China's AI industry carbon footprint more than doubled to 324 Mt CO2 by 2023, study finds

Indirect emissions, those generated across supply chains, reached 187.01 Mt in 2023 and stayed above direct emissions. That shifts attention from model and data-centre electricity use to the industries that supply AI, including electricity and heat, transport, metals and non-metallic minerals.

Between 2020 and 2023, domestic final demand was associated with a 138.82 Mt increase in emissions, while changes in production structure cut 20.24 Mt. The authors argue that efficiency gains alone are not enough and call for demand management, low-carbon electricity, hardware circularity and targeted upstream procurement.

The researchers disaggregated China's AI industry from the broader ICT sector and built comparable 31-sector environmentally extended input–output accounts for 2015, 2017, 2020 and 2023. They combined embodied-emission accounting with linkage analysis, structural path analysis and structural decomposition analysis.

The results show most of the footprint sits in supply chains rather than in direct operations. Roughly half of it was concentrated within the first three production tiers, with other ICT, electricity and heat, transport, metals and non-metallic minerals forming the largest short paths.

In the current-price decomposition for 2020 to 2023, domestic final demand was associated with a 138.82 Mt rise, outweighing a 20.24 Mt reduction from production-structure change. The authors frame AI decarbonisation as a systems-governance problem requiring demand management, low-carbon electricity, hardware circularity and targeted upstream procurement.