One bottleneck challenge for most ML practitioners in Earth science is how to accelerate the training/testing of AI models on large scale datasets which contain numerous files and require massive computational power. It takes a ton of efforts to prepare well-represented, unbiased, and comprehensive training datasets, set up parallel computing environments (e.g., GPUs, multiple CPU cores, Google Earth Engine, SageMaker, Kubeflow, Dask, SparkML, MLaaS, etc), automate the model hyperparameter tuning (AutoML), record the provenance in case the experiment is too expensive to rerun, and merge the results into products and disseminate to stakeholders. This session calls for research that tackles any challenge related to the full-stack Earth AI workflows. We hope to bring together the community experiences to learn how to build practical AI on top of parallel computing and big Earth data and assist the Earth modeling efforts to discover new insights and improve our prediction.