Methodology & Data Sources
NUCarbon uses a bottom-up estimation model: adoption rates × query volume × per-query energy × grid carbon intensity. Every assumption is documented and contestable.
Estimation Model
for each AI tool:
daily_users = total_campus × adoption_rate
daily_queries = daily_users × avg_queries_per_day
daily_kwh = daily_queries × energy_per_query_kwh
daily_co2_kg = daily_kwh × grid_intensity_kg_per_kwh
semester_co2_kg = Σ(daily_co2_kg) × days_since_start
Uncertainty: ±40% due to self-report bias and hardware variability
Key Assumptions
| Parameter | Value Used | Rationale |
|---|---|---|
| Student population | 20,000 | NU Fall 2025 enrollment (public) |
| Faculty & staff | 4,000 | NU HR headcount estimate |
| Queries / person / day | 8 | Avg across all tools, weighted by adoption |
| Base energy (ChatGPT) | 3 Wh | Patterson 2021 + Samsi 2023 midpoint |
| Grid intensity (NE-ISO) | 0.386 kg CO₂/kWh | EPA eGRID 2022 NEWE |
| Semester start | 2026-01-13 | Spring 2026 first day of classes |
Per-Tool Energy Figures
Image generation tools (Midjourney) use diffusion-model inference benchmarks from MLCommons (2023). Audio tools use Whisper-large-v3 A100 measurements.
Literature & Data Sources
Patterson et al. (2021)
Communications of the ACMTraining GPT-3 emitted ~500 tonnes CO₂e; inference per query estimated at ~0.001–0.01 kWh depending on model size.
Strubell, Ganesh & McCallum (2019)
ACL 2019Training a single NLP model can emit as much CO₂ as five cars over their lifetimes. Popularized per-query energy framing.
IEA Data Centres Report (2023)
International Energy AgencyGlobal data center electricity use: 200–250 TWh/yr. AI workloads are fastest-growing segment.
EPA eGRID 2022 — NEWE Subregion
U.S. Environmental Protection AgencyNew England grid emission factor: 0.386 kg CO₂/kWh (includes natural gas, nuclear, hydro, wind mix).
EDUCAUSE AI Horizon Report (2024)
EDUCAUSE65% of students at R1 institutions report weekly ChatGPT use; Copilot adoption in STEM 40–45%.
Samsi et al. (2023) — MLCommons
arXiv / MLCommonsInference benchmarks showing GPT-4 class models at ~0.002–0.005 kWh/query on A100-class hardware.
Limitations & Caveats
- Adoption percentages are extrapolated from peer-institution surveys, not NU-specific measurements.
- Energy figures vary by model version, prompt length, and data center location. Reported values are midpoint estimates.
- This prototype has not been peer reviewed or validated by NU Office of Sustainability.
- Scope 3 attribution methodology is proposed, not settled under any GHG Protocol standard.