Our Methodology and Assumptions.

In our companion blog post, we estimate that AI-enabled coding tools are currently responsible for about 250,000 tonnes of carbon emissions per year, increasing by more than five times over the six months between September, 2025, and March, 2026. In this post, we share our methodology and assumptions to invite feedback, questions, and suggestions. We welcome the opportunity to collaborate and drive more focus on the impact of these emissions.

Methodology

We base our estimation on three basic sub-estimates which we unpack here:

(1) We estimate the amount of carbon emissions per code commit where an AI tool was used.

Per-commit carbon footprint‍

We estimate the carbon footprint of a single AI-generated commit by chaining four factors: the tokens generated by the model, the GPU time required to produce them, the node energy that GPU time consumes, and the carbon intensity of the grid powering the data center.

Tokens. We use SWE-rebench’s measured Claude Code usage on SWE-bench Verified: ~3.06M total tokens per task (Jan–Feb 2026 tests). We follow Jegham et al. (2025) and assume ~25% of these are output tokens (including decode and extended thinking modes). Output tokens drive generation time and compute usage. That translates to ~766K output tokens per commit.

GPU time. At Claude Sonnet's measured throughput of 65 output tokens/sec (Artificial Analysis, Mar 2026) plus 1.5 sec time-to-first-token, 766K tokens takes ~3.27 hours of inference time.

Node energy. We model a 4× H100 SXM node (Jegham et al.'s "Medium" class for 40–70B-parameter models). At 0.7 kW per GPU and 70% utilization, GPU power draws 1.96 kW; non-GPU components (CPU, SSD, network) add 0.1 kW, for 2.06 kW total node power. With 4 concurrent requests sharing the node at 90% utilization, that's 0.57 kW per request, or 1.87 kWh per commit before data center overhead.

Site energy. Until very recently, Anthropic primarily used Amazon Web Services (AWS) primarily for Claude’s infrastructure. Applying AWS's reported Power Utilization Effectiveness of 1.14 (Jegham Table 3 and Amazon’s Sustainability reporting) gives 2.14 kWh per commit.

‍Carbon. Multiplying by AWS US-East's location-based carbon intensity factor (0.287 kgCO2e/kWh, Jegham Table 3): 613 gCO2e per commit.

(2) We estimate the number of commits per year where an AI tool was used.

Commits per year

We estimate annual AI-generated commits two ways: a bottoms-up count of agent signatures in public GitHub activity, and a top-down estimate from developer survey data. The two approaches serve as a cross-check on each other, and we take the average for our topline number that is approximately 250,000 tonnes per year.

Bottoms-up: counting commits with AI signatures.

Detection. Following Robbes et al. (2026), we built a signature list for the ~30 most popular coding agents. For example, commits authored by noreply@anthropic.com are very likely Claude Code. We searched GH Archive, a widely used archive of public GitHub activity, for commits with these signatures between April 1, 2025 and April 1, 2026, finding 11.9M matches. (See our methodology appendix for the full signature list.)

Pushes-to-commits adjustment. GH Archive changed its data policy in October 2025, in part in response to growing AI activity, and now records only push counts rather than commit counts. To bridge the post-policy portion of our window, we analyzed September 2025 push data (the last full month before the change) and winsorized to remove extreme outliers like wholesale codebase copies. We found 1.29 commits per push, which we apply to push counts after October 2025.

Extrapolating beyond public GitHub. Public GitHub represents only ~18.5% of GitHub's overall activity (Octoverse 2025); we also need to account for private GitHub plus competitors like GitLab and Bitbucket. Using JetBrains' 2025 State of Developer Ecosystem survey (~25K developers), we treat personal-project usage as a proxy for public-repo market share and work usage as a proxy for overall market share. This gives GitHub an 88% share of public repositories and a 59% share of private. Combined with the public/private split inside GitHub, public GitHub works out to ~11.63% of all repositories worldwide. Scaling our 11.9M observed commits by this factor: ~102M commits with detectable AI signatures across the full market.

Capture rate. Detection is imperfect. Some agents (notably OpenAI's Codex, representing a bit less than 30% of the market according to most recent measurements) don't leave a commit signature by default, most allow authorship to be turned off, and any commit where a developer pushes after the agent's work loses the signature. We estimate we capture roughly 50% of true agent activity, yielding ~204M AI-generated commits per year. This is likely conservative — coding agents typically require paid subscriptions, so private repos (which we can't directly observe) likely have higher agent adoption than public ones.

Conversion to carbon: Using our 613 gCO2e per commit estimate, that equates to just over 125,000 tonnes of carbon annually.

Top-down: from developer population and adoption rates.‍

We separately estimate tops-down by building up from developer survey data.

Population. Per the 2024 American Community Survey, ~4M people in the US hold software-related jobs (including self-employed).

Baseline commit rate. GitClear's analysis of developer activity puts the average developer at ~673 commits per year.

Productivity adjustment. METR's 2026 developer survey reports self-reported productivity gains from AI tools above 100%. We conservatively assume a 50% increase, yielding ~1,010 commits per year for developers using agents.

Adoption. Stack Overflow's 2025 developer survey finds 15% of professional developers use coding agents daily, with substantially more using them weekly or planning to adopt. We use the daily figure as a conservative floor.

Multiplying. 4M developers × 15% adoption × 1,010 commits/year = ~610M AI-generated commits per year in the US alone.

Conversion to carbon. Using our 613 gCO2e per commit estimate, that equates to just under 374,000 tonnes of carbon annually.

(3) We estimate the current growth rate of AI coding tool usage.

Growth rate‍

While our signature-based detection captures only a slice of total agent activity, trends within that slice are informative of broader trends.

Smoothing. Daily commit counts fluctuate with weekends and holidays, so we compare full-month daily averages rather than point-in-time snapshots.

Comparison. In September 2025, we observed an average of 14,400 AI-signed commits per day on public GitHub. By March 2026, that figure had reached 77,300 per day, representing a 5.4× increase over six months.

Annualization. Without acceleration or deceleration, this implies a 25x+ annual growth rate.

Conclusion

We invite you to read our full post where we discuss our conclusions and our plan to continue examining the emissions associated with AI. We invite you to contact us at feedback@cnaught.com to share feedback on our methodology, to let us know if you’d like to collaborate on these efforts, or to suggest any topics for future exploration.

‍

Appendix - Full AI-Coding Tool Signature List: