in 2039, after which they will enter a long-term decline phase (Fig. 1k). Under the baseline scenario, carbon emissions are projected to reach 1.6 billion tons by 2060, accounting for 12.7% of China’s total carbon emissions in 2024 and 28.4% of emissions from the power sector (Guo et al., 2022; Zhang et al.,2025). The role of data center carbon emissions in China’s overall carbon emissions is clearly significant. This result arises from the assumption of a fixed carbon emission factor of the power grid in the baseline scenario. However, given the rapid development of current AI technology, if no control measures are implemented, such an emission level is highly likely to occur.

Fig. 1: Trend of carbon emissions under baseline and policy scenarios.
Fig. 1: Trend of carbon emissions under baseline and policy scenarios.
Full size image

a–j Cluster - specific carbon emissions comparison between baseline and policy scenarios in 2020-2060. The sequence of the clusters is as follows: Zhangjiakou cluster, Wuhu cluster, Yangtze River Delta cluster, Shaoguan cluster, Tianfu cluster, Chongqing cluster, Guian cluster, Helinger cluster, Qingyang cluster, Zhongwei cluster. k The trend of total carbon emissions of ten data center clusters under the baseline and policy scenarios in 2020–2060.

Our findings indicate that the policy scenario does not result in a significant reduction in carbon emissions compared to the pre-2030 baseline. After 2030, the gap between the two scenarios begins to widen. Given that the officially proposed timeline for the EDWC Project in China is 2022 (Xue et al., 2022), there was no substantial improvement in data center carbon emissions during the early stages of this study. By 2030, however, the implementation of the cross-regional computility transfer strategy will bring total carbon emissions under a certain level of control (Cao et al., 2022; Wan et al., 2023). This will compensate for the slow progress in earlier efforts aimed at reducing carbon emissions. Numerically, by 2030, the policy scenario is projected to reduce carbon emissions by 16.9 MT (5.9%), and by 2060, a reduction of 13.14 MT (77.6%) could be achieved (Supplementary Fig. 4). The cross-regional computility transfer strategy proposed in the EDWC Project offers a feasible pathway for future reductions in data center carbon emissions in China and even globally.

However, as illustrated in Fig. 1a–j, there is a pronounced disparity in the extent to which different data center clusters experience reductions in carbon emissions due to the cross-regional computility transfer strategy (Shan et al., 2018). In the baseline scenario, where the cross-regional computility transfer strategy is absent, carbon emissions from data centers continue to increase. Under the policy scenario, the carbon emissions of data clusters follow a trajectory characterized by initial growth followed by a decline. Most data clusters are projected to reach their growth inflection points around 2040. However, the Helinger cluster and Shaoguan cluster are exceptions, with inflection points anticipated closer to 2050. Notably, all data clusters exhibit a slowdown in carbon emissions growth, effectively bringing it under control. Specifically, in the 2030 policy scenario, the Qingyang cluster shows the largest percentage increase in carbon emissions (35.58%), whereas the Shaoguan cluster demonstrates the most significant percentage decrease (a 4.44% reduction compared to the 2020 level) (Supplementary Fig. 5). Furthermore, an observable decreasing trend exists across all data clusters, with the Tianfu cluster exhibiting the largest reduction magnitude (42.95% by 2060). Although the implementation of the cross-regional computility transfer strategy facilitates overall reductions in carbon emissions, certain clusters in central and west regions still display upward carbon emissions trajectories.

In the baseline scenario, the contributions of various data center clusters to carbon emissions vary significantly. For instance, the Zhangjiakou cluster and the Yangtze River Delta cluster together account for 70.5% of the total carbon emissions from all data center clusters, while emissions from other clusters, especially in the west region, are relatively low. In the policy scenario, the Zhangjiakou cluster, the Yangtze River Delta cluster, the Helinger cluster, and the Qingyang cluster collectively contribute to 74% of the total carbon emissions across all data clusters. Compared with the baseline scenario, certain data center clusters in the west region have experienced an increase in their proportion of carbon emissions. Notably, clusters with an emissions reduction gap exceeding 70% include the Zhangjiakou cluster (86.8%), the Yangtze River Delta cluster (91.6%), the Shaoguan cluster (122.6%), the Qingyang cluster (89.7%), and the Zhongwei cluster (83.6%). These figures reflect substantial improvements in carbon emissions reduction attributable to the cross-regional computility transfer strategy.

