Green AI Practices

• Green AI Practices are defined as AI systems designed to minimize environmental impacts (energy, CO₂, water, materials) across their lifecycle while maintaining performance.
• Lifecycle assessments cover hardware, software, deployment, and end-of-life stages, employing real-time monitoring and energy-aware optimizations.
• Empirical evidence demonstrates up to 98% reductions in energy and CO₂ emissions, validating the cost-saving and scalable potential of sustainable AI systems.

Green AI, as now established in the technical literature, defines the systematic integration of environmental sustainability as an explicit, quantifiable objective throughout the AI lifecycle—from data acquisition and model design to deployment, maintenance, and eventual decommissioning. The central premise prioritizes minimizing energy consumption, carbon and water footprints, and embodied material impacts while maintaining sufficient model utility and performance. The discipline spans not just operational phase energy use, but accounts for hardware, cloud, and lifecycle-level impacts, deploying both engineering and governance strategies to ensure transparency, comparability, and verifiable reductions in environmental burden (Rojahn et al., 10 Nov 2025, Verdecchia et al., 2023, Schwartz et al., 2019).

1. Conceptual Foundations and Scope

Green AI is operationally defined as “a state of best AI-system realization—across algorithms, processes, hardware, and data—within declared lifecycle boundaries, that minimizes cumulative environmental impacts (operational/embodied energy, CO₂, water, materials) and maximizes positive impacts, as evaluated by evidence against published targets and transparent trade-off rules” (Rojahn et al., 10 Nov 2025). This contrasts with “Red AI”, which pursues maximum accuracy without energy or environmental constraint (Sampatsing et al., 12 May 2025, Schwartz et al., 2019).

Key domain distinctions:

• Green AI: Environmental performance of AI systems via quantifiable metrics and lifecycle mapping.
• Sustainable AI: Broader sociotechnical perspective that includes economic and social justice in addition to environmental factors (Rojahn et al., 10 Nov 2025).
• AI-for-Green ("AI4greenAI"): Use of AI algorithms to directly optimize environmental outcomes in other sectors (e.g., renewable integration, cooling optimization) (Clemm et al., 1 May 2024).

2. Lifecycle Phases and System Elements

Lifecycle modeling in Green AI extends from hardware sourcing to end-of-life hardware and model retirement. The five-phase model, mapped directly to Life Cycle Assessment (LCA) stages, embeds environmental targets and Plan-Do-Check-Act (PDCA) governance at every gateway (Rojahn et al., 10 Nov 2025):

1. Green Hardware Selection & Infrastructure Design: Minimize embodied emissions and resource consumption when sourcing CPUs/GPUs, designing server rooms, choosing cooling strategies, and siting data centers. Audit supply chains for environmental product declarations (EPDs).
2. Green AI Development: Integrate low-carbon data acquisition, energy- and carbon-constrained hyperparameter tuning, architecture selection (model-centric Green AI), and transfer/reuse/fine-tuning strategies (Cruz et al., 26 Jun 2024, Verdecchia et al., 2023, Schwartz et al., 2019).
3. Low-Footprint AI Task Realization: Optimize deployment phase via energy-aware model placement (edge/cloud continuum), batch sizing, quantization, runtime resource throttling (DVFS/power-capping), operator fusion, and dynamic scaling (Cruciani et al., 24 Sep 2025, Cruz et al., 2 Jun 2025).
4. Circular AI Maintenance & Administration: Life extension and re-purposing of models and hardware, proactive drift detection and retraining, continual improvement via feedback and empirical impact benchmarking (Santosh et al., 27 Oct 2025, Cruz et al., 26 Jun 2024).
5. Green End-of-Life & Circularity: Decommission models and hardware with recovery of critical materials, closed-loop recycling, and digital product passports for traceability (Rojahn et al., 10 Nov 2025, Clemm et al., 1 May 2024, Ranpara, 28 Mar 2025).

The four-element system thinking framework (model, data, server, cloud) cross-links software and hardware life cycle for an integrated environmental assessment (Clemm et al., 1 May 2024).

