Vertical Evolutionary Game Mechanism between R&D and Application in the AI Supply Chain under Low-Carbon Coordination

• In Event: Lightning Session: Natural Resource, Energy, and Environmental Policy

Abstract

Artificial intelligence (AI) is reshaping energy systems and carbon emission patterns. Yet its rapid growth also relies on energy-intensive computing infrastructure, large-scale model training, and data center operations. AI therefore combines efficiency gains with scale-driven emission growth, creating a classic Jevons paradox, where technological progress expands demand and offsets environmental benefits. This problem is especially important in AI supply chains. Upstream R&D sectors determine the energy efficiency of technologies, while downstream application sectors capture most gains from market expansion. Different incentives across stages may lock the system into a high-performance, high-emission path.
To study this issue, we develop an evolutionary game model with two agents, an upstream AI R&D sector and a downstream application sector. The market expansion effect caused by technological progress is built into the payoff structure. We compare three policy settings: no carbon constraints, uniform carbon constraints, and differentiated carbon constraints. We examine whether regulation can induce low-carbon coordination and which policy design performs better.
Four main findings emerge. First, market forces alone cannot solve the Jevons paradox. Without policy intervention, efficiency gains and demand expansion raise the returns to high-carbon strategies, crowd out low-carbon choices, and create path dependence. Second, governance resistance is concentrated downstream. Downstream sectors directly receive scale gains from AI adoption, so the cost of giving up high-carbon strategies is higher than in upstream sectors. As a result, the critical carbon cost needed to trigger strategic change is systematically higher downstream, making it the main barrier to supply-chain decarbonization. Third, the effect of technological gains on carbon constraints differs across sectors. In upstream sectors, larger technology rents weaken policy effectiveness by raising the critical carbon cost. In downstream sectors, once carbon constraints pass a critical threshold, technological progress strengthens policy effects and lowers the cost of low-carbon transition. Fourth, differentiated carbon constraints perform better than uniform regulation. By matching policy intensity to sector-specific thresholds, differentiated regulation avoids unnecessary compliance burdens under one-size-fits-all rules. It can induce a coordinated low-carbon equilibrium and reduce total policy costs by about 24.6% under the baseline calibration, achieving a Pareto improvement in environmental and economic performance.
This study contributes in three respects. It extends the Jevons paradox to the AI supply-chain context by identifying the micro-mechanism through which technology dividend translates into rebound emissions via vertical interactions. It provides quantitative thresholds linking policy intensity, technological gains, and market expansion, moving beyond the binary view that technological progress necessarily increases or decreases emissions. It also reveals the structural relocation of governance resistance toward downstream sectors and demonstrates the superiority of differentiated carbon constraints, offering practical guidance for targeted environmental governance in the AI industry.

Author

• Yimeng Wang, Harbin Institute of Technology
  Presenting Author