We further conduct attribution analysis to compare the contributions of each data cluster to carbon emissions reduction. The Zhangjiakou cluster and the Yangtze River Delta cluster in the east hub exhibit the largest positive contributions, while the Qingyang cluster in the west hub demonstrates the largest negative contribution (Supplementary Fig. 6). This indicates that although the cross-regional computility transfer promotes carbon emissions reduction overall, there are significant variations in carbon reduction efficiency among different regional data center clusters, which correlates with the direction of the computility transfer.

Cross-regional gap of carbon emissions between the east and west regions

We find that the cross-regional computility transfer strategy significantly reduces carbon emissions from data centers in China, although the extent of this reduction varies between the east and west regions. Analysis of the carbon emissions inventory of the computility hub from 2020 to 2060 reveals that by 2030, the emissions gap for the computility hub in the east region will decrease by 106.6 MT, whereas in the west region it will increase by 89.7 MT (Fig. 2a). By 2060, the emissions gap in the east region will further decrease by 1210.0 MT (85.2%), while in the west region it will decrease by 104.5 MT (38.2%). In 2030, the increase in carbon emissions in the west region, contrary to the overall national trend of declining emissions, indicates that computility transfer from the east to the west increases energy consumption in the latter. Despite the abundance of renewable energy in the west, continuous expansion of computility scale still leads to inevitable carbon emissions (Al Kez et al., 2022; Abderrahim et al., 2024; Al Xu et al.,2024). Furthermore, the carbon emissions gap between the east and west regions under the baseline scenario is projected to reach 194.3 MT by 2030 and 1,146.6 MT by 2060, while under the policy scenario, it is estimated to be 2.0 MT in 2030 and 41.2 MT in 2060. Notably, the cross-regional computility transfer strategy has significantly narrowed the carbon emissions gap between the east and west computility hubs, achieving reductions of 101.0% by 2030 and 96.4% by 2060 compared with the baseline scenario. These outcomes are attributed to the fact that under the cross-regional computility transfer strategy, data processing demands in the east are largely met within the west, effectively curtailing energy usage and related carbon emissions in the east and promoting balanced development between the two regions (Zhu et al., 2023).

Fig. 2: Comparison between the east and west regions under baseline and policy scenarios.
Fig. 2: Comparison between the east and west regions under baseline and policy scenarios.
Full size image

a Comparison of carbon emissions trend of computility hubs in the east and west regions under the baseline and policy scenarios. b Trend of carbon emissions contribution of each computility hub in the east and west regions under the baseline scenario and the policy scenario. Different computility hubs are represented by different colors.

