Hugging Carbon Logo
Hugging Carbon AI Sustainability Research @ INSAIT
Methods
Results
Citation

Hugging Carbon: Quantifying the Training Carbon Emissions of AI Models at Scale

Paper 🐙Code & Data 🧮Emissions Calculator
Accepted Accepted by ICML 2026
Highlight Hugging Carbon is selected among ICML 2026 Papers That Matter by Nebius Science

Authors

Xinlei Wang1,2,* Ruibo Ming1 Jing Qiu2 Junhua Zhao3,4 Jinjin Gu1,*
  1. INSAIT, Sofia University “St. Kliment Ohridski”, Bulgaria
  2. University of Sydney, Australia
  3. The Chinese University of Hong Kong, Shenzhen
  4. Shenzhen Institute of Artificial Intelligence and Robotics for Society (AIRS)
* Corresponding authors.
INSAIT University of Sydney CUHK-Shenzhen AIRS

Summary

🌱 AI’s carbon problem is widely acknowledged but rarely measured at scale. Unlike manufacturing or agriculture, the AI industry still lacks standardized accounting methods that extend beyond individual models. Hugging Carbon uses Hugging Face as a public corpus to estimate aggregate training emissions from energy, compute, and model metadata, even when disclosures are incomplete. It estimates nearly 60,000 metric tons of CO₂e across more than 5,000 popular open-source models.

Why it is important
  • Training is one of the most important stages among the AI lifecycle.
  • AI energy consumption and emissions are no longer invisible externalities.
  • Without quantification, no one knows how large the problem is today, or how large it might be in the near future.
Hugging Carbon equivalents overview: ~60,000 tCO₂e across 5,227 models, compared with cars, EU residents, trees, and a cement plant

Low Self-Disclosed Rate: Less than ~0.2% of Hugging Face Repos Directly Disclose Training Emissions

Total Repos
200M Repos
Emissions Disclosed in HF Carbon Emissions Module
2K Repos
Energy or emissions mentioned in READMEs
126
Note: Data as of August 14th 2025; Model Source is Hugging Face.

Methods

Turn heterogeneous Hugging Face metadata into emission estimates with empirical validation.

Three-Tier Strategies

Tier 1 (Direct Calculations)
Rich disclosures (hardware, GPU hours or FLOPs).
Tier 2 (Regressions with Training Flops)
Partial disclosures with estimated training FLOPs.
Tier 3 (Regressions with Parameters)
Minimal information without training details.

Repository Counts by Tier (Downloads > 5,000)

Tier NLP repos CV/MM repos Total repos
Tier 1 390 352 742
Tier 2 944 1,679 2,623
Tier 3 3,053 220 3,273
Note: Tier definitions and estimation methods are in the paper (§3.3). After deduplication, 5,227 unique models are included in the final results.

Training Emissions Formula for Tier 1 Models

Etrain ≈ (PGPU/θGPU) × PUE × Ftotaltrain × Atime × EFregion
PGPU/θGPU: GPU power consumption / efficiency
PUE: Power Usage Effectiveness
Ftotaltrain: Total training FLOPs
Atime: Time amplification factor capturing parallelization inefficiencies, communication overhead, and system-level delays
EFregion: Regional emission factor

AI Training Carbon Intensity Definition

ATCI = Etrain / Ftotaltrain
ATCI
• Measures emissions per compute
• Enables fair comparison across models
• Accounts for regional energy mix
• Standardized metric for AI carbon footprint

Regression Estimation for Tier 2 and Tier 3 Models

When disclosures are incomplete, we estimate training emissions with log–log regressions fit on Tier 1 models.

+0.83%
emissions for every
+1% training FLOPs
(Tier 2)
−56%
H100/H800 vs.
A100/A800
(Tier 2)
+88%
Other accelerators vs.
A100/A800
(Tier 2)

Tier 2 (FLOPs → emissions). Training emissions increase strongly with computational demand: a 1% increase in training FLOPs is associated with ~0.83% higher emissions.

Tier 2 regression (training FLOPs)
log(Etrain) = −39.25 + 0.85 log(EFregion) + 0.83 log(Ftotaltrain) − 0.83 I{H-family} + 0.63 I{Others}

Tier 3 (parameters → emissions). When training FLOPs are missing, we instead regress emissions on parameter count.

Tier 3 regression (model parameters)
log(Etrain) = −32.13 + 0.34 log(EFregion) + 1.45 log(Nparams) − 1{subtype = finetune}
Tier 2: estimated training emissions versus expected FLOPs with regression fits by accelerator family
Tier 2 scatter: emissions vs. training FLOPs

Results

🌍 Accumulative Training Emissions of HF Models Across the World (Downloads 5,000+)
Interactive visualization of AI model distribution and carbon emissions by region
Data as of August 14th 2025 • Model Source: Hugging Face
Key takeaway
US and China cover 73% of training emissions (United States: 23K tCO₂e; China: 20K tCO₂e)

Mean ATCI, Mean Per-Model Emissions, and Totals

Category Average ATCI (tCO₂e/EFLOP) Mean (tCO₂e) Total (10^4 tCO₂e)
Foundation & Individual Models0.14125.6
Finetuned Models0.2380.4
CV & Multi-Modal (MM) Models0.16122.3
NLP Models0.13113.6
Key takeaway
  • Modality matters: Vision/multimodal models have higher ATCI.
  • Lifecycle matters: Finetuned models tend to have higher ATCI.
  • Hardware matters: H100/H800 reduces emissions by ~56% vs. A100/A800.

Estimated Annual Training Emissions of AI Models in Hugging Face

Downloads > 5,000
Periods: 2020-08–2021-07 … 2024-08–2025-07 (x-axis = start year). Data as of August 2025
Key Takeaways
  • Rapid growth: Annual training emissions increased by nearly 100× within five years.
  • NLP surged: NLP became the largest contributor during 2022–2024.
  • CV/MM rebounded: Vision and multimodal models regained the lead in 2024–2025.

Citation

@inproceedings{wang2026huggingcarbon,
  title     = {Hugging Carbon: Quantifying the Training Carbon Emissions of AI Models at Scale},
  author    = {Wang, Xinlei and Ming, Ruibo and Qiu, Jing and Zhao, Junhua and Gu, Jinjin},
  booktitle = {Proceedings of the Forty-Third International Conference on Machine Learning},
  year      = {2026}
}