Sustainable Open-Source AI Requires Tracking the Cumulative Footprint of Derivatives

Open-source AI is scaling rapidly, and model hubs now host millions of artifacts. Each foundation model can spawn large numbers of fine-tunes, adapters, quantizations, merges, and forks. We take the position that compute efficiency alone is insufficient for sustainability in open-source AI: lower per-run costs can accelerate experimentation and deployment, increasing aggregate environmental footprint unless impacts are measurable and comparable across derivative lineages. However, the energy use, water consumption, and emissions of these derivative lineages are rarely measured or disclosed in a consistent, comparable manner, leaving ecosystem-level impact largely invisible. We argue that sustainable open-source AI requires coordination infrastructure that tracks impacts across model lineages, not only base models. We propose Data and Impact Accounting (DIA), a lightweight, non-restrictive transparency layer that (i) standardizes carbon and water reporting metadata, (ii) integrates low-friction measurement into common training and inference pipelines, and (iii) aggregates reports through public dashboards to summarize cumulative impacts across releases and derivatives. DIA makes derivative costs visible and supports ecosystem-level accountability while preserving openness. https://vectorinstitute.github.io/ai-impact-accounting/

💡 Research Summary

The paper addresses a pressing sustainability challenge in the rapidly expanding open‑source AI ecosystem. Model hubs such as Hugging Face now host millions of artifacts, and a single foundation model can give rise to hundreds or thousands of downstream derivatives—fine‑tuned checkpoints, LoRA adapters, quantized versions, merged forks, distilled variants, and more. While each derivative may consume modest compute, the aggregate resource use can easily surpass the original model’s training footprint, creating a classic “tragedy of the commons” where individual actions collectively strain shared atmospheric and water resources.

The authors first document the “rebound effect” that undermines pure efficiency gains. Techniques like 8‑bit quantization, knowledge distillation, pruning, and mixed‑precision inference dramatically lower per‑query energy, yet they also lower marginal costs, encouraging more experimentation, deployment, and ultimately higher total emissions. Citing International Energy Agency (IEA) forecasts, the paper notes that data‑center electricity demand is projected to double between 2024 and 2030, with AI‑specific servers growing at ~30 % annually. This mirrors the Jevons paradox: efficiency improvements can lead to net increases in resource consumption if not coupled with demand‑side constraints.

Beyond carbon, the work highlights water consumption as a second, often overlooked, externality. Data‑center cooling and upstream electricity generation both draw significant water, and consumption (water not returned to the source) directly impacts local scarcity. The authors illustrate regional water stress in the United States, showing that many high‑density data‑center locations (e.g., Texas, California, Arizona) coincide with water‑stressed basins. By applying a total water‑usage effectiveness (WUE) factor of 1.8–4.0 L/kWh, they translate estimated energy use into megaliters of consumptive water, providing a comparable metric alongside CO₂‑equivalent emissions.

A systematic accounting of emissions for major generative‑AI models (GPT‑3, BLOOM, OPT, Falcon, Llama 2/3/3.1, DeepSeek‑V3, etc.) is presented in Table 1. Where direct disclosures are missing, the authors reconstruct electricity consumption from reported GPU‑hours, average GPU power draw, and Power Usage Effectiveness (PUE). Carbon emissions are then derived using region‑specific grid carbon intensities, and water use is computed via the WUE factor. This methodology reveals that even modest‑size fine‑tunes, when multiplied across thousands of community projects, can dwarf the original model’s carbon and water footprints.

To make these hidden impacts visible, the paper proposes Data and Impact Accounting (DIA), a lightweight, non‑restrictive transparency layer consisting of three pillars:

1. Standardized Impact Cards – Model metadata is extended with a structured “impact card” that records hardware specifications, total kWh consumed, CO₂‑eq, and water‑consumption figures for both training and inference. These cards are attached to every model release, enabling downstream attribution.

2. Automated Tracking Tools – Existing libraries such as CodeCarbon, as well as cloud‑provider APIs, are integrated into common training pipelines (e.g., PyTorch Lightning, Hugging Face Trainer) to capture real‑time emissions with minimal friction.

3. Ecosystem Dashboards – Public dashboards aggregate impact cards across model hubs, visualizing cumulative emissions, water use, and trends over time. The dashboards differentiate between base‑model and derivative contributions, offering the community a shared “environmental ledger” that can inform governance decisions, funding allocations, and policy advocacy.

DIA is deliberately designed to be opt‑in but incentivized: community‑driven certification badges signal responsible reporting, and repositories that adopt DIA gain higher visibility on model hubs. The authors argue that such a coordination mechanism is essential for open‑source AI, which currently lacks any corporate‑level sustainability reporting obligations that constrain closed‑source providers.

In the discussion, the authors stress that efficiency improvements remain necessary but insufficient; without transparent accounting of derivative lineages, the net environmental impact may continue to rise. They suggest extending DIA to capture additional dimensions—regional electricity carbon intensity, cooling technology, renewable‑energy share—and to integrate with emerging regulatory frameworks (e.g., EU AI Act, carbon‑border adjustments).

In conclusion, the paper makes a compelling case that sustainable open‑source AI requires ecosystem‑level impact accounting, not just per‑model efficiency. By standardizing impact metadata, automating measurement, and aggregating results in public dashboards, DIA offers a practical roadmap to render the invisible visible, enabling collective stewardship of AI’s carbon and water footprints.