Research map / Large language models / AI and society
Holistically Evaluating the Environmental Impact of Creating Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Estimates the holistic environmental impact (embodied and operational carbon, water, development costs, and simulated inference) of developing the OLMo model series (20M to 13B active parameters), reporting 493 tCO2eq and 2,769 kL water.
Problem
Environmental impact of LLM development is poorly reported: most estimates cover only the final training run, ignoring development experiments, hardware manufacturing, water consumption, and inference.
Approach
Measures holistic footprint of building the OLMo series by combining sub-second GPU power logging, PUE/carbon-intensity/WUE factors, amortized embodied hardware impacts, and development runs, plus simulated SGLang inference breakeven analysis.
Key results
- Series of OLMo models released 493 tCO2eq and 2,769 kL of water in total, including development, manufacturing, and final training.
- Model development accounted for roughly 50% of the environmental impact of final training (159 tCO2eq and 843 kL water versus 312 tCO2eq and 1,921 kL for training).
- Power draw during training fluctuated between about 15% and 85% of an H100's maximum, with sharp drops during checkpoint saving.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase model size and training scale | OLMo model training · Final training runs | carbon emissions (tCO$_2$eq): 0.3 -> 101 | Hurt |
| Train 7B and 13B models during development | OLMo model development · Development runs | Carbon Emissions (tCO$_2$eq): 65; 46 | Hurt |
| Include model development in the environmental accounting | OLMo model creation · Development and final training | carbon emissions (tCO$_2$eq): 159; 312 | Hurt |
| Use the MoE model instead of the dense 1B model | OLMo training · Training per trillion tokens | energy use | Hurt |
| Increase request rate from 1 req/s to saturated batching | Inference serving · ShareGPT inference | GPU Power Usage (kWh): 0.160 -> 0.003; 0.386 -> 0.033 | Helped |
Surprising findings
- Model development costs were roughly half the emissions of final training, a portion typically undisclosed by developers.
- Inference breakeven for most models requires hundreds of millions to tens of billions of inferences, but smaller models in some scenarios reach this within weeks to months.
Limitations
- Embodied hardware impacts rely on assumptions because GPU manufacturers do not disclose data.
- Only GPU power is measured, so estimates are lower bounds; inference simulations cover a limited setting.
- Excludes transportation and end-of-life disposal impacts.
Experimental setup
Datasets: ShareGPT
Benchmarks: ShareGPT inference simulation (2400 prompts)
Metrics: GPU energy (kWh) carbon emissions (gCO2eq) water consumption (L) seconds per 100 requests breakeven inference count
Baselines: Llama 2 7B Llama 2 13B Llama 3.1 8B Llama 3.2 1B Gemma 2B & 9B Qwen 2.5 7B OLMo 2 7B OLMo 2 13B
Builds on: Greenhouse Gas Protocol BLOOM environmental report OLMo SGLang
Closest papers by meaning
- LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models (ICLR 2024 · similarity 0.86)
- What can large language models do for sustainable food? (ICML 2025 · similarity 0.84)
- Assessing Large Language Models on Climate Information (ICML 2024 · similarity 0.84)
- AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.83)
- Promoting Sustainable Web Agents: Benchmarking and Estimating Energy Consumption Through Empirical and Theoretical Analysis (AAAI 2026 · similarity 0.83)
- Cost-of-Pass: An Economic Framework for Evaluating Language Models (ICLR 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.