Based on the carbon emissions inventories of the eight computility hubs located in the east and west regions, there are significant variations in the contributions of different hubs to carbon emissions across regions (Fig. 2b). Under the baseline scenario, the Beijing-Tianjin-Hebei hub (36.2%) and the Yangtze River Delta hub (35.2%) contribute the most to carbon emissions in the east, followed by the Guangdong-Hong Kong-Macao hub (10.6%) and the Chengdu-Chongqing hub (1.8%), with a gap of 34.4% between the highest and lowest contributors. In the west, the Inner Mongolia hub (9.1%) and the Guizhou hub (4.5%) contribute the most, followed by the Gansu hub (1.7%) and the Ningxia hub (0.9%), with a gap of 8.2%. Under the policy scenario, the Beijing-Tianjin-Hebei hub (20.6%) and the Yangtze River Delta hub (19.7%) remain the largest contributors in the east, followed by the Guangdong-Hong Kong-Macao hub (5.1%) and the Chengdu-Chongqing hub (4.2%), with a gap of 16.4%. In the west, the Gansu hub (17.8%) and the Inner Mongolia hub (16.5%) become the largest contributors, followed by the Guizhou hub (10.4%) and the Ningxia hub (5.6%), with a gap of 12.2%. It is evident that the cross-regional computility transfer strategy has resulted in reduced carbon emissions in some east hubs while increasing emissions in certain west hubs. Among the east computility hubs, except for the Chengdu-Chongqing hub, the contributions to carbon emissions from the other three hubs have decreased to varying degrees. For the Chengdu-Chongqing hub, receiving computility transmission from other hubs is also an integral part of the cross-regional computility transfer strategy. The changes in carbon emissions contribution results suggest that the allocation of computility resources between the Chengdu-Chongqing hub and neighboring regions is effectively balanced (Abderrahim et al., 2024; Tareen et al., 2024; Istrate et al., 2024). Furthermore, it is noteworthy that the four computility hubs located in the west region have exhibited a substantial increase in their contribution to carbon emissions. This phenomenon may be closely associated with the direction of computility transfer from each computility hub, as depicted in Supplementary Fig. 1.

Comparison of the carbon emissions in different scenarios

In accordance with the findings of the previous study (Uddin et al., 2015), we hypothesize that the scale of computility will sustain its expansionary trajectory, accompanied by a corresponding advancement in technological sophistication over the course of time. Based on the policy scenario, we quantify five varied learning rate scenarios relative to the baseline scenario (see Supplementary Table 1 for details). The following is a brief introduction (Table 1).

Table 1 Brief description of different scenarios.
Full size table

As illustrated in Fig. 3, the dynamic changes in carbon emissions of data centers under various scenarios are clearly depicted. In the BS, carbon emissions increase sharply over time, reflecting the substantial growth trend of carbon emissions driven by the unmitigated expansion of the data center industry. In contrast, the PS, CsS, LurS, WcetS, PueS, and RerS scenarios all effectively suppress the rise in carbon emissions to varying degrees. This demonstrates that the measures associated with these scenarios play a critical role in reducing carbon emissions from data centers. Specifically, we find that the carbon emissions of CsS will exceed those of BS until 2034 and remain higher than those of PS throughout the period from 2020 to 2060. In the years leading up to 2033, the carbon emissions of LurS will consistently exceed those of BS, and from 2020 to 2060, they will remain higher than those in PS. Given the current green level of data centers, the continuous expansion of the computility scale is expected to cause a short-term increase in carbon emissions during the implementation of the cross-regional computility transfer strategy. Furthermore, although improvements in ITLUR are anticipated to mitigate increases in energy consumption, they may still result in greater-than-expected short-term carbon emissions. However, by 2034, it is projected that the carbon emissions under both CsS and LurS will be lower than those under BS but higher than under PS. This suggests that over the long term, both scenarios have the potential to facilitate emissions reduction. Considering the prospects for China’s global computility competitiveness, scale, and status, reducing data center carbon emissions requires advance preparation (Ding et al., 2018; Xu et al., 2024; Fujimori et al., 2021).

Fig. 3: Carbon emissions under the baseline scenario, policy scenario and five varied learning rate scenarios.
Fig. 3: Carbon emissions under the baseline scenario, policy scenario and five varied learning rate scenarios.
Full size image

In addition to the baseline and policy scenarios, five variables containing data center development have been added to compare their impact on carbon emissions.

