Overview
I've spent a good amount of time on economic policy modeling using computable general equilibrium modeling. The core concept there is that instead of using traditional macroeconomics tools, you create a network model of microeconomic relationships between different sectors in different regions and either solve for equilibrium conditions using a system of equations or simulate dynamic evolution using differential equations.
I don't have the expertise to independently develop novel models, but I have experience evaluating policy outcomes using these models while leveraging principles from decision-making under uncertainty. That combination matters more than it might sound: CGE results are highly sensitive to elasticity assumptions and baseline projections that nobody can pin down precisely, and reporting a single headline number from one parameterization gives a false impression of precision. Treating those assumptions as an uncertainty space to be explored, rather than inputs to be fixed, produces conclusions that survive contact with the real world.
Recent Work: AI and "Enabled Emissions"
In August 2026, npj Climate Action published AI-driven productivity gains enable more CO2 emissions than they avoid in a global energy–economy model, which I co-authored with Will Alpine, Holly Alpine, and Maksym G. Chepeliev.
The conversation about AI and climate has centered on two things: the direct energy consumption of data centers, and prospective AI-driven boosts to renewable productivity and end-use efficiency. It has largely ignored a third: fossil fuel producers get to use the same technology. We modeled AI as a bidirectional productivity amplifier in a global computable general equilibrium model, quantifying enabled emissions from fossil fuel productivity gains alongside avoided emissions from renewables productivity gains.
What we found:
- Under parallel adoption across pathways, net annual CO2 emissions increase by 0.47–1.8 gigatonnes—1.2–4.8% of 2024 global energy-related CO2 emissions.
- Enabled emissions exceed avoided emissions whenever fossil-sector gains are nonzero. Reaching net emissions reductions requires renewables productivity gains 4–5× greater than fossil fuel gains.
- Enabled emissions (0.6–2.4 Gt CO2 annually) exceed the IEA's 2025 estimate of data center emissions by 3.3–13.3×, and its projected 2035 data center emissions by 1.2–8×.
Absent policy steering, AI's modeled effects increase the carbon intensity of the global economy and reinforce fossil fuel incumbency—outcomes that current analytical and governance frameworks do not fully capture.
How This Paper Happened
Story time! Around a year ago I randomly attended a talk by Holly Alpine and Will Alpine, with the Enabled Emissions Campaign.
They argued that the conversation around AI and climate was focused on a) the direct energy consumption of data centers, and b) prospective AI-driven boosts to renewable productivity and end-use efficiency. But the conversation ignored that c) fossil fuel producers can access the same technology.
I think it was Will who said something like "If you're a researcher with an idea of how to quantify the emissions consequences here, we need your help." So I raised my hand and made the "call me" 🤙 gesture.
I was familiar with the Global Trade Analysis Project (GTAP) at Purdue, which developed a model (the GTAP model) which—to simplify perhaps beyond the point of usefulness—strings together many microeconomic relationships to construct a model of global trade. The GTAP model doesn't have the capability to measure what impact AI will have on productivity in renewables or the fossil fuel sector, but you can set specific productivity changes as inputs and examine how prices and consumption change while accounting for downstream impacts and complex market interactions. So it seemed obvious how we could attempt to quantify "enabled emissions." I'd taken a similar approach with a related model that focused more on agriculture to examine climate and environmental impacts of rolling back ethanol fuel mandates in 2023.
Issue: The publicly available version of GTAP didn't sufficiently disaggregate fuels and power generation technologies, the more detailed version was inaccessible to me, and even if it wasn't I don't have deep enough expertise in computable equilibrium modeling to generate a more appropriate aggregation. And for that matter, I didn't have deep enough computable equilibrium modeling experience to confidently publish on it without a specialist on the team. Don't get me wrong—I've got strong general proficiency in quantitative modeling for policy and decision analysis under uncertainty to the point that I can productively use those models, but computable equilibrium modeling is a very deep pool that I've basically only dangled my legs into.
Resolution: Ask a specialist. Maksym Chepeliev joined to help with the model aggregation, direct our sensitivity analysis, and broadly ensure the work and reporting were rigorous.
In a lot of ways, I was in my element with this one. Finding passionate people in the world with deep expertise in their area of interest and a problem to solve, helping them come up with a scientifically rigorous way of addressing that problem while leveraging their subject matter expertise, and when I didn't have the necessary chops in a highly complex and deep sub-speciality, I connected with someone who does.
Further Reading
- Alpine, W., Geldner, N., Alpine, H., & Chepeliev, M. G. (2026). AI-driven productivity gains enable more CO2 emissions than they avoid in a global energy–economy model. npj Climate Action, 5, 71. [PDF]
- Press release summarizing the study's findings.
- An earlier example of policy evaluation using these methods. Energy Policy, 2023.