Research
Paper: AI Increases Net Global Emissions
A plain-language summary of findings published in npj Climate Action, a Nature Portfolio journal.
Title
AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy–economy model
Headline Result
Conventional assessments of AI’s climate impact focus on datacenter energy demand, AI-optimized renewables, and demand-side efficiency, but largely omit AI’s effect on the economics of fossil fuel supply. This is the first peer-reviewed model to quantify that pathway alongside the others in a single framework. It finds that AI-driven productivity gains increase net global emissions by 0.47 to 1.8 gigatonnes of CO₂ annually (1.2–4.8 percent of 2024 global energy-related emissions), because AI’s fossil fuel applications enable more emissions than its renewables applications avoid, even under parallel adoption. Isolated fossil fuel productivity gains generate emissions 3.3 to 13.3 times larger than today’s datacenter emissions, and up to 8 times larger than projected 2035 datacenter emissions.
Grey bars show datacenter operations, blue bars show emissions avoided through renewables applications, the red bar shows emissions enabled through fossil fuel applications, and the purple bar shows the net effect. Under parallel adoption, the fossil fuel effect is approximately 4 to 5 times the renewables effect.
Key findings
AI-driven productivity gains enable more CO₂ emissions than they avoid – a net increase of 0.47 to 1.8 gigatonnes of CO₂ annually.
Isolated fossil fuel productivity gains generate emissions 3.3 to 13.3 times larger than today’s datacenter emissions, and up to 8 times larger than projected 2035 datacenter emissions.
For net emissions to break even, renewables productivity gains would need to be roughly 4 to 5 times fossil fuel productivity gains.
Without a carbon price, net emissions fall only in scenarios where AI’s fossil fuel productivity gains are zero.
AI-driven grid and demand-side efficiency gains reduce net emissions by only about 0.1 gigatonnes of CO₂ a year, even in the high-adoption scenario.
Carbon pricing narrows the net increase but does not eliminate it, even at $308 per tonne of CO₂ – a level substantially above the emissions-weighted carbon tax rates in effect in any country.
Under parallel adoption, AI-driven productivity increases emissions faster than economic output – the opposite of the absolute decoupling associated with environmentally sustainable growth.
What Are Enabled Emissions?
Enabled emissions are the additional greenhouse gas emissions that result when oil and gas companies use AI products to make fossil fuel production faster, cheaper, and more productive across the value chain, from extraction, to refining, to power generation.
Enabled Emissions vs. Datacenter Emissions
Datacenter emissions and enabled emissions are separate but coupled issues within the same fossil-intensive energy system. Datacenter emissions reflect AI’s direct operational energy use. Enabled emissions come from what AI is used to do: applications that enhance the productivity of fossil fuel supply.
Why It Matters
Current climate, ESG, and AI governance frameworks assess technology companies on their operational emissions, not on how their products are used. A company can reduce its own emissions while its AI products help expand global fossil fuel production. AI is widely promoted as a tool for decarbonization, but under the study’s modeled conditions its productivity gains enable more emissions than they avoid – reinforcing fossil fuel incumbency rather than displacing it.
Methodology
The study uses Purdue University's computable general equilibrium (CGE) model, GTAP-E-Power, representing AI adoption as productivity shocks across fossil fuel extraction, refining, and generation; renewables generation; and grid and demand-side efficiency pathways. Datacenter energy demand is shown for reference only and is not modeled within the equilibrium framework. Shock magnitudes are drawn from IEA analysis and other empirically grounded industry and academic sources, with results tested for robustness across elasticity, baseline, and carbon pricing assumptions.
Two modeling choices likely understate the net increase: renewables gains are set at the upper bounds of documented technical potential, and adoption rates are assumed equal across fossil fuel and renewables applications. The model tracks CO₂ only, and results describe the direction and approximate scale of economic incentives, not time-resolved forecasts.