Furthermore, in response to the sustained increase in electricity demand in western regions, some propose that the utilization of local energy resources in the west should be enhanced while reducing the transmission of clean energy from west to east (Johnson et al., 2024). Taking the aforementioned factor into account, we established a scenario involving reduced clean energy transmission from west to east. Our findings indicate that the carbon reduction effect of this approach on data centers is comparable to that of PS but results in slightly higher carbon emissions than PS. In the context of the Third Industrial Revolution dominated by computility (China Energy News, 2024), the centralized and large-scale development of computility presents a complex challenge. China has long promoted the “West-to-East Power Transmission” Project, aiming to optimize energy resource allocation and enhance clean energy consumption capacity in the east region. However, WcetS contradicts this initiative, lacking both policy support and practical feasibility. Moreover, reducing the eastward transmission of clean energy from the west may compel east data centers to rely more heavily on locally produced high-carbon electricity, further increasing carbon emissions and intensifying environmental pressure. Meanwhile, clean energy production areas in the west region may encounter challenges such as energy surplus and restricted industrial development, which could further widen the digital economic divide between the east and west regions (Glasnovic et al., 2016). Therefore, the WcetS scenario not only deviates from the core direction of energy structure optimization under the “carbon peaking and carbon neutrality” goals but also conflicts with the regional coordinated development strategy. Our results also demonstrate that reducing west-to-east clean energy transmission has a negligible impact on emission reduction. Given its conflict with national strategies, the WcetS scenario is both unfeasible and fraught with multiple risks, necessitating careful consideration in actual policymaking.

Notably, transitioning from PS to PueS and RerS could lead to respective reductions of 5.9% and 6.2% by 2030, with further reductions of 80.0% and 82.0% expected by 2060. Although these two scenarios exhibit similar levels of emission reduction, their mechanisms of action and influencing factors differ significantly. From a regional perspective, the high construction density of data centers in the east and the resulting heat dissipation pressure make it challenging to reduce PUE. However, cross-regional allocation can provide access to green electricity, thereby increasing the proportion of renewable energy. The west region, with its natural cooling conditions, holds substantial potential for PUE optimization. As a region rich in renewable energy, the RerS scenario can better leverage its resource advantages and significantly enhance emission reduction efficiency. From the standpoint of reliance on the baseline energy structure, the proportion of fossil fuel-based power generation in the east region’s baseline energy structure is high. RerS can directly replace energy sources to reduce carbon emission intensity, while PueS can decrease total energy consumption from the demand side and indirectly lower reliance on high-carbon energy. Combining the two approaches comprehensively can establish a low-carbon development path for data centers. Our research findings also indicate that, driven by technological progress, reducing carbon emissions from data centers requires support from increased green electricity usage and reduced PUE. These results are consistent with previous studies (Fan et al., 2023; Istrate et al., 2024). We hypothesize that, against the backdrop of an increased computility scale, IT load utilization rate or reduced west-to-east transmission of clean energy, advancements in renewable energy consumption or PUE optimization in data centers may potentially offset the increase in carbon emissions caused by these factors. However, this hypothesis requires further investigation.

Embodied carbon emissions flow in different scenarios

In comparison with the baseline scenario, both the policy scenario and the five varied learning rate scenarios result in carbon emissions flowing from the east to the west regions (Fig. 4). Numerically, the flows (Unit: MT) for the six scenarios are 41.45, 53.88, 51.02, 33.16, 41.45 and 41.05 (in 2030), respectively, and 19.01, 24.72, 23.40, 15.21, 17.43 and 13.72 (in 2060), respectively. Notably, by 2060, compared to the PS, the CsS and LurS can increase the carbon emissions flow by 30% and 23.08%, respectively. In contrast, the WcetS, PueS, and RerS can decrease the carbon emissions flow by 20%, 0.83%, and 2.98%, respectively. The volume of carbon emissions is largely influenced by the transition to renewable energy sources, improvements in PUE, and other factors. However, the primary drivers of carbon emissions flow remain the transfer of computility and the integration of green energy into the infrastructure of different computility hubs (Uddin et al., 2012; Zhao et al., 2024). It is worth noting that the expansion of the computility scale and the optimization of IT load utilization will lead to an increase in carbon emissions for each computility hub, thereby exacerbating the carbon emissions flow from east to west. Furthermore, when computility is transferred from the east to the west, which is rich in renewable energy, the carbon emissions flow will also decrease accordingly. Although the decline of clean energy from west to east does not reduce the overall carbon emissions of China’s data centers, the increasing proportion of green energy consumption in the west contributes to a reduction in carbon emissions flow. This effect is particularly pronounced as the availability of self-use energy in the west continues to rise.

