Sustainable AI
Research Intelligence
Explore the latest research on AI energy consumption, carbon footprint, and sustainable computing. Powered by Claude — structured as a research brief, not a chatbot.
Suggested Research Queries
Custom Query
Press Enter to submit · Shift+Enter for new line · Response structured as an academic research brief
Core Literature
Six landmark papers that define the field of AI energy accounting. Every number on this dashboard traces back to one of these sources.
Energy and Policy Considerations for Deep Learning in NLP
Strubell, Ganesh & McCallum · ACL 2019
First major paper to quantify the financial and environmental cost of training state-of-the-art NLP models, revealing costs orders of magnitude larger than previously understood.
Key Stat
626,155 lbs CO₂
to train one NLP model with full NAS
Carbon Emissions and Large Neural Network Training
Patterson, Gonzalez, Le, et al. · arXiv / Google
Google researchers show that GPT-3's training carbon footprint is substantial, but that hardware choice and grid carbon intensity can reduce emissions by up to 100×.
Key Stat
552 t CO₂
to train GPT-3 (could be 100× lower with optimization)
Energy Usage Reports for Neural Network Training
Lottick, Susai, Friedler & Wilson · NeurIPS Climate Workshop
Introduced the concept of Energy Usage Reports (EURs) for ML experiments, providing a framework for tracking and communicating energy costs analogous to financial reporting.
Key Stat
52.4 Wh
average energy per ML experiment measured
Green AI
Schwartz, Dodge, Smith & Etzioni · Communications of the ACM
Coined the 'Red AI vs Green AI' framing and argued that efficiency should be a first-class research metric — not just accuracy — to counter unsustainable growth in AI compute.
Key Stat
10× per 18 months
compute cost growth 2012–2019 (vs Moore's Law)
Electricity 2024 — Data Centres and AI
International Energy Agency · IEA Report
Landmark IEA forecast projecting global data center electricity demand doubling by 2026, driven primarily by AI workloads — with the US and China accounting for most of the growth.
Key Stat
1,000 TWh
projected global AI data center demand by 2026
Measuring the Carbon Intensity of AI in Cloud Instances
Dodge, Prewitt, Tachet des Combes, et al. · FAccT 2022
Showed that the carbon intensity of identical AI workloads varies by up to 40× depending on cloud region and time of day, enabling significant reductions through smarter scheduling.
Key Stat
40× variance
in carbon intensity across cloud regions and time of day
Literature note: This field is moving fast — these six papers represent the methodological foundation, but dozens of follow-on studies have refined the estimates. The IEA (2024) figures are the most current aggregate demand forecasts available. All NUCarbon estimates are derived from these sources applied to Northeastern-specific population and adoption rate assumptions.