AI Carbon Footprint Assessment

• AI carbon footprint assessment is a method that quantifies GHG emissions from all AI model lifecycle stages, including training, inference, and hardware overheads.
• It integrates system boundaries, quantitative metrics, and regulatory frameworks to provide a comprehensive measure of environmental impact.
• The assessment supports AI sustainability by informing mitigation strategies such as model optimization, renewable-powered infrastructure, and lifecycle reporting.

AI carbon footprint assessment quantifies the greenhouse gas (GHG) emissions—typically expressed as CO₂-equivalent (CO₂e)—associated with all phases of AI model lifecycles, including model training, inference (deployment), and supporting infrastructure. The assessment integrates system-level boundaries, rigorous quantitative metrics, regulatory mandates, and technical levers for mitigation, aiming to operationalize sustainability in environments where AI is a critical function such as financial services, large-scale cloud infrastructure, and manufacturing. Documented variability in methods and system scope, rising regulatory scrutiny, and rapidly evolving hardware and model architectures make the AI carbon footprint a complex, high-priority measurement and risk domain (Tkachenko, 15 Sep 2024, Kim et al., 21 Nov 2025).

1. Definitions, System Boundaries, and Regulatory Context

AI carbon footprint denotes total GHG emissions attributable to all phases of an AI model’s operation, including energy for compute hardware (CPUs, GPUs, TPUs), data center overheads (cooling, power distribution), and data storage/movement. This includes both single-event (training) and ongoing (inference) emissions, measured in CO₂e (Tkachenko, 15 Sep 2024, Schneider et al., 1 Feb 2025).

• System boundaries distinguish between training and inference, compute overhead (Power Usage Effectiveness, PUE), and data-related emissions (storage, transfer).
• Regulatory frameworks (EU AI Act, CSRD, CSDDD, PRA SS1/23) mandate integration of environmental risk into model governance, obliging organizations to identify, assess, mitigate, monitor, and disclose AI-related GHGs at portfolio scale (Tkachenko, 15 Sep 2024).

2. Quantitative Metrics, Core Formulas, and KPIs

Total carbon emissions are calculated using phase-specific hardware power, runtime, and the carbon intensity of the regional grid, scaled by PUE:

$$E_{\rm AI} = \sum_{i\in\{\text{training,\,inference}\}} \left(\sum_{h\in\{\text{CPU,GPU,TPU}\}} P_{h,i} \times t_{h,i} \times \mathrm{CF}_{\rm grid}\right) \times \mathrm{PUE}$$

Where:

• $P_{h,i}$: average power (kW) of hardware $h$ in phase $i$
• $t_{h,i}$: runtime (hours)
• $\mathrm{CF}_{\rm grid}$: grid carbon intensity (kg CO₂e/kWh)
• $\mathrm{PUE}$: Power Usage Effectiveness

With data movement included: $$E_{\rm AI}^{\rm full} = \left(\sum_{h} P_{h} t_{h} \times \mathrm{CF}_{\rm grid} + E_{\rm net} \times \mathrm{CF}_{\rm grid}\right) \times \mathrm{PUE}$$

Key KPIs include:

• Energy per training run (kWh/run)
• CO₂e per inference (g CO₂e/call)
• Total annual AI emissions (t CO₂e/year) (Tkachenko, 15 Sep 2024).
• Compute Carbon Intensity (CCI): $\mathrm{CCI} = \frac{E_{\mathrm{total}}}{F}$ (gCO₂e per FLOP) (Schneider et al., 1 Feb 2025)

Amortized per-inference values:

$$E_{\mathrm{CO₂,amortized}} = \frac{E_{\mathrm{CO₂,training}}}{N_{\mathrm{infer\,requests}}}$$

Life-cycle assessment (LCA):

Full assessments also assign GHGs to hardware manufacturing (embodied emissions), infrastructure construction, and end-of-life stages using ISO 14040/44 and GHG Protocol standards (Schneider et al., 1 Feb 2025, Clemm et al., 1 May 2024).

3. Risk Management Integration and Governance

A structured workflow for banking and regulated sectors integrates AI carbon management into existing risk frameworks:

1. Risk Identification: Catalog models by type and energy profile; align with regulations (EU AI Act Art. 7, CSRD ESRS E1, CSDDD) (Tkachenko, 15 Sep 2024).
2. Risk Assessment & Quantification: Compute emissions per model using the above formulas, perform scenario comparisons (baseline vs. energy-efficient architectures).
3. Mitigation Strategies: Apply model pruning, quantization, knowledge distillation, and adopt energy-efficient model architectures; select renewable-powered cloud vendors and set lifecycle management policies (e.g., scheduled decommissioning).
4. Monitoring & Controls: Embed carbon KPIs in risk dashboards, automate threshold-based alerts, and support continuous compliance reporting (PRA SS1/23, EU AI Act post-deployment) (Tkachenko, 15 Sep 2024).
5. Governance & Reporting: Use dashboards for portfolio tracking, facilitate committee oversight, and map all activities to regulatory reporting lines (CSRD, CSDDD) (Tkachenko, 15 Sep 2024).

4. Model, Hardware, and Infrastructure Efficiency

Energy-efficient AI architectures consistently yield order-of-magnitude reductions in operational emissions:

• Open Mixture-of-Experts (OLMoE): Sparse activation (~10-20% of parameters per token), achieving up to 7× reduction in FLOPs versus standard transformers, with equal or higher accuracy.
• Agentic RAG (Retrieval-Augmented Generation) for Time-Series: Modular sub-agent architecture for inference reduces mean absolute error (MAE) by ~15%, MAPE by ~10%, and halves inference energy versus monolithic models.