Fig. 4: Carbon emissions flow under different scenarios.
Fig. 4: Carbon emissions flow under different scenarios.
Full size image

The key to the EDWC Project is the transmission of computility, transferring from the east to the west regions. The inequality impact of the implementation of the cross-regional computility transfer strategy on carbon emissions flow under different scenarios are demonstrated, which is positively correlated with the scale of the computility transfer.

Figure 5 shows that in all scenarios, not all east computility hubs flow their own carbon emissions to the corresponding west hubs, and correspondingly, not all west computility hubs are affected by the carbon emissions flow from the east. The implementation of the cross-regional computility transfer strategy has resulted in carbon emissions flows, with the most significant flow occurring in the Gansu hub and Ningxia hub. The Chengdu-Chongqing hub, classified as an east computility hub, is also subject to varying degrees of carbon emissions flow due to its role in computility transfer from the Yangtze River Delta hub. In contrast, the Inner Mongolia hub and the Guizhou hub in the west remain unaffected by carbon emissions flow, suggesting that they may potentially offset emissions generated by computlity transfers from other hubs through the utilization of their abundant renewable energy resources. Importantly, while these scenarios exhibit similar levels of carbon emissions flow, the PUE and green energy absorption capabilities implied by the computility hubs vary significantly.

Fig. 5: The scale of carbon emissions flow in different computility hubs under various scenarios.
Fig. 5: The scale of carbon emissions flow in different computility hubs under various scenarios.
Full size image

a Inflow and outflow of carbon emissions in the eight computility hubs under the policy scenario. b The impact of computility scale expansion on the inflow and outflow of carbon emissions in the eight computility hubs. c The impact of increased IT load utilization on the inflow and outflow of carbon emissions in the eight computility hubs. d The impact of a 20% reduction in clean energy transmission from west to east on the inflow and outflow of carbon emissions in the eight computility hubs. e The impact of PUE optimization on the inflow and outflow of carbon emissions in the eight computility hubs. f The impact of enhanced renewable energy utilization and green power consumption capacity on the inflow and outflow of carbon emissions in the eight computility hubs.

Sensitivity tests

Based on the analysis of different scenarios, we identified that the PUE of data centers, the carbon emission factors of the power grid, and the renewable energy utilization rate at different hubs are three critical factors influencing the carbon emissions of data centers. A two-factor sensitivity analysis—encompassing PUE—power grid carbon emission factor and PUE—renewable energy utilization rate combinations—was conducted. As shown in Fig. 6, it illustrates the sensitivity of data center carbon emissions with respect to these factors. The positioning symbols in each panel represent benchmark values for 2020. For example, when PUE is 1.5 and the carbon emission factor of the power grid is 0.539 kg CO2/kWh, the total carbon emissions of data centers with a specific computility scale are ~117 MT CO2. However, if the PUE is reduced to 1.3 and the carbon emission factor of the power grid decreases to 0.37 kg CO2/kWh, the total carbon emissions will be reduced to 70 MT CO2 (Fig. 6a). From the variable changes in Fig. 6a, it can be observed that data center carbon emissions increase with higher PUE and carbon emission factors of the power grid, indicating that these two variables are positively correlated with carbon emissions. In Fig. 6b, a higher renewable energy utilization rate and a lower PUE correspond to reduced carbon emissions, suggesting that the renewable energy utilization rate is negatively correlated with carbon emissions, while PUE remains positively correlated with carbon emissions. This is consistent with our initial expectations. Furthermore, in conjunction with relevant studies (China Intelligent Computing Industry Alliance, 2022; Wang et al., 2023), we agree that in regions abundant in renewable energy, changes in the utilization rate of renewable energy have relatively low sensitivity to final carbon emissions. In contrast, for east regions where the proportion of non-fossil energy is relatively high, an increase in the utilization rate of renewable energy has a more significant impact on their carbon emissions.

