Estimating AI's Energy Use and Emissions Impact Explained

Understanding the energy consumption and carbon emissions of artificial intelligence (AI) is a complex but critical task as AI technologies become increasingly pervasive. Unlike traditional appliances or vehicles, AI models lack standardized metrics or regulatory frameworks to measure their energy use, making transparency and accurate estimation difficult. This challenge is compounded by the fact that many leading AI companies operate closed-source models and do not disclose detailed energy data.

MIT Technology Review tackled this issue by focusing on open-source AI models as proxies to estimate energy consumption during two key phases: training, where models learn from data, and inference, where models respond to user queries. While training was initially the primary energy consumer, inference now accounts for the majority of AI’s energy use.

To estimate energy use during inference, researchers measured the electricity drawn by GPUs—specialized chips that perform the bulk of AI computations—and doubled that figure to account for other server components such as CPUs and cooling systems. For text models, the Meta Llama family was analyzed across different sizes, while for image generation, the popular Stable Diffusion 3 model was examined. Video models were assessed using data from the CogVideoX model variants.

A critical aspect of understanding AI’s environmental impact is tracing where the energy powering data centers originates. The carbon intensity of electricity varies widely depending on the regional power grid mix, which includes renewables, natural gas, coal, and nuclear sources. Researchers used data from Electricity Maps, which provides real-time and historical carbon intensity metrics, focusing on major US grid regions that host large data center clusters.

By combining energy consumption estimates with carbon intensity data, the analysis translated AI queries into tangible environmental metrics, such as equivalent miles driven by a gasoline car or hours of microwave use. This approach offers a relatable context for understanding AI’s carbon footprint.

Looking ahead, estimating AI’s total energy demand involves both bottom-up methods—summing energy per query across all uses—and top-down approaches that analyze overall data center energy trends. However, the scarcity of transparent data from AI companies and the absence of AI as a distinct economic sector in emissions reporting complicate accurate forecasting.

This investigation underscores the urgent need for standardized measurement frameworks, increased transparency from AI providers, and integration of sustainability goals into AI development and deployment. As AI’s role expands, understanding and managing its energy and emissions footprint is essential for aligning technological progress with climate objectives.