Earth observation satellites collect massive amounts of unlabelled imagery every day, which researchers use to train artificial intelligence (AI) models. These models can help monitor land use, ecosystems, and even estimate population. However, there's currently no clear guideline on where this training data should come from; should it be sampled evenly across the globe, or focused on certain continents or ecosystems, etc? In this study, we tested how the geographic origin of training data affects how well AI models perform on global tasks. We trained several popular Earth observation models using data sampled from different regions and then measured how well they worked on problems like land-use classification, map segmentation, and estimating surface-level data. We found that where the data comes from, its geographic and environmental diversity has a big impact on how useful the models are. Our findings can help researchers build better, more general-purpose models for monitoring the planet.