ICML 2026 Position Paper Track — Spotlight (top 5% of submissions)

📄 Paper (arXiv) · 🌐 Project Page · 📦 PyPI

Authors: Shaina Raza, Iuliia Zarubiieva, Ahmed Y. Radwan, Nate Lesperance, Deval Pandya, Sedef Akinli Kocak, Graham W. Taylor

Vector Institute for Artificial Intelligence / University of Guelph

Open-source AI has a hidden cumulative footprint problem. Hugging Face now hosts over 2 million models, and a single foundation model like Llama can spawn hundreds of fine-tunes, LoRA adapters, quantizations, and forks within months of release. Each derivative consumes energy and water that goes largely untracked.

Compute efficiency alone is not enough. Lower per-run costs encourage more experimentation and deployment, which can increase aggregate emissions even as individual runs get cheaper. This is the rebound effect, and it means the open-source ecosystem is collectively exposed to a tragedy of the commons: individually rational actions accumulate into a shared environmental cost that no single actor observes or manages.

• Global data centre electricity is projected to grow from 415 TWh (2024) to 945 TWh by 2030, a 15% annual rate
• AI-specific servers are growing at 30% annually
• Pretraining Llama 3 emitted approximately 2,290 tCO2eq; one documented model family already has 146+ derivatives
• GPT-4 training is estimated at 11,000–18,870 tCO2eq and 76–170 megalitres of water consumed
• One in four data centre assets may face increased water scarcity by 2050

DIA is a lightweight, voluntary transparency layer with three components:

1. Standardized impact schema embedded in model cards: hardware, GPU-hours, estimated kWh, CO2, water use, and model lineage
2. Low-friction instrumentation via existing tools (CodeCarbon, ML CO2 Impact, cloud provider APIs) integrated into training pipelines
3. Ecosystem dashboards that aggregate reported footprints across model families and derivatives over time

DIA is non-regulatory. It does not restrict who can train or release models. The goal is visibility into trends and relative impacts, not auditing individual projects.

DIA ships as the ai-impact-accounting package on PyPI (Python 3.12+) and the dia CLI. A PyPI install is enough for track(), dia, Store, and Hub ingest. Clone this repo for training demos (scripts/), crawl, Space deploy, and the full lab workflow in LAB.md.

pip install ai-impact-accounting

# optional extras
pip install "ai-impact-accounting[measure]"              # CodeCarbon / NVML
pip install "ai-impact-accounting[dashboard]"            # local FastAPI dashboard
pip install "ai-impact-accounting[measure,dashboard]"    # both

On shared HPC login nodes, system pip may point at an older Python or a site wheelhouse. Use a Python 3.12+ venv and uv pip install if bare pip fails with Requires-Python >=3.12.

Requires Python 3.12+. Use this path for scripts/train_*.py, ingest batch jobs, and Space deploy.

git clone https://github.com/VectorInstitute/ai-impact-accounting.git
cd ai-impact-accounting

# recommended — reproducible lockfile
uv sync --all-extras --dev
source .venv/bin/activate

# alternative: pip editable install
# python -m venv .venv && source .venv/bin/activate
# pip install -e ".[measure,dashboard,examples]"

hf auth login   # once; needs a write token for Hub push / ingest

from ai_impact_accounting import track

with track(base_model="distilbert-base-uncased", relation="finetune") as t:
    trainer.train()

t.write("out/README.md")              # local model card + dia_report
t.push("your-org/your-model")         # update the Hub card (needs HF token)

export DIA_REGION=ca-on
export DIA_CI=0.03

After install + hf auth login:

export DIA_REGION=ca-on
export DIA_CI=0.03
export DIA_DATASET=DIA-MVP/dia-state-lab-2026

REPO=your-org/my-bert-sentiment python scripts/train_bert_demo.py

All demos share scripts/dia_finalize.py for Ctrl+C / partial-run handling. See LAB.md for A100 / A40 / CPU recipes, ingest, and batch jobs.

export DIA_DATASET=DIA-MVP/dia-state-lab-2026
export DIA_BASES=distilbert-base-uncased
python scripts/view_local.py

pip install "ai-impact-accounting[dashboard]"
export DIA_DATASET=DIA-MVP/dia-state-lab-2026
export DIA_BASES=distilbert-base-uncased
python -c "
import os
from ai_impact_accounting import Store
from ai_impact_accounting.dashboard.server import serve
serve(Store(os.environ['DIA_DATASET']), default_base=os.environ['DIA_BASES'].split(',')[0])
"

./scripts/run_local.sh your-org/my-bert-sentiment

python scripts/deploy_space.py
# or: python scripts/deploy_space.py --space your-org/dia-dashboard --dataset your-org/dia-state

interrupted = False
with track(base_model="...", relation="finetune") as t:
    try:
        trainer.train()
    except KeyboardInterrupt:
        interrupted = True

trainer.save_model("out/")
t.write("out/README.md")
if not interrupted:
    t.push("your-org/your-model")

dia validate path/to/README.md    # or a Hub repo id
dia report   path/to/README.md

See LAB.md for setup, A100 / A40 / CPU training, ingest, nightly crawl, webhooks, and the web dashboard (local or HF Space).

The docs/ folder contains the companion website (index.html, style.css, main.js) with an interactive model footprint table, carbon vs. water scatter plot, DIA overview, and an embedded read-only dashboard (?embed=1). It is a static site with no build step required and is published as the project page.

This research was funded by the European Union's Horizon Europe research and innovation programme under the AIXPERT project (Grant Agreement No. 101214389), which aims to develop an agentic, multi-layered, GenAI-powered framework for creating explainable, accountable, and transparent AI systems.

Resources used in preparing this research were provided in part by the Province of Ontario, the Government of Canada through CIFAR, and companies sponsoring the Vector Institute (vectorinstitute.ai/#partners).

Graham W. Taylor acknowledges support from the Natural Sciences and Engineering Research Council (NSERC), the Canada Research Chairs program, and the Canadian Institute for Advanced Research (CIFAR) Canada CIFAR AI Chairs program.

@article{raza2026sustainable,
  title={Sustainable Open-Source AI Requires Tracking the Cumulative Footprint of Derivatives},
  author={Raza, Shaina and Zarubiieva, Iuliia and Radwan, Ahmed Y and Lesperance, Nate and Pandya, Deval and Kocak, Sedef Akinli and Taylor, Graham W},
  journal={arXiv preprint arXiv:2601.21632},
  year={2026}
}