Energy Transparency in Open-Source AI: Training Carbon Footprints and Power Consumption Reporting Standards

Abstract #

Open-source AI model development has transformed machine learning deployment, yet energy transparency remains inconsistent. This article surveys 250 publicly released models from leading repositories, quantifies reporting gaps relative to the EU AI Act’s upcoming sustainability mandates, and proposes a standardized Energy Disclosure Framework (EDF) for future releases. We find that only 12% of models disclose training carbon footprints, while 68% provide insufficient power consumption data. The EDF integrates a three-tiered reporting schema, mandatory metadata fields, and an open-source audit toolkit. By unifying metrics and verification protocols, the framework enables researchers, regulators, and industry stakeholders to compare environmental impacts across models, facilitating informed decision‑making and accountability. Our results underscore the urgency of standardized disclosure and present a actionable pathway toward greener AI ecosystems.

1. Introduction #

The rapid proliferation of open-source AI models has democratized access to state‑of‑the‑art capabilities, but it has also obscured a critical dimension of responsible AI: environmental impact. As AI systems consume increasing amounts of computational resources, their associated carbon emissions and electricity usage have become subjects of intense scrutiny from regulators, investors, and the public. The European Union’s Artificial Intelligence Act (EU AI Act) is set to introduce mandatory sustainability reporting for high‑risk AI systems, including provisions for energy consumption disclosure. However, the current landscape of open-source model publishing lacks standardized mechanisms for conveying these metrics, leaving a fragmented and often incomplete picture of AI’s environmental footprint.

In this context, three research questions guide this study:

RQ1: To what extent do open-source AI model repositories currently report training carbon footprints and power consumption metrics? RQ2: How do existing disclosure practices align or misalign with the forthcoming EU AI Act sustainability requirements? RQ3: What core components constitute a robust Energy Disclosure Framework that can be adopted universally by model developers and publishing platforms?

Answering these questions requires a systematic examination of publicly available model documentation, analysis of reported energy metrics, and synthesis of best‑practice principles for transparent AI energy reporting.

2. Background and Related Work #

The intersection of AI and sustainability has attracted growing attention in recent years. Early studies such as Strubell et al. (2019) highlighted the carbon intensity of training large neural networks, sparking a wave of research into the environmental costs of AI development. Subsequent works, including Patterson et al. (2021) and Liu et al. (2022), quantified the energy expenditure of training state‑of‑the‑art models across multiple domains, establishing benchmarks for carbon equivalence and energy‑efficient architecture design.

More recently, policy‑oriented research has begun to map AI’s environmental externalities onto regulatory frameworks. The EU AI Act’s draft amendments propose a “Sustainability Impact Assessment” for high‑risk AI, mandating disclosure of energy consumption, hardware efficiency, and lifecycle emissions. Parallel initiatives, such as the OECD AI Principles on Green AI and the UN’s Global Pulse on Climate Change, emphasize the need for standardized metrics and transparent reporting. Yet, as highlighted by recent surveys (e.g., Zhang et al., 2024), the implementation of these principles within open-source ecosystems remains ad‑hoc, with no consensus on measurement protocols or verification procedures.

A critical gap persists between empirical measurement capabilities and practical disclosure requirements. While tools like CodeCarbon and Experiment Impact Tracker enable fine‑grained energy monitoring during training, they are seldom integrated into model release pipelines. Moreover, the heterogeneity of computing environments — spanning on‑premises clusters, cloud providers, and heterogeneous hardware — complicates cross‑institution comparisons. This article addresses these challenges by constructing a comprehensive dataset of model release metadata, evaluating current disclosure practices, and proposing a unified framework that reconciles technical measurement with regulatory compliance.

graph LR
    A[Data Collection] --> B[Energy Metric Extraction]
    B --> C[Compliance Scoring]
    C --> D[Framework Specification]
    D --> E[Open‑Source Toolkit]

3. Methodology #

Our research adopts a mixed‑methods approach combining data collection, quantitative analysis, and design synthesis.

3.1 Data Corpus Construction #

We compiled a corpus of 250 open‑source AI models selected from leading repositories (e.g., Hugging Face, GitHub, ModelScope) spanning natural language processing, computer vision, and reinforcement learning domains. Selection criteria included: (i) public availability of model cards, (ii) documented training processes, and (iii) relevance to at least one of the three primary AI subfields. Metadata extraction employed the huggingface_hub Python library to programmatically retrieve model cards, README files, and associated training logs where accessible.

3.2 Energy Metric Extraction #

For each model, we extracted reported carbon footprint values (in kg CO₂e) and power consumption figures (kWh) as advertised in model cards or linked technical reports. In cases where explicit metrics were absent, we applied the CodeCarbon surrogate model to estimate energy usage based on hardware specifications, training duration, and regional electricity intensity factors. All estimates were normalized to a common baseline (e.g., per‑epoch energy consumption) to facilitate comparability.

3.3 Alignment with EU AI Act Requirements #

We mapped the extracted metrics against the EU AI Act’s draft sustainability disclosure checklist, which delineates required fields such as total training energy, inference energy per query, and hardware utilization rates. Alignment was assessed using a binary scoring scheme: 1 for full compliance, 0 for partial or missing information. Scores were aggregated to produce a compliance index ranging from 0 to 100.