Enabled Emissions Campaign | holly@enabledemissions.com
Press Highlights
-
The Wall Street Journal
AI’s Biggest Energy Impact Might Be in the Oil Patch, Not the Data Center
-
The Guardian
AI’s potential climate benefits outweighed by role in boosting fossil fuels, study finds
-
Financial Times
AI will boost oil and gas production more than green energy, report finds
-
Axios
AI could help unlock more oil — and emissions
-
Sierra Club
The Hidden AI Carbon Emissions That Aren’t Being Counted
-
Inside Climate News
AI Applications for Oil and Gas Companies Worsen Climate Pollution
-
Wired
AI Could Help Fossil Fuel Companies Create More Emissions
-
Heated
AI’s climate problem is worse than we thought
-
Fast Company
AI’s biggest climate problem may not be data centers
-
Grist
AI could help fossil fuel companies create more emissions
-
Public Citizen
New Peer-Reviewed Study Finds that AI Boosts Fossil Fuels More than Renewables
-
DesignWhine
AI Emissions: New Study Models Up to 1.8 Billion Tonnes of Added CO₂ a Year
-
Climate Home News
“Dangerous consequences” – how AI’s climate framing lets Big Tech off the hook
-
AI Weekly
AI May Add Up to 1.8 Gigatonnes of Emissions a Year, Study Finds
-
HEATED
We asked Big Tech about AI’s fossil fuel problem. The responses were atrocious.
-
Greenpeace
Greenpeace USA responds to report showing that AI’s net impact is a disaster for global emissions
-
ReNews
AI renewables gains face fossil challenge
-
DeSmog
AI Was Supposed to Cut Emissions. Instead, Fossil Fuel Companies Are Using It to Find More Oil and Gas.
Microsoft 2017 Chevron deal vs. Largest AI datacenter
Summary
In 2017, Microsoft signed a seven-year cloud and AI partnership with Chevron, and Microsoft worked with Chevron’s Tengiz operation in Kazakhstan to adopt its technology at a time when the field aimed to raise output from about 600,000 to 1 million barrels a day. Tengiz now produces about 1 million barrels of oil equivalent a day. Burning those roughly 400,000 added barrels a day releases an estimated 51.2 million metric tons of CO₂ a year – about 21 times our estimate for Microsoft’s largest AI datacenter, Fairwater Atlanta, at 2.4 million metric tons.
This comparison doesn’t claim Microsoft caused the whole expansion (see Attribution below). It shows the scale of the fossil fuel production that tech companies’ oil and gas deals support – emissions that appear in no tech company’s climate reporting.
Methodology
Tengiz oil-field expansion: 51.2 million metric tons of CO₂ per year
Approximately 400,000 added barrels of oil equivalent per day × 365 days × 433 kg CO₂ per barrel (EPA) × 81% combustion share. The combustion share is conservative: IEA data show roughly 16% of oil is used for non-energy purposes such as petrochemical feedstock (Endrava, citing IEA Key World Energy Statistics 2016).
Tengiz production rose from approximately 600,000 barrels per day (Logic(s), 2019) to 1 million barrels of oil equivalent per day following completion of the Future Growth Project (Chevron 2025 Form 10-K). Chevron attributes 260,000 barrels per day of crude oil to the project itself, a figure that excludes associated gas; using that figure alone yields approximately 33 million metric tons of CO₂ per year, or 14 times Fairwater Atlanta.
Chevron and Microsoft announced their partnership in October 2017 (World Oil), and Microsoft subsequently worked with Chevron’s Tengiz operation to adopt its technology (Logic(s), 2019).
Microsoft Fairwater Atlanta: 2.4 million metric tons of CO₂ per year
636 MW of IT capacity (Epoch AI) × 1.12 power usage effectiveness (Microsoft) × 8,760 hours × 842.3 lb CO₂ per MWh, the SERC South grid average (EPA eGRID2023). The estimate assumes full-year operation at current capacity and uses location-based emissions; Microsoft’s market-based reporting, which credits renewable energy purchases, would yield a lower figure. Epoch AI ranks the site third worldwide and first among Microsoft facilities by IT capacity as of October 2026 (Epoch AI).
Both figures are Enabled Emissions Campaign estimates, not peer-reviewed results. For global estimates, see Alpine, Geldner, Alpine & Chepeliev, AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy–economy model, npj Climate Action (2026), doi.org/10.1038/s44168-026-00411-0.
Attribution
The 51.2 million metric tons reflects all the added production at Tengiz, not Microsoft’s share of it. The expansion is mainly a physical megaproject – tens of billions of dollars in processing, injection, and pressure-management infrastructure – and Microsoft’s technology was one contributor among many. Its exact contribution can’t be determined from public information, because no accounting standard requires tech companies to measure or disclose the emissions their technology enables.
For scale: if Microsoft’s technology accounted for about 5% of the added production, that would roughly match the annual emissions of Fairwater Atlanta. For cases where the companies attribute production gains to the technology themselves, see Microsoft and ExxonMobil’s 2019 Permian partnership (Microsoft) and our documented examples.