The OlmoEarth Platform: Geospatial inference at planetary scale
18:27 · July 28, 2026 · Hugging Face Blog
Summary
The OlmoEarth Platform supplies the production infrastructure needed to move Earth observation foundation models from fine-tuning through evaluation to continent-scale inference. The underlying models are pretrained on roughly 10 terabytes of multimodal satellite imagery and are already being adapted by governments and NGOs for tasks such as deforestation monitoring, food-security assessment, and wildfire-risk mapping.
Inference at this scale differs sharply from typical machine-learning workloads. A single job may ingest dozens of terabytes of multi-band, multi-sensor, multi-temporal data drawn from several providers, each with distinct projections, resolutions, and catalog formats. Data acquisition and preparation often dominate runtime, so the platform splits every job into three hardware-matched stages: CPU-bound ingestion and reprojection, GPU-bound model execution, and output assembly. Multiprocess loaders keep the GPUs saturated while completed tiles stream directly to object storage.
Geographic partitioning enables the required parallelism. Each continent-scale job is divided into hundreds or thousands of independent partitions sized for individual workers; adjacent partitions overlap slightly so that seams can be reconciled during final raster assembly. In one recent North America wildfire-risk run the system used up to 19,600 CPUs and 994 GPUs concurrently, sustaining more than 168 GB/s of network throughput and reducing an estimated 4,737 serial hours to roughly 30.5 wall-clock hours—a 155-fold speedup.
To avoid overwhelming external STAC catalogs, the platform maintains its own metadata index that is refreshed from change notifications or periodic polling. At query time the index returns pointers to cloud-optimized files (COG, Zarr), allowing windowed reads that fetch only the required bytes. Failures are treated as routine: every task is reentrant and idempotent, so the system can retry transient errors, fall back to alternate data sources, or restart stalled workers without manual intervention.
Taken together, these design choices let the platform deliver inference across continent-scale regions in about a day at a marginal cost measured in fractions of a cent per square kilometer, while exposing tunable knobs for resolution, model size, and caching that let operators balance accuracy against budget.
Why it matters
Strong coverage of production-level challenges (distributed setups, hardware optimization, MLOps practices, cost at scale) and metrics for ML engineers handling foundation model inference, with actionable insights on geospatial systems applicable beyond the domain.
