Digital technologies measurably raise energy efficiency at China’s listed firms, with the largest improvements among the worst-performing companies; combining AI, blockchain, cloud and big data delivers bigger gains than any single technology, but increased usage also produces a detectable rebound in energy consumption.

Main Finding

Digital technologies — measured as firm-level adoption of AI, blockchain, cloud computing, and big data — significantly improve firm-level energy efficiency (EE) across the distribution of firms (10th–90th quantiles). Effects are largest for the least efficient firms (10th quantile). A composite digital-technology index yields stronger EE gains than any single technology. However, a measurable energy rebound effect is observed, partially offsetting efficiency gains. Results are robust to multiple estimation approaches.

Key Points

• Sample and scope: 2,003 Chinese A‑share listed firms; panel observations 2013–2021 (18,027 obs.).
• Technologies studied: artificial intelligence (AI), blockchain, cloud computing, and big data; also a composite DT index combining these dimensions.
• Main empirical approach: method of moments quantile regression (MMQR) to estimate heterogeneous effects across EE quantiles (10th to 90th).
• Principal empirical findings:
  • Individual DTs and the composite DT index are positively and significantly associated with firm-level EE across nearly all quantiles (mostly at the 1% level).
  • The largest marginal EE gains appear for firms in the lower tail of the EE distribution (10th quantile).
  • The composite DT index consistently outperforms any single-technology indicator in magnitude of EE improvement.
  • An energy rebound effect is detected, indicating some offsetting increase in energy use associated with DT adoption/scale.
• Robustness: results confirmed using alternative estimators — fixed-effects unconditional quantile regression (FEUQR), panel fixed-effects quantile regression (PFEQR), and instrumental variables (IV) approaches.

Data & Methods

• Data: Firm-level panel of 2,003 Chinese listed companies (A‑shares), 2013–2021; 18,027 observations. (Paper does not provide detailed variable-by-variable measurement in the excerpt—technology indicators are treated as firm-level DT adoption indices for AI, blockchain, cloud, and big data, plus a composite index.)
• Outcome variable: firm-level energy efficiency (EE) — defined broadly as reduced final energy consumption while maximizing energy services.
• Key explanatory variables: four firm-level DT indicators (AI, blockchain, cloud, big data) and a composite DT index.
• Estimation strategy:
  • Primary: Method of moments quantile regression (MMQR) following Machado & Silva (2019) to capture conditional quantile heterogeneity in the DT–EE relationship (estimated for 10th–90th quantiles).
  • Robustness: FEUQR, PFEQR, and IV methods to address potential bias and endogeneity concerns.
• Findings emphasize heterogeneity (distributional impacts) and compare single-technology vs composite-technology effects.

Implications for AI Economics

• AI as strategic firm resource: Empirical support for the resource-based view — AI and related DTs function as strategic capabilities that materially improve firm-level EE, implying AI investments should be valued not only for productivity but also for energy-efficiency returns in firm-level cost–benefit analyses.
• Heterogeneous returns: The largest EE gains occur for low‑efficiency firms, so economic models and policy evaluations should account for distributional heterogeneity (quantile effects). Aggregate average effects understate benefits to the most energy‑inefficient firms.
• Composite adoption and complementarities: Composite DT adoption (integration of AI with blockchain, cloud, big data) yields larger EE gains than isolated AI deployment. Models of technology adoption and diffusion should incorporate complementarities among digital technologies when estimating returns on AI investments.
• Rebound/externalities: The observed energy rebound implies that energy-saving technologies (including AI systems) can induce additional energy demand elsewhere (e.g., via increased activity, data-center loads). AI-economic analyses must internalize these second‑order effects (net energy impacts), and welfare calculations should include the costs of rebound.
• Policy design for AI & climate objectives:
  • Encourage integrated digitalization strategies and cross-technology complementarities rather than narrow AI subsidies alone.
  • Pair AI/digital adoption incentives with measures to limit rebound (e.g., energy pricing, carbon pricing, energy-efficiency standards for data centers, green procurement).
  • Target support toward the least energy-efficient firms to obtain the largest EE dividends per unit of digital investment.
• Research implications:
  • Future microeconomic models should incorporate quantile-heterogeneous treatment effects of AI adoption.
  • Need for better measurement of firm-level DT adoption (granular indicators) and tracking of energy use of AI-related infrastructure (data centers, edge devices).
  • Explore interactions between AI deployment, energy prices, regulatory regimes, and the lifecycle energy footprint of AI systems to estimate net climate impacts.

Limitations (noted): sample limited to Chinese listed firms and 2013–2021 period; measurement details of DT indicators not fully specified in the excerpt; observational design despite IV robustness checks. Future work should test generalizability across countries, sectors, and post-2021 technological developments (e.g., large foundation models).