Low-Carbon Cities Boost AI Growth in China by Linking Clean Energy and Innovation

Researchers found that pilot cities hosted approximately 16.9% more artificial intelligence (AI) enterprises than comparable non-pilot cities, indicating that well-designed environmental policies can foster technological innovation while supporting green-transition objectives.

The findings challenge the assumption that stricter environmental regulations hinder the growth of technology-intensive industries by increasing operational costs. Instead, they provide evidence that targeted regulation and a low-carbon policy framework can create favorable conditions for digital transformation and the expansion of high-tech industries.

Balancing Climate Commitments and Energy Demands

Modern metropolitan areas are significant centers of economic activity, but also account for approximately 70% of global energy-related carbon dioxide emissions, while Chinese urban areas account for about 70% to 80% of national carbon emissions. In response, China has set ambitious climate goals, including peaking carbon emissions before 2030 and achieving carbon neutrality before 2060.

The rapid growth of AI increases the demand for energy-intensive computing infrastructure, including large data centers. Balancing these climate commitments with rising electricity needs poses a significant challenge for urban planning. This necessitates cleaner energy systems and supportive policies.

Methodology: Analyzing the LCCP's Influence

To evaluate the influence of low-carbon policies on high-tech development, researchers analyzed panel data from 285 Chinese cities spanning 2007 to 2022. They employed a staggered difference-in-differences framework, treating the phased implementation of the LCCP in 2010, 2012, and 2017 as a quasi-natural experiment.

To measure industry changes, the study used Python-based web scraping and fuzzy keyword matching to collect data on over 3.8 million enterprises from the Qichacha corporate database. The researchers also used the QWEN-7B large language model to classify companies into three value-chain segments: the upstream infrastructure layer, the midstream technology layer, and the downstream application layer.

The econometric models controlled for various city characteristics, including population size, economic density, GDP (gross domestic product) per capita, financial development, trade openness, education levels, and fixed-asset investment.

Furthermore, to strengthen causal inference and reduce selection bias, researchers combined Propensity Score Matching with two-stage least-squares instrumental-variable estimation. They used local government environmental rhetoric and historical levels of industrial pollution as instruments.

Key Findings: Channels of AI Development

The baseline regression analysis found that the LCCP was significantly associated with the growth of AI enterprises, with a policy coefficient of 0.156 in the fully specified model. Parallel trend tests indicated that the policy effect emerged with a one-year lag and became statistically significant in the second year. This reflects the adjustment time required for local industries to respond to and leverage the new policy framework. Robustness tests, including placebo analyses and the exclusion of data from the 2020 pandemic period, supported the robustness of the findings.

Further analysis showed that the policy promoted AI development through two proposed channels: cleaner energy use and increased green technology innovation.

The LCCP was associated with a reduction in coal-based energy reliance (coefficient = -0.112), while lower coal reliance was associated with higher AI industry development (coefficient = -0.037). The policy was also associated with increased green technology innovation (coefficient = 0.161), and green technology innovation was linked to overall AI industry growth (coefficient = 0.035).

The impact varied across the AI value chain, with downstream application firms benefiting more from cleaner energy adoption. In contrast, midstream firms focused on model training and algorithm development experienced short-term adjustment frictions due to their reliance on stable computing resources. The effects also varied by local context, with stronger associations reported in some innovation hubs and transition-oriented regions.

Practical Implications for Energy and AI Infrastructure

These outcomes provide practical guidance for balancing clean energy goals with the growth of the AI industry. Upstream companies that operate large data centers and midstream firms engaged in algorithm development and large-scale model training depend on a stable electricity supply and could benefit from dedicated renewable energy sources.

Regional planners can support these industries by providing reliable green power supplies and time-of-use electricity pricing that reduce energy costs during continuous computing workloads. 

Additionally, incentives for energy-efficient chip design could accelerate the adoption of smart microgrids and distributed computing infrastructure, helping to reduce emissions while supporting continued growth in the AI sector.

Conclusion: Integrating Sustainability with Technology

In summary, this study shows that well-designed low-carbon policies can promote both environmental sustainability and high-tech industrial growth. Rather than acting solely as regulatory constraints, carbon reduction policies can encourage industrial upgrading, accelerate technology adoption through green knowledge spillovers, and effectively support the transition toward a more sustainable digital economy in China’s urban policy context.

The findings suggest that climate policies should be integrated into broader industrial development strategies while accounting for the different energy needs across the technology value chain.

Although the analysis is limited to city-level data, historical business registrations, and AI-related enterprise counts rather than firm output, employment, revenue, or patent quality, future research could enhance these results by incorporating firm-level production data and long-term indicators such as patent quality. The authors also noted that the findings may not directly generalize to other countries or to the post-2022 generative AI period.