Fig. 6: Results of sensitivity test.
Fig. 6: Results of sensitivity test.
Full size image

a Two-factor sensitivity test of PUE and grid carbon emission factors. b Two-factor sensitivity test of PUE and renewable energy utilization rate. c Carbon emission changes within the range of data deviation caused by different influencing factors compared with the 2020 baseline. d The absolute change rate of carbon emissions within the data deviation range of different influencing factors.

Furthermore, due to the lack of publicly disclosed data at the computility hub level, this study employs publicly available provincial statistical data (including the carbon emission factor of the provincial power grid and the proportion of renewable energy consumption) as a substitute for each hub. To assess potential deviations arising from data substitution, we performed a sensitivity analysis on key indicators (Fig. 6c, d). The results demonstrate that when various factors fluctuate within their realistic ranges, the deviation in final carbon emissions remains stable within ±20%. Additionally, although deviations in certain indicators may affect the results, the core conclusion—that the renewable energy utilization rate and PUE are the primary determinants of carbon emissions—remains robust. For future research, we aim to further validate this conclusion by incorporating measured data at the hub level.

Discussion and conclusions

Our scenario setting is primarily focused on the implementation of the EDWC Project, establishing five varied learning rate scenarios closely associated with the variables of carbon emissions from data centers. Through our analysis, we have confirmed that the adoption of the cross-regional computility transfer strategy significantly reduces carbon emissions in data centers across China, leading to a westward shift in carbon emissions flow. Furthermore, we have quantified the specific contributions of various factors to this reduction in carbon emissions flow.

There is no denying the importance of policy design decisions (cross-regional computility transfer strategy) in promoting the reduction of national data center carbon emissions and narrowing the carbon emissions gap between the east and west regions. However, our scenario results reveal that while achieving these significant outcomes, the strategy also intensifies the flow of carbon emissions from the east region to the west, which is also an undeniable fact. The emissions reduction potential of the strategy is based on the policy overlay effect (Shan et al., 2018; Tong et al., 2019; Xu et al., 2024), and is associated with factors such as rack scale, rack design power, IT load utilization, PUE value, energy supply, and renewable energy utilization. Under the guidance of macro policies, computility hubs with more stringent targets will identify feasible paths for carbon emissions reduction tailored to their own resource characteristics (Caggiano et al., 2024; Verpoort et al., 2024). The rapid expansion of the computility scale in China has led to increasing demand for computility in the west region. However, energy utilization efficiency in the west remains relatively low. Even new energies such as wind power and solar energy face challenges related to randomness and volatility (Isler-Kaya and Karaosmanoglu, 2023; Lin et al., 2023). Consequently, this situation has resulted in the west being unable to fully leverage 100% low-carbon new energy on the demand side while simultaneously absorbing carbon emissions originally attributed to the east region. Therefore, it is imperative to explore further strategies aimed at reducing carbon emissions and mitigating the flow of carbon emissions from the east to the west during cross-regional computility transfer strategy implementation, which is a matter worthy of thoughtful consideration.

As a crucial performance indicator with a substantial impact on data centers, reducing the PUE value has become a significant challenge for both current and future data centers (Shan et al., 2018). The introduction of the cross-regional computility transfer strategy has increased the proportion of green electricity consumption in data centers. However, energy consumption still continues to rise daily (Lal et al., 2023). This can be attributed to several factors, including the technical challenges of equipment cooling and the need to maintain overall operational efficiency (Ghaddar et al., 2024; Carlsen et al., 2024). Overcoming these challenges poses a considerable obstacle for data center operators. It is reasonable to anticipate that with the continuous advancement of technology and growing awareness of energy conservation, the adoption of advanced energy-saving technologies and strategies could effectively lower the PUE value, thereby contributing to reduced carbon emissions (Lu et al., 2022; Chen et al., 2024).