Quantitative illustration (per 1M tokens inference, grid CF = 0.5 kgCO₂e/kWh):

Model	GPU-hours/1M tokens	CO₂e/1M tokens
Baseline	50 kWh	25 kg
OLMoE	7 kWh	3.5 kg
RAG	25 kWh	12.5 kg

High-performance GPUs (e.g., NVIDIA A100) reduce wall time and absolute emissions by up to 83% compared to older devices (e.g., T4), even with higher nameplate power draw (Liu et al., 1 Apr 2024). Fine-tuned, compressed, or distilled models routinely achieve 40–60% emission savings per task.

Cloud and Data Center Strategies:

• Select low-PUE, renewable-powered zones for AI workloads (global range: PUE ≈ 1.1–2.0; grid intensity varies by 2–5×) (Tkachenko, 15 Sep 2024).
• Implement dynamic scheduling (e.g., “Follow-the-Sun”: carbon-aware regional migration) for up to 14.6% additional reduction in compute GHGs, with static regional selection reducing emissions up to 75% (Vergallo et al., 20 Mar 2025, Dodge et al., 2022).
• Embodied hardware emissions, though a minority under dirty grids, dominate total footprint in low-carbon grids and must be amortized per functional unit (FLOP, training run) (Schneider et al., 1 Feb 2025).

5. Measurement Tools, Methodologies, and Audit Criteria

Tool types and boundaries:

• Model-based estimation (Green Algorithms, ML CO₂.Impact): User-provided runtimes and hardware—low precision, ignores system overhead.
• Sensor-based telemetry (eco2AI, WattsOnAI, CarbonTracker): Real-time device power monitored (GPU/CPU/RAM), combined with grid carbon factors and PUE. High granularity; system-level reporting (Budennyy et al., 2022, Huang et al., 25 Jun 2025, Kim et al., 21 Nov 2025).
• Hybrid approaches (CodeCarbon, Eco2AI): Combine measurement and fallback (TDP-based) models.
• Standardization remains incomplete; ISO 14067 and GHG Protocol are general, but variation in system boundaries and component inclusion (e.g., network, storage, cooling) is the norm.

Validation and audit:

Systems must benchmark against expert-labeled data; report accuracy, recall, completeness, and uncertainty; and maintain transparent documentation and audit trails (Ulissi et al., 29 Aug 2025). Uncertainty quantification (e.g., variance propagation, repeatability scores) and lifecycle completeness are increasingly required for credible claims. Human-in-the-loop and deviation analyses (reported error ≈10–20% for best automated tools) are best practice (Zhang et al., 22 Jul 2025, Deng et al., 2023).

6. Lifecycle, Environmental Impacts, and Mitigation

Lifecycle monitoring covers both operational and embodied emissions with spatially disaggregated, policy-relevant reporting:

• Operational: Training and inference energy, with per-batch, per-query, or per-FLOP normalization (Tkachenko, 15 Sep 2024, Jegham et al., 14 May 2025).
• Embodied: Hardware manufacturing, data center construction, and decommissioning impacts, amortized by hardware lifetime (Schneider et al., 1 Feb 2025, Clemm et al., 1 May 2024).

Mitigation levers:

• Algorithmic: Pruning, quantization, distillation, minimizing hyperparameter sweeps, maximizing batch size, early stopping (Tkachenko, 15 Sep 2024, Liu et al., 1 Apr 2024).
• Hardware: Prefer the latest accelerators (lower Compute Carbon Intensity), optimized for FLOPs/Watt; extend hardware lifetimes, recycle components (Schneider et al., 1 Feb 2025).
• Cloud scheduling: Region and timing selection for low-carbon grid supply, dynamic pause/resume, and workload migration (“Follow-the-Sun”) (Dodge et al., 2022, Vergallo et al., 20 Mar 2025).
• Lifecycle policies: Decommission stale models, virtualize and consolidate workloads, avoid idle-resource leakage (serverless/container orchestration) (Tkachenko, 15 Sep 2024).

7. Future Directions and Outstanding Challenges

Standardization and transparency are primary research and governance imperatives. Universal measurement protocols should specify system boundaries, measurement intervals, and reporting units (per-task, per-inference, per-FLOP) (Kim et al., 21 Nov 2025).

• Dynamic frameworks: Integrate user behavior, workload patterns, and rebound effects into carbon predictions.
• Lifecycle observability: Spanning edge–cloud–data center, capturing both hardware and operational emissions.
• Multi-criteria sustainability: Extend assessments to water use, land use, and broader environmental metrics; establish benchmarks (e.g., max gCO₂e/1,000 tokens) as industry targets.
• AI for Green AI: Leverage AI to optimize its own efficiency: automated carbon-aware scheduling, energy-efficient code generation, and real-time workload dispatch to lowest-emission regions (Clemm et al., 1 May 2024).

Challenges include definition heterogeneity, incomplete data on proprietary or closed-source models, lack of end-to-end telemetry, and underreporting of embodied emissions. Accurate allocation of shared infrastructure and validation of automated carbon-footprint estimates remain active areas of research and debate (Kim et al., 21 Nov 2025, Ulissi et al., 29 Aug 2025).

---

References:

(Tkachenko, 15 Sep 2024, Kim et al., 21 Nov 2025, Schneider et al., 1 Feb 2025, Chakraborty, 22 May 2024, Budennyy et al., 2022, Jegham et al., 14 May 2025, Vergallo et al., 20 Mar 2025, Ulissi et al., 29 Aug 2025, Clemm et al., 1 May 2024, Deng et al., 2023, Zhang et al., 22 Jul 2025, Huang et al., 25 Jun 2025, Liu et al., 1 Apr 2024, Dodge et al., 2022)