Coding agents are already large emitters, and their staggering growth means we need to pay attention.
Claude Code, Codex, Cursor, Copilot and other AI coding agents are having a moment. According to Stack Overflow’s 2025 annual survey, 51% of professional software engineers were using AI tools daily. And today, in May 2026, public code repos are showing that agents produce greater than five-times more daily code commits than they did six months ago.
What does that mean for carbon emissions? You don’t have to look very far to find concerns and even existential alarmism. The UN says “AI has an environmental problem,” and Greenpeace goes further to assert that the environmental impact may “undermine democracy.” But we couldn’t find any data on the actual emissions attributable to the use of coding agents specifically. So we took a crack at estimating using real data, the best available models, and the best reasonable assumptions that we could make–and transparently disclose.
This was way harder than it should have been. Too much of this information is siloed in private organizations and proprietary tools, and the pace of change is so fast that it is hard to capture. But our conclusion is nonetheless straightforward: a reasonable approach shows that the emissions from coding agents are large, growing astronomically, and deserve our immediate focus.
What we found
- 250,000 metric tonnes today: As we write this, we believe that AI coding agents are responsible for emitting approximately 250,000 tonnes of carbon per year. That’s equivalent to burning more than 278,000,000 pounds of coal or fully charging more than 18,000,000 smart phones.
- 25x annualized growth rate: Our research in public GitHub repos shows that average daily commits by agents in March, 2026 were 5x higher than in September, 2025. That kind of growth annualizes to 25x. While this is astronomical, it doesn’t strike us as out of line: Anthropic, the creator of Claude Code, reported 80x annualized growth in Q1 2026.
- 6,250,000 metric tonnes by this time next year: 250,000 tonnes is meaningful; 6,250,000 tonnes is equivalent to the annual emissions of fast fashion giant Shein and 10% greater than the entire state of Vermont. That is where we will be in a year if the growth we are seeing continues.
How we estimated
We share significantly more detail about our methodology in this companion blog post, but, in brief, we performed four estimations that are driving these findings. Reasonable changes to our assumptions could drive our estimates lower or much higher. We invite feedback and discussion about each of them at feedback@cnaught.com.
- 613 g of CO2e per AI code commit: The Jegham et al. 2025 framework we use is the current best-available public methodology for LLM inference energy estimation. In short, we used it to take AI code commits through best estimates of token usage, energy consumption, and grid emissions factors. We estimate that each commit is approximately 613g of CO2e.
- Tonnage by estimating AI-driven commits per year: We estimated this two separate ways, confirmed they were the same order of magnitude, and took an average.
- Method 1, extrapolating from Github data, 125,000 tonnes of carbon per year: We estimate that there were at least 11.8M code commits by AI coding tools on public GitHub repositories between April 1, 2025 and April 1, 2026. Using reasonable assumptions about the percentage of all code in public GitHub repositories and the percentage of code that lacks a clear commit signature from an AI-coding tool but nonetheless was written using one, we estimate 204 million commits per year and 125,000 tonnes of carbon per year.
- Method 2, based on the number of software engineers, 373,000 tonnes of CO2e per year: Before coding assistants, the average software engineer produced an estimated 673 commits per year, and we conservatively estimate that AI-coding tools 1.5x productivity–meaning 1,010 commits per year. In the United States alone, the Census Bureau’s American Community Survey estimates approximately 4 million software engineers and other professionals who regularly write code as part of their professions. Conservatively ignoring the rest of the world and assuming the 15% adoption of AI tools estimated by Stack Overflow, this gets you to approximately 610 million code commits per year using AI coding tools and approximately 373,000 tonnes of carbon per year.
- Growth Rate in AI Code Commits: Using best available data on Github, we found that the number of daily code commits by AI agents in March, 2026 was more than five times higher than the number six months prior, in September, 2026. That annualizes to a current growth rate of 25x.
What this means
The time to address the climate impacts of AI coding tools is right now. When you start with a substantial base and layer on an off-the-charts growth rate, AI-generated code becomes a meaningful driver of emissions in months, not years.
And AI coding agents are just one part of a larger AI ecosystem that is growing every bit as quickly. That’s why this is the first post in a new series about AI and emissions. We want to shine light on how AI adoption affects emissions so that the broader community of AI users and providers can take action.
We recognize that we are not going to be perfect. Data limitations create unavoidable error bars, and we are unlikely to get everything right working alone. But, given AI’s growth rate, we cannot wait for perfect data to get started, and we seek community input so we can improve together and in real time. Email us at feedback@cnaught.com with your thoughts on what we missed or what we should take on next.
The one thing we know for sure is that adoption of these powerful tools is not slowing down. Climate solutions need to catch up fast.
