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

TL;DR

Open-source AI derivatives create a cumulative, hidden environmental footprint that current per-model efficiency metrics cannot reliably reduce. The authors propose Data and Impact Accounting (DIA), a lightweight transparency layer that standardizes carbon and water reporting metadata, adds low-friction instrumentation, and aggregates results via public dashboards to reflect ecosystem-level impact. DIA comprises a minimal footprint schema embedded in model cards, automated measurement tooling, and phased rollout across norms, tooling, and community dashboards, with equations for energy and emissions guidance such as $E_{\mathrm{train}} = \frac{H_{\mathrm{GPU}} \times P_{\mathrm{avg}} \times \mathrm{PUE}}{1000}$ and $C_{\mathrm{train}} = E_{\mathrm{train}} \times \mathrm{CI}$, plus $W_{\mathrm{train}} = E_{\mathrm{train}} \times \mathrm{WUE}_{\mathrm{total}}$. While DIA can support compute budgeting and curb rebound effects, it remains a foundation for voluntary, community-driven governance rather than a regulatory solution. The work emphasizes ecosystem-level coordination as essential to sustainable open-source AI, balancing openness with accountability through standardized measurement and shared dashboards.

Abstract

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/

Figures (2)

• Figure 1: The hidden environmental reality of the AI ecosystem, illustrating localized water stress, disparities in base model footprints, and the cumulative, often invisible, impact of derivative models and the rebound effect. (A) Localized water stress across the United States, according to wri_2023. Circles illustrate the number of data centres per state (https://www.datacentermap.com/usa/), with the top 10 states labelled with their respective counts. Texas (392), California (288), and Arizona (155) are among the states that have both a high number of data centres and high water stress levels. (B) Estimated order-of-magnitude comparison of training-related carbon and water footprints. Closed-model values (e.g., GPT-4) are approximate and based on secondary public estimates rather than audited disclosures iea2025energyAI. Open-model values (e.g., Llama 3) are drawn from official Meta documentation meta2024llama3 when available. Water consumption values for both model types are estimated using reported/inferred energy consumption and average water usage effectiveness (WUE) factors.
• Figure 2: Overview of Data and Impact Accounting (DIA). Top: Base-model training emissions may be reported, but derivative artifacts (e.g., fine-tunes, LoRA adapters, quantizations, merges) are typically untracked, making aggregate ecosystem impact unobservable. Bottom: DIA introduces a low-friction visibility layer with (1) standardized impact reporting in model metadata, (2) automated tracking via existing tools, and (3) ecosystem-level aggregation through public dashboards.