What the study measured
The rapid build-out of AI data centres has produced a large new source of electricity demand, and the obvious question is where that power comes from. Researchers led by Chenglin Li and Yangsiyu Lu examined global datasets covering 9,379 coal-fired generating units and more than 9,000 data centre facilities, then tracked which plants stayed open.
Their survival analysis covered 259,044 unit-year observations and 2,913 retirement events. The finding is that coal plants exposed to data centres built after 2020 have a measurably lower risk of shutting down. Each standard-deviation increase in that exposure was associated with roughly a 1.7% extension of plant lifespan. The paper was published in Nature Climate Change on 10 November.
Emissions from delayed retirement
A 1.7% extension sounds small until it is multiplied across a fleet. The authors project that the 937 exposed units will emit an additional 1.96 gigatonnes of CO2 between 2025 and the end of their operating lives, with a 95% confidence interval of 1.01 to 2.92 gigatonnes. That is roughly the annual emissions of a large industrial economy, added to the atmosphere because plants kept running instead of closing.
The relationship was stronger for units with lower CO2 emission intensity and higher thermal efficiency. That pattern suggests the mechanism is economic rather than political: efficient plants are the ones a grid operator or utility would prefer to keep online when demand rises, and new load nearby makes the case for keeping them.
Pressure on net-zero plans
The result complicates the assumption that data centre demand can be met with new renewables and storage alone. Grid connections take years; existing coal capacity is already wired in. Companies that sign long-term power agreements for AI facilities are now being asked to account for the local generation mix they support, not just the clean energy they buy.
The authors argue that digital infrastructure planning has to be folded into national climate pathways. If it is not, the emissions from the AI build-out will show up as an unexplained gap between modelled coal retirement schedules and what utilities actually do.