Most of the argument about AI and the climate has focused on hungry data centers. A new peer-reviewed study says that is the small half of the problem: by making oil and gas extraction cheaper and faster, AI may push global energy emissions up by as much as 4.8 percent – more damage than the entire data center buildout, and more than AI saves through better clean-energy design.
The Research Behind the AI Emissions Warning
The paper, published in the journal npj Climate Action, comes from Will and Holly Alpine, a married pair of former Microsoft sustainability staffers who resigned in early 2024 over the company’s continuing commercial work with the oil and gas sector. Since leaving, they have campaigned publicly on the argument that the cloud and AI industry’s climate footprint cannot be measured honestly while ignoring who its biggest industrial customers are.
Their modelling treats AI not as an energy consumer but as a productivity multiplier applied to the fossil fuel supply chain – exploration, drilling, refining and thermal power generation. Feeding in the efficiency gains that oil and gas firms themselves have publicised, the authors estimate the extra global energy-related emissions unlocked by that productivity at between 1.2 and 4.8 percent per year.
What Those Numbers Look Like in Practice
Percentages are easy to shrug off, so the authors translate them. At the conservative end, the additional annual emissions are roughly equivalent to the total output of Mexico. At the upper end they approach the annual emissions of Russia, the world’s fourth-largest emitter. That is an entire industrial nation’s worth of carbon added to the atmosphere as a side effect of software that never appears on any tech company’s carbon ledger.
Enabled Emissions: The Category Nobody Reports
The study’s central concept is what the Alpines call “enabled emissions” – pollution that a technology makes possible elsewhere in the economy. Corporate sustainability reporting today is built around operational emissions and supply-chain emissions. A cloud provider will count the electricity in its own halls and, increasingly, the carbon embedded in its hardware. Nothing in that framework asks whether the machine-learning models sold to a drilling company caused more hydrocarbons to be lifted out of the ground.
Why the Accounting Gap Matters
Because the gap is structural, it survives good intentions. A firm can hit a net-zero operational target, buy clean power for every rack, publish an immaculate disclosure – and still be the single largest efficiency contributor to an expanding extraction industry. Until enabled emissions are named and estimated, that outcome is invisible by design rather than by dishonesty.
The Feedback Loop Between AI and Oil
What makes the finding awkward is that the relationship runs both ways. Will Alpine describes it as “a self-reinforcing effect between supply and demand,” arguing the two cannot be assessed separately. AI helps produce cheaper gas; cheap gas is then used to power the data centers that train and serve more AI. Chevron and Microsoft have confirmed a large behind-the-meter gas plant in Texas built to supply Microsoft data centers, and Chevron executives have indicated the company expects to use some of that compute for its own internal AI work. The loop closes neatly.
Does AI’s Clean-Tech Upside Cancel It Out?
This is the strongest counter-argument, and the paper addresses it head-on. Machine learning genuinely helps with grid balancing, wind and solar siting, battery chemistry search and materials discovery. But the modelled benefits on that side of the ledger are smaller than the modelled fossil-side gains. Extraction is a mature, capital-intensive industry where marginal efficiency converts almost immediately into extra barrels; clean-tech gains arrive slowly and often bottleneck on permitting, transmission and manufacturing rather than on analysis.
Oil and Gas Has Used AI for Decades
It is worth stressing that none of this is new technology in the oilfield. Seismic interpretation, reservoir simulation and predictive maintenance have used statistical learning for a long time. The generative AI boom did not create the practice, it accelerated and commercialised it – packaging capability that once required in-house specialists into services any operator can buy. Scale, not novelty, is what turns an old tool into a climate variable.
What Would Change This
Practical responses are less exotic than the headline suggests: disclosure standards that require enabled-emissions estimates, contract-level transparency about which industrial sectors are being served, and procurement rules from public buyers that treat fossil-enablement as a factor. Any of these would force the conversation past the electricity bill of the data hall.
Outlook
Expect pushback on the modelling assumptions – economy-wide models are sensitive inputs, and the 1.2-to-4.8 percent band is wide for a reason. But the framing is likely to outlive the argument over the exact figure. Once “enabled emissions” enters the vocabulary of regulators and institutional investors, tech companies will find it much harder to present AI as a straightforwardly green technology while quietly selling efficiency to the industry the climate transition is meant to shrink.
Source: Original report. Rewrite for Your News Website.