Considerably, increasing the west region’s own energy consumption and reducing the clean energy transmission from west to east can effectively reduce the carbon emissions flow caused by the cross-regional computility transfer strategy. However, despite the stark contradiction of electricity price, there remains a significant demand for power in the east due to its continued role as the focal point of national economic development and the primary contributor to China’s GDP (Di et al., 2023; Jia et al., 2023). In addition, the change of energy supply and demand pattern and the increase of environmental pressure make it a topic worthy of discussion to reduce the clean energy transmission from west to east (Jin et al., 2021). On this basis, we propose that the eastern region could fully leverage its own energy resources to enhance self-sufficiency (Zhang et al., 2024). In response to the carbon emissions flow, the western region could expedite the construction of multi-energy complementary new energy bases (National Energy Administration, 2024). This could be achieved by coordinating the construction of these bases with data centers, thereby enhancing new energy utilization and improving power security capabilities. By establishing data center related industrial links, the aforementioned strategies could not only bolster the west’s competitive edge in low-carbon growth but also achieve coordinated development with the east (Sun et al., 2023).

There are certain limitations and uncertainties in our work. Due to the constraints of data availability, we estimated the carbon emissions of the cross-regional computility transfer strategy by using provincial data from the locations of each computility hub as proxy data, as the detailed data lists for individual computility hubs have not been disclosed. Although the expanded carbon emissions of each computility hub may lead to higher final estimates, this has minimal impact on the conclusions of this study. We anticipate that relevant authorities will release such information in the future as the cross-regional computility transfer strategy becomes more mature. Furthermore, in conjunction with previous studies, we can confirm that the transmission and utilization costs of renewable energy, energy storage costs, and the resource endowments of different regions significantly influence the adoption of local renewable energy. However, in this study, we did not investigate the sub-variables of these influencing factors but instead focused on identifying which factors have a more substantial impact on the carbon emissions of data centers. Therefore, the carbon emissions assessment in our research was based on the general trends of changes in each factor in reality to determine the ranking of influencing factors. In future research, we will emphasize integrating more accurate prediction trends of various factors for the final carbon emission predictions, thereby providing theoretical references for related studies.

Methods

Data center energy consumption and carbon emissions estimation

There are two primary calculation models used for estimating data center energy consumption: based on extrapolation (Chen et al., 2024; Vallejo et al., 2024) and bottom-up analysis (BUA) method (Wang et al., 2023; Istrate et al., 2024). The extrapolation method estimates total energy use by scaling bottom-up values based on market growth metrics such as data center investment and global IP traffic. While this approach is relatively simple, it tends to be less accurate (Wang et al., 2024; He et al., 2024). In contrast, the BUA method estimates total energy consumption by analyzing the detailed energy usage characteristics of each device in the data center. However, this method requires substantial data and time investment (Yu et al., 2024; Zhang et al., 2024). Given these considerations, we have chosen the BUA model for our energy consumption calculations. Nevertheless, due to the limited development time of the EDWC Project, key indicators such as the rack scale and other specifications of each computility hub remain undisclosed. Therefore, we use provincial-level data from the locations of the computility hubs as a proxy for estimation (China Intelligent Computing Industry Alliance, 2022). Additionally, data center energy consumption is influenced by several factors, including rack scale, single rack design power, IT load utilization, and power usage effectiveness (Istrate et al., 2024; Verpoort et al., 2024). Based on these factors, the total energy consumption is calculated using Eq. (1).

$${E}_{dc}=\mathop{\sum }\limits_{i=1}^{n}{N}_{i}\cdot PU{E}_{i}\cdot I{T}_{i}\cdot {P}_{dci}\cdot t$$
(1)

where \({E}_{dc}\) is the data center electrical energy consumption (kWh), \(n\) is the number of regions, \({N}_{i}\) is the total number of racks in the \(i\) region, \(PU{E}_{i}\) is the average power usage effectiveness of the data center in the \(i\) region, \(I{T}_{i}\) is the average IT load utilization of data centers in the \(i\) region (%), \({P}_{dci}\) is the average single-rack design power in \(i\) region (kW), and \(t\) is the time (h).

