Ai2 describes the engineering behind the OlmoEarth Platform, infrastructure for running Earth observation foundation models at planetary scale. The platform processes dozens of terabytes of satellite imagery per day at fractions of a penny per square kilometer by splitting jobs into three hardware-matched stages: CPU-based data acquisition, GPU inference, and CPU postprocessing. A recent North America wildfire-risk map used ~19,600 CPUs and ~994 GPUs in parallel, achieving a 155× speedup over serial compute. Key engineering challenges covered include managing satellite metadata across providers via a custom STAC index, windowed reads from cloud-optimized formats (COG, Zarr), idempotent task execution for fault tolerance, and overlap reconciliation for seamless output rasters. The roadmap includes automated scheduled runs, change-detection alerts, agentic interfaces, precomputed global embeddings, and multi-cloud deployment.
