# Power Hungry Processing: Watts Driving the Cost of AI Deployment? **Authors**: Alexandra Sasha Luccioni, Yacine Jernite, Emma Strubell **Year**: 2024 **Source**: arXiv:2311.16863 **Published at**: ACM FAccT 2024 **URL**: https://arxiv.org/abs/2311.16863 ## Summary A systematic measurement of the energy consumption of AI inference across a range of task types, from text classification to image and video generation. Running on professional A100 GPUs, the study found that **image generation** is among the most energy-demanding tasks at roughly **2.9 kWh per 1,000 inferences**, approximately 1,450 times more than a text classification task (0.002 kWh per 1,000). Generating 1,000 images with SDXL was measured at around 1,594 g CO₂ equivalent at average grid intensity. The paper argues that the choice of task and model architecture, not just hardware efficiency, is the primary lever for reducing AI's energy footprint at deployment scale.