Summary
- FT analysis estimates 60 major planned facilities could be associated with 101.5m tonnes of annual CO₂ at full operation on the current US power mix.
- Around three-quarters of the utilities serving the analysed projects are reported to be building additional gas generation.
- The comparison is relevant to Europe because data centre carbon performance increasingly depends on what generation is added to serve new load, not only on contractual renewable procurement.
Sixty of the largest US data centres planned by Amazon, Microsoft, Google, and Meta could be associated with 101.5m tonnes of carbon dioxide emissions annually if powered by the current US generation mix, according to a Financial Times analysis.
The estimate is equivalent to roughly 7% of US power-sector emissions in 2025, based on the methodology used in the analysis. It does not represent measured emissions from operating facilities, and the generation mix supplying the projects could change substantially before all 60 sites reach full operation.
The value of the exercise is therefore not as a precise forecast but as an indication of the scale of electricity demand being added to the system. Data centre emissions increasingly depend on what generators are built or kept operating because of new load, rather than simply the annual renewable-energy contracts reported by individual technology companies.
The FT found that around three-quarters of the utilities serving the analysed sites are building new gas-fired generation, while some are delaying coal-plant retirements. That creates an uncomfortable interaction between corporate decarbonisation targets and the reliability decisions utilities are making to accommodate rapid demand growth.
The technology companies continue to procure large amounts of renewable power and invest in new energy technologies. But their own reported emissions have also been rising as data centre construction and electricity consumption accelerate.
Additional demand changes the accounting
Traditional renewable procurement can match a company’s annual electricity use with contracted clean generation, renewable certificates, or power-purchase agreements. That accounting does not necessarily show what happens on the local grid during the hours when a new data centre is consuming electricity.
A large facility requiring round-the-clock power can increase demand for firm generation even where its owner has contracted renewable supply elsewhere. If transmission constraints prevent that clean generation from reaching the site, or if wind and solar output is unavailable during peak hours, gas or other dispatchable plant may still be needed to balance the system.
This is becoming a central issue for European data centres as well. The EU’s energy-performance reporting regime is increasing visibility of facility-level efficiency and water use, but broader carbon consequences depend on electricity-system development outside the data centre boundary.
Markets with abundant low-carbon generation can accommodate new load with a different emissions profile from regions where additional demand triggers new fossil capacity. That partly explains the appeal of the Nordics, France, and other markets with comparatively low-carbon electricity, although grid connection availability remains a separate constraint.
The distinction also makes additionality important. A developer signing a renewable PPA can support new generation, but the strongest claim comes where the project demonstrably causes new low-carbon capacity, storage, or network infrastructure to be built rather than allocating existing clean output to a new customer.
Behind-the-meter generation complicates the picture further. Gas turbines can allow a data centre to operate without waiting for grid reinforcement and can provide firm capacity, but they move emissions directly into the project’s infrastructure strategy.
Nuclear, storage, demand flexibility, and long-duration clean-power contracts are all being explored as alternatives, yet none can be deployed everywhere at the same speed as AI compute demand.
The 101.5m-tonne estimate should therefore be treated as a scenario rather than a fixed outcome. The relevant question is what the power system looks like by the time the facilities operate. If clean generation, transmission, and storage expand quickly enough, the realised emissions can be lower. If utilities respond primarily with gas and delayed fossil retirements, the carbon consequence of the AI build-out becomes materially larger.
Facility efficiency remains important, but PUE alone cannot answer the environmental question. The source, location, and timing of the additional electricity required for new data centre capacity increasingly determine the outcome.
