Key Takeaways
- Data-centre exposure after 2020 is associated with a measurable increase in the operating life of coal-fired generating units.
- The 937 exposed coal units could produce an additional 1.96 GtCO₂ between 2025 and retirement.
- Amazon, Microsoft, Google, and Meta face growing pressure to connect AI expansion with local grid emissions and retirement decisions.
A rapid expansion of data-centre capacity is beginning to alter the economics of coal power, complicating efforts to retire older generating units as artificial intelligence pushes electricity demand higher.
A study published in Nature Climate Change examined 9,379 coal-fired units and found that exposure to data-centre growth after 2020 was associated with a 1.7% longer plant lifespan for every standard-deviation increase in exposure. The analysis covered 259,044 unit-year observations and 2,913 retirement events.
That percentage may sound modest. Across hundreds of power plants, however, even relatively short extensions can translate into substantial emissions. The researchers project that 937 exposed coal units will emit an additional 1.96 GtCO₂ between 2025 and their eventual retirement. The 95% confidence interval ranges from 1.01 GtCO₂ to 2.92 GtCO₂.
The finding does not mean every new server hall automatically saves a nearby coal plant. Electricity markets differ, grid connections take time, and utility planners consider reliability, transmission capacity, regulation, and fuel costs. Still, the association suggests that concentrated data-centre demand can change retirement calculations, particularly where replacement generation and grid infrastructure are not arriving quickly enough.
Interestingly, the effect was stronger among coal plants with lower emissions intensity and better thermal efficiency. Those facilities may be comparatively attractive to utilities looking for dependable capacity while they connect large new loads. In practical terms, the coal units most likely to survive are not necessarily the oldest or least efficient. They may be the plants that remain economically credible for a few more years.
The demand trajectory adds urgency. The International Energy Agency reported that global data-centre electricity consumption rose 17% in 2025, with AI-focused facilities growing even faster. Consumption is projected to increase from 485 TWh in 2025 to roughly 950 TWh in 2030. That is close to a doubling within five years.
In the United States, the generation mix serving data centres in 2024 was more than 40% natural gas, about 25% renewables, roughly 20% nuclear, and close to 15% coal, according to the IEA. So, while coal represents a minority share, it remains meaningful. Natural gas expansion also introduces a second emissions challenge as utilities respond to large, round-the-clock loads.
Amazon, Microsoft, Google, and Meta are the principal hyperscale developers behind the demand surge. Utilities serving their projects have delayed some coal retirements and expanded gas capacity. This creates a tricky gap between corporate clean-energy procurement and the physical operation of regional grids. A company can contract for renewable electricity while its growing load still contributes to keeping dispatchable fossil generation online.
Annual renewable-energy matching does not necessarily reveal what happens in a specific grid during constrained hours. Are new clean resources being added in the same region and at the same time as the computing demand, or are certificates offsetting consumption on paper while local fossil plants run longer?
The University of Chicago Climate Systems Engineering has also highlighted the risk that Big Tech’s data-centre expansion could increase emissions by affecting power-sector investment and retirement choices. For technology buyers, investors, and regulators, this shifts attention from corporate procurement totals toward additional generation, hourly matching, transmission availability, and the marginal power plants serving each project.
Existing standards provide a starting point, though not a complete answer. The GHG Protocol Scope 2 Standard supports accounting for purchased electricity, while ISO 50001:2018 offers a structured approach to energy management. Data-centre operators can build on those frameworks by disclosing regional load growth, hourly electricity use, expected generation additions, and any retirement delays connected to their projects.
That said, the commercial pressure to deploy AI infrastructure remains intense. The near-term question is not whether data centres will consume more power, but how quickly clean generation, storage, transmission, and flexible computing can catch up. If those investments lag, utilities may keep efficient coal plants running and add gas capacity, turning the AI boom into a longer-lived fossil-fuel commitment than corporate emissions reports currently suggest.
⬇️