Carbon emissions are embedded in all stages of the data center life cycle, including energy consumption during operation as well as the construction, production, transportation and scrapping of data centers (Dayarathna et al., 2016; Habibi et al., 2017; Xu et al., 2024). The emissions from these non-operational stages can be reduced by using green building materials and enhancing building energy efficiency (Li et al., 2023; Ding et al., 2024). In this study, we focus exclusively on the carbon emissions primarily associated with electricity consumption during data center operation (Eq. (2)).

$$C={E}_{dc}\cdot G\cdot (1-\alpha )$$
(2)

where \(C\) is the data center carbon emissions (kg), \({E}_{dc}\) is the data center energy consumption (kWh), \(G\) is the grid emissions factors of carbon emissions (kgCO2/kWh), \(\alpha\) is the renewable energy utilization rate (%).

Full set of scenarios

Baseline scenario

To investigate the impact of implementing the cross-regional computility transfer strategy on data center carbon emissions, we constructed a baseline scenario that excludes the influence of this strategy. In this scenario, the rack scale is determined according to the decreasing mechanism (see Supplementary Table 1 for detailed parameters) (Uddin et al., 2015). The IT load utilization, rack power, and grid carbon emission factors are assumed to remain constant at their 2020 levels. The PUE value is annually optimized in line with the current trend, and the renewable energy utilization rate is allocated based on the photovoltaic and wind power consumption responsibility targets proposed by the National Energy Administration for each province until 2035 (National Energy Administration, 2021). This means that the scenario is only influenced by other national policies unrelated to the cross-regional computility transfer strategy.

Policy scenario

The policy scenario is constructed by incorporating the changes in various indicators proposed by the cross-regional computility transfer strategy and adheres to the setting mode of the baseline scenario, with adjustments made to the rack scale, IT load utilization, rack power, and renewable energy utilization rate. The rack distribution between the east and west regions gradually shifts from ~8:2 to ~6:4 (Guy et al., 2023; Istrate et al., 2024). The power grid carbon emission factors for each cluster during 2020-2035 are predicted based on Ref. (China Hua-Neng Carbon Neutral Research Institute, 2023), and the emissions factor after 2035 continues to follow the current trend. According to the PUE targets set by the cross-regional computility transfer strategy, the PUE of east data centers will decrease to 1.25 by 2030, while the PUE of west data centers will decrease to 1.2 by 2025 (The Ministry of Industry and Information Technology, 2021). The continuous evolution of these indicator values reflects the sustained efforts following the implementation of the policy.

Varied learning rate scenario

The varied learning rate scenario is constructed based on the policy scenario, and the following five varied learning rate scenarios are sequentially established, assuming that all favorable practices in the policy scenario remain unchanged (see Supplementary Table 1 for additional sources of scenario settings). A. The growth rate of the computility scale is set to 30%. B. Due to advancements in networking, resource scheduling, and new technologies (Ding et al., 2018), the IT load utilization is set to reach 80%. C. The clean energy transmission from west to east will be reduced by 20%. D. The PUE in the east and west regions will reach 1.1 in 2050 and 2040, respectively. E. The renewable energy utilization in the east (west) region is gradually increasing to 50% (60%) in 2040 (National Energy Administration, 2021). Additionally, the renewable energy utilization in the Tianfu cluster reaches 90% by 2040, and the Chongqing cluster reaches 60% by 2040. Although these scenarios may deviate from realistic trends, they allow us to explore the drivers behind differences in emissions reduction and potential pathways to address carbon emissions flow.