Product
Solutions
Our Approach
Resources
Company
AI coding agents are a fast-growing source of carbon emissions that won't show up on any cloud sustainability dashboard, run on someone else's infrastructure, and aren't yet standard in any Scope 3 inventory. They belong in yours anyway.
The International Energy Agency (IEA) and cloud providers tell us how much electricity AI infrastructure consumes at the macro level. What they don't tell us is which downstream workloads are driving that growth, or how those workloads should be accounted for in the Scope 3 inventories of the companies using them. AI coding agents are a useful place to start: the activity is measurable, the growth is documented, and the emissions belong somewhere in your Scope 3 inventory whether or not you've found a home for them yet.
Global data center electricity consumption is projected to roughly double, from 415 TWh in 2024 to around 945 TWh by 2030 — slightly more than Japan's current total electricity consumption. Data center electricity demand grew 17% in 2025, with AI-focused data centers growing 50% — well outpacing the 3% growth in global electricity demand overall.
The hyperscalers are seeing emissions climb in step: Microsoft's emissions are up 23.4% since 2020 against a 168% increase in energy use, with the company attributing the increase to AI and cloud expansion. Google's data center electricity demand is up 27% year-over-year, even as the company reduced its data center energy emissions by 12% through efficiency gains and clean energy procurement.
Training emissions for individual frontier models keep rising too: the Stanford AI Index 2026 estimates training emissions for xAI's Grok 4 alone at 72,816 tCO2e, equivalent to driving roughly 17,000 cars for a year.
Unlike chat agents like ChatGPT, Copilot, or Gemini, which output code in a browser window for a human to copy and paste, agents make actual changes to your company's code base just like a developer would. Examples include Claude Code from Anthropic, Copilot SWE Agent from GitHub, and Codex from OpenAI. We track 30+ others in our dataset.
Agents are also only the tip of the iceberg. AI autocomplete tools, such as the original GitHub Copilot, Tabnine, and Cursor's editor mode, are much more pervasive and have their own significant environmental footprint. We focus on agents specifically because their activity is more measurable: many agents sign their own commits, while autocomplete leaves no unique trace because the human is still the one saving the change. Agent usage is also predicted to eclipse autocomplete in the years ahead.
For scale, GitHub reported more than 20 million all-time Copilot users by mid-2025 (a figure that includes both autocomplete and agent users), with adoption at 90% of Fortune 100 companies. Our own pipeline data on agent-attributed commits gives a more focused view of agent activity specifically.
If you've encountered AI coding agents through the tech news cycle, you're likely already familiar with a few of the big ones: Claude Code, GitHub Copilot, and Cursor. These are the ones that get frequent coverage. They're not the whole category.
The distribution is top-heavy: Copilot SWE Agent and Claude Code together account for roughly 85% of the commit volume in our analysis. The remaining 15% is split across 30+ tools, each with their own users, enterprise contracts, and growth trajectories.
AI coding tool adoption also follows a different pattern than most SaaS. Where a company will typically standardize on a single CRM or HR system, AI coding tools get picked up at the team level, and often at the individual developer level. We see this consistently in conversations with customers and engineering peers: a single company's backend team defaults to one tool, the platform team uses another, and individual developers pick whatever fits their workflow.
When tracking AI coding emissions in your inventory, it's tempting to assume that consolidating around a single vendor will reduce your footprint. The public commit data suggests otherwise.
Anthropic's October 2025 web launch of Claude Code grew the total AI coding agent market sharply above its prior trajectory, with Claude Code capturing nearly all of the new demand and competing agents holding steady. A vendor switch within your developer pool doesn't reduce your AI code emissions if the launches grow the category, and they have, repeatedly, over the past year.
For the full analysis of what the launch revealed about market dynamics, see our companion piece Distribution, Not Capability: What Claude Code's Web Launch Revealed About AI Coding Adoption.
Precise per-developer numbers would require granular data — model choice, session length, codebase size, task complexity — and would still vary widely. We worked through an estimate step-by-step in our companion piece, How to Estimate the Carbon Footprint of AI-Written Code: A Methodology Walkthrough. The headline numbers are below.
Applying the Jegham et al. (2025) LLM inference energy methodology to a heavy Claude Code user generating roughly 50 agent-attributed commits per month, we estimate:
This is a conservative measure. We only estimate output token usage, not the input tokens used to create the tasks, and we don't include the very compute-intensive model training.
For a 200-developer organization, this is a small but real new line item today. And the trajectory is what matters. At 5x growth every six months, this is the kind of category sustainability teams will wish they'd started measuring early.
The data covered so far is precise enough to demonstrate that AI coding agent emissions are real, growing, and measurable. The same data is not precise enough to support a per-engineer benchmark in your annual report.
What works instead is a measurement system that captures raw inputs today and applies the best-available estimation methodology at reporting time. Four practical steps:
For engineering teams that want to start measuring AI coding emissions today, Carbonlog is our open-source Claude Code plugin that tracks CO2 and energy consumption per AI coding session. It's based on the same Jegham et al. methodology referenced throughout this piece, and it's the most direct path to producing the raw inputs your sustainability team will need for Scope 3 reporting.
For sustainability teams ready to offset what they measure, CNaught's verified carbon credit portfolios are designed for fast-moving Scope 3 line items like AI code emissions. The full detection pipeline, methodology, and historical data behind this analysis are also open source.
Per GitHub's Octoverse reporting, roughly 80% of developer contribution volume happens in private repos, which this analysis doesn't see. The volume and emission numbers should be read as the observable trajectory; the full-category footprint is plausibly four to five times larger. Autocomplete tools, water use, hardware manufacturing emissions, and model training are also outside the scope of these numbers. The published numbers are a floor, not a ceiling.
Nine days after Claude Code's web launch, Cursor shipped version 2.0. Nine days after that, OpenAI released GPT-5.1. Within the same four-week stretch: Gemini 3 Pro, GPT-5.1-Codex-Max, Claude Opus 4.5. The post-launch period is the single most event-dense stretch in the year of data we have. Our "no cannibalization" finding is about the first two weeks, not the next six months. Full caveats are in the methodology walkthrough.
Commit counts in this analysis come from CNaught's open-source pipeline, which queries the GitHub Search API and attributes commits to AI agents that sign their own author identity. The dataset covers public repositories only and inherits roughly 5% sampling error on large result sets. Energy and carbon estimates apply the Jegham et al. (2025) inference framework. Full methodology, source code, and historical data are available in our open-source repository. For the step-by-step walkthrough, see our companion piece, How to Estimate the Carbon Footprint of AI-Written Code.
