Deep learning on climatic data can benefit global ecological tasks, but the application of such methods is currently not accessible to a majority of scientists outside of the deep learning field due to high storage, compute and know-how requirements. To combat this, we present Climplicit, the first spatio-temporal geolocation encoder pretrained to produce implicit climatic representations anywhere on Earth. This model significantly reduces the storage and compute requirements for downstream task learning compared to using the original climatic data together with a fully-trained network. We validate the quality of the Climplicit embeddings for global biomes classification, species distribution modeling, and global plant trait regression. They consistently rank among the top two compared to both training from scratch and alternative, learned geospatial representations, while any other method shows significant variation in performance depending on the task.