A review finds that energy estimates vary widely and that most AI releases still disclose little about their environmental impact.
The paper examines the environmental costs of developing and using AI models, with a focus on energy efficiency and how those costs are measured. It reports that inference is the main source of energy use over a model’s lifetime, accounting for about 90%, and that estimates can differ by up to 2.4 times depending on which parts of the system are included in the calculation.
The review also identifies limited environmental reporting: more than 84% of AI models released since 2022 provide no information about their environmental impact. It highlights 4-bit quantization, a method that uses less numerical detail in a model, as one approach that could reduce emissions by up to 55% without affecting accuracy in the results discussed by the paper.
Where AI energy goes
The analysis found that inference, or answering requests after an AI model has been developed, accounts for about 90% of the model’s total lifetime energy use. It also found that energy estimates can vary by up to 2.4 times when researchers use different system boundaries—different decisions about which parts of the hardware, software and operation to count.
More than 84% of AI models released since 2022 do not provide information about their environmental impact. The paper reports that 4-bit quantization could reduce emissions by up to 55% without affecting model accuracy, identifying it as one possible Green AI practice.
Why efficiency counts
The findings suggest that improving AI should involve more than making models accurate or capable. Because most lifetime energy use occurs during inference, the efficiency of systems in everyday operation is an important part of their environmental cost.
The paper argues that reporting performance alongside energy efficiency, using consistent measurement methods and disclosing environmental impacts would make comparisons more transparent and improve accountability as AI systems are deployed more widely.
Evidence and caveats
This is an analysis of existing research and benchmark studies, rather than a new direct measurement of every AI model or deployment. The abstract does not specify the individual studies, models or operating conditions included in the analysis.
The reported energy totals are sensitive to system boundaries, with estimates differing by up to 2.4 times. The emissions reduction associated with 4-bit quantization is presented as potential from the reviewed evidence, not as a result that necessarily applies to every model or use case.
// Source
International Journal of Technology & Emerging Research · 2026 · DOI: 10.64823/ijter.2621020
Authors: K S Anaswara, Krishna Mukundan M
Institutions: Little Flower Hospital & Research Centre
