ARTFEED — Contemporary Art Intelligence

Ai2 Launches OlmoEarth Platform for Planetary-Scale Geospatial Inference

ai-technology · 2026-07-28

The OlmoEarth Platform has been introduced by the Allen Institute for AI (Ai2) as a robust infrastructure for executing geospatial foundation models on a planetary scale. This platform facilitates the transition of models from fine-tuning to extensive inference while tackling issues related to data acquisition, preprocessing, and postprocessing. It organizes tasks into three phases—data acquisition and preprocessing (CPU), inference (GPU), and postprocessing (CPU)—with each phase optimized for specific hardware. Capable of conducting continent-scale inference in about a day, it processes numerous terabytes of imagery at a minimal cost per square kilometer. Utilizing a custom metadata index to streamline external STAC APIs, it also accommodates windowed reads from cloud-optimized formats. The platform automatically provisions virtual machines for tasks and manages failures seamlessly. Ai2's future plans include automated model executions, change detection alerts, agentic tools, enhanced models, diverse modalities, embeddings, and multi-cloud capabilities. Currently, the platform operates on Google Cloud but is built for versatility across various environments.

Key facts

  • OlmoEarth models are pretrained on roughly 10 terabytes of multimodal satellite data.
  • The platform can run inference across continent-scale areas in roughly a day.
  • Processing cost is fractions of a penny per square kilometer.
  • A recent wildfire risk map covering all of North America used 19,600 CPUs and 994 GPUs in parallel.
  • Network throughput exceeded 168 GB/s during that run.
  • The platform divides jobs into three stages: data acquisition/preprocessing (CPU), inference (GPU), postprocessing (CPU).
  • The platform maintains its own metadata index updated via SNS notifications or polling.
  • Ai2 operates existing platforms Skylight and EarthRanger.
  • The platform supports cloud-optimized formats such as COG and Zarr.
  • Future plans include embeddings, change detection, and agentic tools.

Entities

Institutions

  • Allen Institute for AI (Ai2)
  • Skylight
  • EarthRanger
  • ESA
  • Microsoft Planetary Computer
  • Amazon Web Services (AWS)
  • Google Cloud

Locations

  • North America
  • San Francisco

Sources