3. Measurement, Reporting, and Metrics

3.1 Primary Environmental Metrics

The literature establishes direct, proxy, and composite metrics:

• Energy Consumption: $E = \int_{t_0}^{t_f} P(t)\,\mathrm{d}t$, with P average power in W and E in (J, kWh) (Rojahn et al., 10 Nov 2025, Yarally et al., 2023).
• Carbon Emissions: $C = E\,\times\,EF$, (EF = carbon intensity, kg CO₂e/kWh, spatially and temporally granular) (Agarwal et al., 27 Nov 2025, Schwartz et al., 2019, Verdecchia et al., 2023).
• Water Footprint: $W = E \times WUE$, (WUE = water usage effectiveness, L/kWh) (Goff, 4 Jan 2024, Rojahn et al., 10 Nov 2025).
• Embodied Impact: $$I_{\mathrm{emb}} = \sum_{h}\frac{I_{h}^{\mathrm{EPD}}}{L_h}$$, with $L_h$ hardware lifetime (Rojahn et al., 10 Nov 2025, Clemm et al., 1 May 2024).
• Normalization Per Functional Unit: Results reported per inference, per training run, per million tokens, or per user (Wegmeth et al., 16 Sep 2025, Agarwal et al., 27 Nov 2025).

3.2 Experimental and Reporting Standards

Direct measurement tools (RAPL, nvidia-smi, power meters, PDU/PMC) and estimator-based software (CodeCarbon, Carbontracker, GreenMiner, EMaaS) are deployed for reproducibility; calibrated estimation aligns indirect software to direct hardware readings to minimize error (Rojahn et al., 10 Nov 2025, Verdecchia et al., 2023, Cruz et al., 2 Jun 2025).

Reporting guidelines require:

• full hardware and location specification,
• inclusion of total kWh/CO₂e/Water,
• performance–energy/CO₂ trade-off plots,
• peer-reviewed metadata,
• uncertainty intervals and provenance disclosure.

Model cards and benchmarking suites (e.g., MLPerf with energy columns, Hugging Face cards with CO₂ fields) standardize reporting for public and private deployment (Cruz et al., 26 Jun 2024, Clemm et al., 1 May 2024, Wegmeth et al., 16 Sep 2025).

4. Core Green AI Practices and Optimizations

4.1 Footprint Monitoring

Continuous monitoring of training and inference phases, with reporting of total kWh and CO₂e. Real-time dashboards and location-aware grid carbon lookups allow teams to identify idle resource usage, reduce retraining, and tune scheduling for low-carbon intervals (Verdecchia et al., 2023, Zhao et al., 2023).

4.2 Model and Data Efficiency

• Model Compression: Pruning, quantization, and knowledge distillation lower inference and training FLOPs, reducing both energy and carbon footprints by 2–15×, sometimes with <1% accuracy loss (Schwartz et al., 2019, Clemm et al., 1 May 2024, Verdecchia et al., 2023).
• Energy-Aware Hyperparameter Tuning: Multi-fidelity and energy-regularized Bayesian optimization minimize cumulative retraining energy; early stopping and zero-shot transfer eliminate redundant runs (Yarally et al., 2023, Verdecchia et al., 2023).
• Data-Centric Approaches: Instance selection or “elite sampling” identifies minimal high-impact training subsets, enabling up to 98% energy savings for negligible generalization loss (Alswaitti et al., 19 Feb 2024, Sabella et al., 23 Jul 2025).
• Dynamic Model Selection/Routing: Cascading or routing policies adaptively select the cheapest model per inference meeting accuracy constraints (~25% energy reduction with ~95% of maximal accuracy) (Cruciani et al., 24 Sep 2025, Nijkamp et al., 21 May 2024).

4.3 System and Software Engineering

• Edge/Fog/Cloud orchestration: Intelligent load partitioning reduces central data transfer and localizes compute for optimal energy source and cooling (Cruz et al., 26 Jun 2024, Cruz et al., 2 Jun 2025).
• Energy-Aware Architectures: Client–server fallback, modular plug-and-play components, and dynamic scaling patterns enable rapid adaptation based on workload and budget (Cruz et al., 2 Jun 2025, Ranpara, 28 Mar 2025).
• Circularity and Hardware Lifecycle: Prolong hardware life, choose repairable/disassembly-friendly designs, and archive/re-use models for multiple endpoints (Rojahn et al., 10 Nov 2025, Ranpara, 28 Mar 2025, Clemm et al., 1 May 2024).

