Artificial Intelligence (AI) and Machine Learning (ML) are increasingly consuming energy resources, especially during the training and operational phases of ML algorithms. The concept of “Sustainable AI” aims to reduce or better predict the energy consumption of AI systems, often relying on software measurement tools provided by chip manufacturers (e.g., NVIDIA, Intel). However, initial studies show that these tools often deliver inaccurate estimates, which can lead to counterproductive measures that increase, rather than decrease, resource consumption. This holds true especially for the usage of Graphical Processing Units (GPUs), which, as of today, deliver the biggest contribution in terms of computational power for ML. This work reviews the results of these initial studies and discusses the influence of their findings on real-world ML workloads in industrial applications by conducting measurements on three different close-to-industry use cases. In doing so, it contributes to a better understanding of energy consumption of AI-based industrial applications.