The OlmoEarth Platform: Geospatial inference at planetary scale

The OlmoEarth Platform is infrastructure that enables large-scale inference using Earth observation foundation models trained on satellite data, allowing organizations to process geospatial information across continent-sized areas. Most machine learning models process small amounts of data quickly, but satellite inference operates at a vastly different scale, requiring the acquisition and alignment of terabytes of imagery from multiple sources with different projections and resolutions, as well as the ability to stitch results into geographically consistent maps. The platform divides inference jobs into three hardware-matched stages: data acquisition and preprocessing on CPUs, model inference on GPUs, and postprocessing on CPUs, while partitioning geographic regions into independent work units that can be processed in parallel. Governments, NGOs, and environmental organizations are already using OlmoEarth models for applications including deforestation monitoring, food security assessment, and wildfire risk evaluation.