The Invisible Monster: The Carbon Footprint of AI
summary
Generative AI's training and deployment already consume large amounts of electricity and CO2, with a cited rough estimate of about 1 tonne CO2 per billion model parameters and a projection that widespread Nvidia GPU use could demand 85–134 TWh/year by 2027. The talk uses tools like ElectricityMaps, MLCO2, CodeCarbon, Ecologit and Bloom transparency to quantify impacts (e.g. Bloom needing ~422 GB GPU RAM and ~15,000 kWh/year, equivalent to several transatlantic flights depending on grid mix) and describes a Rebase Green prototype that uses AI to find wasteful code; mitigation strategies shown include distillation, mixture-of-experts (MoE), batching and carbon-aware scheduling.
AI-written summary · about
chapters (10)
- 0:00Rebase Green project
- 1:09IT environmental impact
- 2:55AI generative footprint
- 8:52Training vs inference
- 12:24Carbon per token
- 20:55Measuring and reducing tools
- 23:23Optimizing model training
- 31:33Green AI best practices
- 34:42AI for environment
- 37:59Summary
related talks
14mGreener Intelligence: Reducing AI’s Carbon Footprint
01h 02mJoint Keynote Address: Scaling AI Sustainably: An Uncharted Territory- 01h 02mJoint Keynote Address: Scaling AI Sustainably: An Uncharted Territory
45mGreen AI: Making Machine Learning Environmentally Sustainable
43mGenerative AI and Global Warming: co…
20mSustainability of AI
20mBuilding GenAI That Doesn't Cost the Earth
36mLet’s program to fight the impacts of climate change!