The exponential growth of large-scale Artificial Intelligence (AI) models, or "Red AI" , has led to a 300,000-fold increase in computational demand since 2012 , raising significant environmental and sustainability concerns. While the high carbon cost of model training (e.g., GPT-3's estimated 550 metric tons of CO2e) is well-documented, this focus obscures the dominant environmental burden: model inference, which can account for up to 90% of a model's total lifecycle energy consumption. A critical research gap exists in the unified analysis of carbon cost versus performance metrics across this entire AI lifecycle. Furthermore, the field lacks a standardized, comprehensive framework for Green AI reporting, hampering transparent and verifiable comparisons. This paper addresses this gap through a systematic review and quantitative synthesis of Green AI. We systematically categorize and evaluate three pillars of technical optimization: (1) model compression, (2) hardware-aware AI, and (3) low-power inference techniques. This analysis reveals that high-level architectural choices—such as using general-purpose generative models for discriminative tasks—are orders of magnitude (e.g., 14.6x to 30x) less efficient than task-specific models. We also highlight a "measurement crisis," where common reporting tools like CodeCarbon underestimate true energy consumption by 20-40% compared to ground-truth measurements. We conclude by proposing a comprehensive, lifecycle-based Green AI reporting framework, designed to integrate with existing GHG and ISO standards. This framework mandates unified cost-performance metrics (e.g., CO2e/ inference / performance-unit) to enable transparent, verifiable, and-informed decision-making for sustainable AI development.

Green AI, Sustainable AI, Energy-Efficient AI, Model Compression, Hardware-Aware AI, Carbon Footprint, AI Lifecycle Assessment, LLM

General Purpose AI Uses 20 to 30 Times More Energy than Task-Specific AI - Proof News, https://www.proofnews.org/general-purpose-ai-uses-20-to-30-times-more-energy-than-task-specific-ai/

How AI Uses Energy - Third Way, https://www.thirdway.org/memo/how-ai-uses-energy

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference - arXiv, https://arxiv.org/pdf/2505.09598

Smarter sustainability: How technology can transform climate metrics and disclosure, https://ccli.ubc.ca/smarter-sustainability-how-technology-can-transform-climate-metrics-and-disclosure/

Measuring AI's Energy/Environmental Footprint to Access Impacts, https://fas.org/publication/measuring-and-standardizing-ais-energy-footprint/

Lower Numerical Precision Deep Learning Inference and Training - Intel, https://www.intel.com/content/dam/develop/external/us/en/documents/lower-numerical-precision-deep-learning-jan2018-754765.pdf

PQK: Model Compression via Pruning, Quantization, and Knowledge Distillation - arXiv, https://arxiv.org/abs/2106.14681

A Survey on Neural Network Hardware Accelerators - IEEE Computer Society, https://www.computer.org/csdl/journal/ai/2024/08/10472723/1ViYSMvUFI4

Inference, Low-Cost Models, and Compression - CS@Cornell, https://www.cs.cornell.edu/courses/cs6787/2018fa/Lecture11.pdf

Edge Intelligence: A Review of Deep Neural Network Inference in ..., https://www.mdpi.com/2079-9292/14/12/2495

Towards a Methodology and Framework for AI Sustainability Metrics - HotCarbon, https://hotcarbon.org/assets/2023/pdf/a13-eilam.pdf

he Hidden Cost of an Image: Quantifying the Energy Consumption of AI Image Generation, https://www.researchgate.net/publication/392918101_The_Hidden_Cost_of_an_Image_Quantifying_the_Energy_Consumption_of_AI_Image_Generation

[2506.17016] The Hidden Cost of an Image: Quantifying the Energy Consumption of AI Image Generation - arXiv, https://arxiv.org/abs/2506.17016

The Hidden Cost of an Image: Quantifying the Energy Consumption of AI Image Generation, https://arxiv.org/html/2506.17016v1

Comparative analysis of model compression techniques for ..., https://pubmed.ncbi.nlm.nih.gov/40604122/

Energy-Efficient Transformer Inference: Optimization Strategies for Time Series Classification - arXiv, https://arxiv.org/html/2502.16627v4

Energy-Efficient Transformer Inference: Optimization Strategies for Time Series Classification - arXiv, https://arxiv.org/pdf/2502.16627

Reducing Carbon Footprint of Machine Learning Through Model ..., https://www.ijisrt.com/assets/upload/files/IJISRT25AUG970.pdf

mlco2/codecarbon: Track emissions from Compute and recommend ways to reduce their impact on the environment. - GitHub, https://github.com/mlco2/codecarbon

How to estimate and reduce the carbon footprint of machine learning models, https://towardsdatascience.com/how-to-estimate-and-reduce-the-carbon-footprint-of-machine-learning-models-49f24510880/

Ground-Truthing AI Energy Consumption: Validating CodeCarbon Against External Measurements - arXiv, https://arxiv.org/pdf/2509.22092

Life-Cycle Emissions of AI Hardware: A Cradle-To-Grave Approach and Generational Trends - arXiv, https://arxiv.org/html/2502.01671v1

Sustain AI: A Multi-Modal Deep Learning Framework for Carbon Footprint Reduction in Industrial Manufacturing - MDPI, https://www.mdpi.com/2071-1050/17/9/4134

Criteria for Credible AI-assisted Carbon Footprinting Systems: The Cases of Mapping and Lifecycle Modeling - arXiv, https://arxiv.org/html/2509.00240v1