Ai2's OlmoEarth Platform scales geospatial AI inference to continents

Ai2 has introduced the OlmoEarth Platform, an infrastructure layer built around its OlmoEarth foundation models—pretrained on roughly 10 terabytes of multimodal satellite imagery—that takes Earth observation models from fine-tuning through continent-scale inference. Many environmental organizations lack the engineering resources to manage a full ML lifecycle, and the platform is designed to close that gap.

The architecture splits every job into three stages matched to hardware: CPU-heavy data acquisition and preprocessing, GPU-based inference, and CPU-based postprocessing and stitching. Jobs are partitioned geographically into worker-sized chunks and further divided into windows the model can process independently, allowing massive parallel fan-out across thousands of compute instances.

In a recent North America wildfire-risk map run, the platform used roughly 19,600 CPUs and 994 GPUs in parallel at peak, sustained network throughput above 168 GB/s, and cut an estimated 4,737 hours of serial compute down to about 30.5 hours—a 155x speedup. To avoid overwhelming external STAC catalogs during bursts of metadata queries, the platform maintains its own continuously updated index via SNS notifications or periodic polling, then performs windowed reads against cloud-optimized formats like COG and Zarr.

For engineers building large-scale geospatial or remote-sensing inference systems, the design offers concrete patterns for hybrid CPU/GPU pipelines, partition-based parallelism, and rate-limit-aware metadata indexing.