AI’s Environmental Footprint: Insights and Actions

About this resource

This forum explores how ML/AI practitioners can measure and reduce the environmental costs of AI. It pairs two complementary efforts: one that retrieves emissions and cost data from sustainability reports using RAG, and another that benchmarks energy, water, and carbon footprints across large language models.

WattBot: Estimating AI Emissions and Costs with RAG — Chris Endemann 02:24

Chris introduces WattBot, a Kaggle challenge and retrieval-augmented generation (RAG) framework for estimating AI emissions and compute costs. Using 35+ papers and 300+ curated Q&A pairs, teams build systems that return citation-backed answers or explicitly state when evidence is missing—promoting transparency and reproducibility in sustainability reporting.

• Kaggle challenge: kaggle.com/competitions/WattBot2025/overview

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference — Nidhal Jegham 09:07

Nidhal presents a reproducible framework to estimate per-request energy, water, and carbon use for open and proprietary LLMs. The method combines hardware assumptions (A100–H200 GPUs), data center multipliers (PUE, WUE, CIF), and a DEA-style efficiency score that balances model accuracy against environmental cost.

• Preprint: https://arxiv.org/abs/2505.09598
  
• Dashboard: https://app.powerbi.com/view?r=eyJr9

Key points

• Data center efficiency and GPU generation (A100–H200) drive impact as much as model size.
  
• Environmental multipliers like PUE (Power Usage Effectiveness) and WUE (Water Usage Effectiveness) are critical to cross-site comparisons.
• Efficiency is not absolute: the Jevons paradox applies—lower per-query cost can increase overall usage.
  
• U.S. regulation remains minimal, making voluntary transparency efforts (like Mistral’s) especially important.
  
• Renewable energy sourcing and liquid cooling are among the most actionable interventions.
  
• Academic and industry collaborations can close data gaps through open benchmarking.
• Aggregate usage, not single-query cost, drives total environmental footprint.
  
• Reporting environmental impact alongside accuracy metrics is an emerging best practice.

See also

• Notebook: Exploring RAG with Romeo and Juliet: Learn how to build an end-to-end retrieval augmented generation (RAG) pipeline using Shakespeare’s Romeo and Juliet as example text.
• Video Archive: ML+X forum archive: Check out other recorded forums from ML+X.
• Machine Learning Marathon: Learn about the annual Machine Learning Marathon (3-month AI/ML hackathon) hosted by ML+X each fall.