AI model training emissions
Training one large AI model runs tens of thousands of chips for weeks or months. Few labs say what that emitted. Where a lab publishes its own figure, we print it. Where it does not, we estimate a range from the compute it took and say so. How we estimate.
| Model | Lab | Released | Training emissions (t CO2e) | Figure | gasoline cars driven for a year |
|---|---|---|---|---|---|
| gpt-oss-120b | OpenAI | August 5, 2025 | 410–870 t | Our estimate | 96–200 |
| Phi-4-reasoning | Microsoft | April 30, 2025 | 15–160 t | Our estimate | 3–37 |
| Llama 4 Scout | Meta | April 5, 2025 | 1,354 t | Reported by Meta | 320 |
| Llama 4 Maverick | Meta | April 5, 2025 | 645 t | Reported by Meta | 150 |
| Grok 3 | xAI | February 19, 2025 | 29,000–62,000 t | Our estimate | 6,800–14,000 |
A single number is the lab’s own. A range is our estimate. Cars: EPA, Greenhouse Gas Equivalencies Calculator: 4.29 metric tons CO2e per passenger vehicle per year (EPA). Last checked October 17, 2026.
Per prompt
Training happens once. Answering prompts goes on every day. Two labs have said how much electricity one text prompt uses:
- Gemini Apps, median text prompt (Google): 0.24 Wh, about 0.11 g CO2e on the world’s average grid. Google Cloud blog, Measuring the environmental impact of AI inference (August 2025)
- ChatGPT, average query (OpenAI): 0.34 Wh, about 0.16 g CO2e on the world’s average grid. Sam Altman, The Gentle Singularity (June 2025)
World average grid: 458.3 g CO2e per kWh, Ember via Our World in Data, carbon intensity of electricity, World, 2025.
What is not here
We found no training emissions figure from Google, Anthropic or Amazon for a recent model. Nor could we source every input an estimate needs: Anthropic and Amazon do not say which chips or how much compute, and Google does not publish its TPUs' power. We add a model when we can. Company-wide emissions are on the tracker.