3.4 Framework Design Workshop #

The design of the Energy Disclosure Framework (EDF) was informed by a collaborative workshop involving 12 domain experts from academia, industry, and regulatory bodies. The workshop employed a structured design‑thinking process, beginning with empathy mapping of stakeholder needs, followed by iterative prototyping of reporting schemas, and culminating in a consensus‑driven specification of mandatory metadata fields, verification protocols, and an open‑source audit toolkit. The resulting framework adopts a three‑tiered architecture: (i) Core Metrics, (ii) Optional Extensions, and (iii) Audit Mechanisms, each documented in a machine‑readable JSON schema.

4. Findings #

Our empirical analysis reveals stark disparities in energy transparency across the sampled models.

• Reporting Coverage: Only 30 out of 250 models (12%) explicitly disclosed training carbon footprints. An additional 45 models (18%) mentioned energy considerations in vague terms but provided no quantifiable data.
• Metric Completeness: Among the 30 models with disclosed footprints, merely 9 (30% of the disclosed subset) included granular power consumption details such as per‑hour electricity usage or hardware‑specific efficiency metrics.
• EU AI Act Alignment: Compliance scores ranged from 0 to 28 out of a possible 100, with a mean alignment of 14%. The majority of models failed to meet even the most basic disclosure expectations (e.g., total training energy).
• Geographic Distribution: Models originating from European institutions demonstrated a modestly higher alignment (mean score 18) compared to those from North America (mean 12) and Asia (mean 10), suggesting nascent regional awareness but overall insufficient adoption.

These findings underscore a pronounced gap between emerging regulatory expectations and current industry practices, highlighting the necessity for a standardized disclosure mechanism.

5. Discussion #

The observed deficit in energy transparency can be attributed to several interrelated factors. First, the absence of a universally accepted measurement protocol leads to heterogeneous reporting formats, discouraging model developers from publishing detailed energy data. Second, the computational overhead associated with precise energy monitoring — particularly in distributed training environments — may deter smaller research groups from investing in comprehensive reporting. Third, the perceived lack of regulatory incentives for early adopters creates a market equilibrium where sustainability disclosures remain optional.

To bridge this gap, we propose the Energy Disclosure Framework (EDF), comprising the following core components:

1. Mandatory Metadata Schema – A JSON‑based schema that captures training energy (kWh), carbon equivalent (kg CO₂e), hardware specifications, and inference energy per query. The schema is designed to be machine‑readable, enabling automated verification and integration with model registry APIs.
2. Verification Protocol – An open‑source audit tool that cross‑validates reported metrics against independent energy estimation tools (e.g., CodeCarbon) and hardware telemetry APIs. The tool generates a verification badge that can be attached to model cards.
3. Publication Checklist – A set of required fields that must be present in the model card frontmatter, including energy.trainingkwh, energy.trainingco2e, and energy.inferenceperquery. Each field must be accompanied by a citation anchor linking to the source of the metric.
4. Progressive Disclosure Path – An optional extension mechanism that allows developers to incrementally adopt EDF components, starting with core metrics and later adding optional extensions such as lifecycle assessment data.

Implementation of the EDF would require collaboration with platform providers (e.g., Hugging Face, ModelScope) to embed schema validation into their submission pipelines, as well as coordination with regulatory bodies to recognize EDF‑compliant disclosures as compliant with EU AI Act sustainability clauses.

Our survey also revealed that models with higher compliance scores tended to originate from institutions with established sustainability offices, suggesting that organizational culture and dedicated resources play a pivotal role in adopting transparent practices. This insight points to the importance of coupling framework deployment with capacity‑building initiatives, such as workshops and guidelines, to accelerate uptake across the broader research community.

sequenceDiagram
    participant ModelCard
    participant Developer
    participant Platform
    participant Auditor

    ModelCard->>Developer: Publish metadata with energy fields
    Developer->>Platform: Submit model with EDF schema
    Platform->>Auditor: Trigger verification checklist
    Auditor->>Platform: Return compliance badge
    Platform->>User: Display badge on model page

6. Conclusion #

Addressing the energy transparency deficit in open‑source AI requires a coordinated effort that blends technical standardization, regulatory alignment, and community engagement. This article documented a comprehensive survey of current disclosure practices, identified critical gaps relative to forthcoming EU AI Act mandates, and introduced the Energy Disclosure Framework — a pragmatic, open‑source solution that defines mandatory metadata, verification protocols, and a pathway for progressive adoption. By instituting a machine‑readable reporting schema and embedding verification mechanisms directly into model publication pipelines, the EDF empowers developers, regulators, and end‑users to make informed, sustainability‑aware decisions. Future work will focus on piloting the framework across major AI repositories, evaluating its impact on reporting consistency, and extending the audit toolkit to support inference‑phase energy monitoring. The ultimate goal is to foster a culture of environmental accountability that permeates the entire AI lifecycle, from research laboratories to production deployments.

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18]