The AI emissions nobody counts are the ones it helps oil companies produce

Data centre power draw has absorbed most of the argument about AI and the climate. A paper published last week in npj Climate Action makes the case that it is the smaller half of the problem — and that the larger half is barely being measured at all.

The research models what happens as the oil and gas industry becomes more productive through AI. Its central finding is that the resulting emissions could add somewhere between 1.2% and 4.8% to global energy-related emissions. At the bottom of that range, the yearly addition is roughly what Mexico emits. At the top, it is roughly Russia, the fourth-largest emitter on earth.

Both ends of the range exceed the projections for the data centre buildout itself. Both also, on the authors' figures, outweigh whatever AI contributes to developing solar, wind and other clean technology — which is the argument usually deployed to close the question down.

The authors are Will and Holly Alpine, who worked in sustainability at Microsoft and left at the start of 2024 over the company's continuing business with oil and gas. They have campaigned publicly on the connection since.

Their framing is that supply and demand here reinforce each other and cannot sensibly be assessed apart — two sides of one coin, as Will Alpine puts it. The industry supplying the power to train the models is also the industry the models are making more efficient.

That efficiency is not a new development. Oil and gas companies have used forms of AI for decades to locate and develop reserves at lower cost, which deepens global reliance on the fuel and pushes climate targets further out of reach.

What is new is the attempt to put a figure on it. Technology companies account for their own emissions and those of their supply chains; almost none account for how much extra fossil fuel their tools help bring out of the ground. The Alpines call that category enabled emissions, and their argument is that corporate sustainability reporting, by concentrating on operational numbers, omits the part doing the most damage.

The estimate comes from a macroeconomic model, fed with oil and gas companies' own published claims about what their AI deployments have delivered, and applied across extraction, refining and power generation. Will Alpine describes the scale that came out of it as staggering.

The Chevron arrangement announced with Microsoft illustrates the loop about as plainly as it could be illustrated. Chevron is putting up a large gas plant in Texas, behind the meter, to supply Microsoft's data centres. On an analyst call in June, Jeff Gustavson, who heads Chevron's New Energies division, indicated the benefit would run both ways — that Chevron would take some of the compute the plant enables and use it to run AI inside its own business. A Chevron spokesperson, asked about the relationship, pointed to years of joint work on digital transformation and cloud capability.

Jon Koomey, an energy researcher with no involvement in the study, told WIRED its conclusions look solid, and was blunter about the reasoning it displaces. Claims that AI will solve the climate problem, and that development should therefore proceed as fast as possible, ignore what the technology does across every industry rather than only in renewables and efficiency. Machine learning can make cooling a data centre 30 to 40 percent more efficient; it can equally make extracting oil cheaper and quicker. Which effect wins, he says, nobody yet knows for certain.

That uncertainty is the finding rather than a hedge on it. The emissions AI causes are counted, argued over and offset. The emissions it enables are not counted at all — and on this modelling, they are the bigger number.