AI Emissions and Mitigation Calculator

This calculator estimates the CO₂ emissions of AI model use and the number of trees needed to mitigate that carbon footprint. It combines published research on AI energy use with public electricity grid data and ecological studies.

Different AI models require different amounts of energy per token or per query. Because providers do not publish precise numbers, we used estimates from independent research, industry reports and provider announcements:

• Jegham et al. (2025)– How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.

• Strubell et al. (2019)– quantify the training and inference costs of NLP models.

• Patterson et al. (2021, Google)– examine the emissions of large-scale AI systems.

• Lacoste et al. (2019)– propose a framework for measuring ML emissions.

• OECD AI reports and recent LCA studies (2023–2024)– offered updated ranges for inference efficiency.

• Ren et. al. (2024)– Reconciling the contrasting narratives on the environmental impact of large language models

• Mistral AI (2025)- Our contribution to a global environmental standard for AI

• Anthropic (2024)- Claude 3 announcement

• Google (2023)- Introducing Gemini

• Meta- Meta Llama 3 (2024)

• Meta- Meta Llama 2 (2023)

From these sources, we extracted reported ranges of CO₂ per 1,000 tokens and then took the median value for each model. Estimates vary hugely between sources, especially more recently, so we looked for concurrances and source authority, but this data can be regarded as a "best guess" based on available information.

A query is roughly assumed to be 400 tokens.

Note there is no published data on Gpt-5, Claude-3-opus or Claude-4, so we based their values on inferences.

Electricity emissions depend on where the AI runs. To reflect the differing grid intensities of different regions, we apply regional multipliers based on electricity grid intensity (grams of CO₂ per kWh):

Ember Global Electricity Review 2024, IEA data, and Our World in Data were used.

These values capture the differences in regional power mixes — coal-heavy grids emit more CO₂ per kWh, while renewables-heavy grids emit less.

The user's datacenter region is inferred from the region they select. This is a simplification of the way data routing works - in actuality a user's data could be split up and processed in multiple locations for efficiency. But for the purposes of the calculator we chose the nearest region for illustrative purposes.

We convert emissions into the number of trees required to mitigate that carbon footprint. We assume:

• A mature native tree planted via Trees That Count is estimated to sequester ≈220 kg CO₂ over its 50 year lifespan.

• Accounting for a typical 85% survival rate for plantings, the effective sequestration per planted tree is about 3.74 kg CO₂/yr.

This means one tree is treated as mitigating about 3.74 kg CO₂ per year.

1. User input (queries or tokens) × model emission factor = baseline emissions.

2. Multiply by the regional grid intensity multiplier.

3. Scale to daily or annual totals depending on the user's timeframe.

4. Divide annual emissions by 3.74 kg CO₂/tree to estimate the number of trees needed.

All results are estimates, not exact measurements. They incorporate AI usage, not the carbon costs of building the infrastructure for AI, which are enormous. Values here should be treated as giving you an indicative, order-of-magnitude view to help you plant trees, not an exact measurement.