4.4 Tooling and Integration

Despite maturity in research tooling, industry lag persists due to lack of integration in ML orchestration and CI/CD systems. A small proportion of Green AI studies publish their measurement or optimization tools; uptake and open contribution is strongly recommended (Verdecchia et al., 2023, Sampatsing et al., 12 May 2025).

5. Governance, Standards, and Socio-Technical Barriers

Explicit governance via PDCA cycles and “phase completion/target” gateways controls progression at each AI lifecycle phase (Rojahn et al., 10 Nov 2025). Regulatory momentum (EU AI Act, CSRD) remains weak: negligible practitioner awareness and operational effect, with voluntary reporting driven mostly at large organizations through broader ESG compliance (Sampatsing et al., 12 May 2025). Key challenges are lack of user-friendly measurement APIs, AI-service-provider transparency, and cross-disciplinary collaboration.

Community recommendations:

1. Integrate Green AI practices from project inception; make sustainability a formal requirement.
2. Publish energy/CO₂ metrics as first-class citizens in research outputs and product dashboards.
3. Push for standardized, provider-agnostic cloud measurement and scope 1–3 supply chain transparency.
4. Structure project and code contribution cycles to minimize redundant retraining/deployment events by batching, reusing, and open-sourcing resources.
5. Drive public awareness and offer incentive structures for compliance with carbon, energy, and reporting targets (Goff, 4 Jan 2024, Zhao et al., 2023, Rojahn et al., 10 Nov 2025).

6. Practical Impact and Field Results

Empirical studies consistently report that by adopting Green AI best practices—including real-time monitoring, energy-aware optimization, right-sized and pruned models, and intelligent scheduling—energy and CO₂ emissions can be reduced by 50–98% for typical tasks, with minimal or fractional loss in utility. For large-scale deployments (LLMs, recommender systems, circular economy workflows), validated reductions in absolute terms range from tens to thousands of tons CO₂e per year (Wegmeth et al., 16 Sep 2025, Ranpara, 28 Mar 2025, Alswaitti et al., 19 Feb 2024, Verdecchia et al., 2023).

Field deployment data echo these gains: dynamic model selection in live ensemble systems cut CPU usage by up to 90% relative to naïve full-ensemble operation with only negligible drop in F1 (Nijkamp et al., 21 May 2024); federated learning configurations reduced carbon emissions by 56–90% while raising accuracy (Sabella et al., 23 Jul 2025); incremental, HITL-guided model adaptation delivered >70% carbon savings during continual learning (Santosh et al., 27 Oct 2025).

7. Future Directions, Limitations, and Open Research

Opportunities include:

• Standardizing emissions reporting APIs and metadata templates.
• Rigorous, sample-level carbon and water attribution.
• Extending AI4greenAI (“self-improving” sustainability) for model and infrastructure adaptation.
• Incorporating embodied hardware, value-chain, and supply-chain emissions into continuous measurement frameworks (Clemm et al., 1 May 2024, Rojahn et al., 10 Nov 2025).
• Regulatory research on the effectiveness and adoption incentives of policy interventions.
• Extension to model families beyond classification, e.g., generative, sequential decision, and large-scale RL.
• Tackling organizational silos and aligning IT–sustainability–policy across sectors (Sampatsing et al., 12 May 2025).

Limitations persist in industry awareness, lack of transparent service-provider metrics, and under-developed toolchains for holistic, provider-agnostic, and regionally-harmonized lifecycle measurement and scheduling.

---

By embedding rigorous lifecycle modeling, direct/standardized measurement, energy- and carbon-aware design at every AI pipeline stage, and strong governance, Green AI offers a technically proven, cost-saving, and globally impactful roadmap for aligning AI innovation with planetary climate and resource goals (Rojahn et al., 10 Nov 2025, Verdecchia et al., 2023, Schwartz et al., 2019, Agarwal et al., 27 Nov 2025, Sabella et al., 23 Jul 2025, Cruz et al., 26 Jun 2024, Sampatsing et al., 12 May 2025).