The OlmoEarth Platform: Geospatial inference at planetary scale

AI2 Launches OlmoEarth Platform to Operationalize Geospatial Models

The Allen Institute for AI (AI2) has introduced the OlmoEarth Platform, an infrastructure service designed to run planetary-scale inference using its family of Earth observation foundation models. The platform aims to bridge the significant gap between the capabilities of powerful geospatial models and the operational capacity of most environmental organizations. While groups focused on deforestation, food security, and disaster response are best positioned to use these tools, they often lack the specialized engineering teams and infrastructure required for terabyte-scale data processing, model fine-tuning, and large-scale deployment.

The Engineering of Planetary-Scale Inference

The core of the OlmoEarth Platform is its architecture, which is engineered to handle the unique challenges of geospatial data. Unlike typical machine learning tasks, a single Earth observation job can process terabytes of satellite imagery from multiple providers, each with different formats, projections, and resolutions. To manage this complexity cost-effectively, the platform distributes workloads across thousands of processes, as demonstrated in a recent run that mapped wildfire risk for all of North America. That job utilized approximately 19,600 CPUs and 994 GPUs in parallel, reducing an estimated 4,737 hours of serial compute time to just 30.5 hours.

• Three-Stage Processing: Work is divided into distinct stages matched to optimal hardware: CPUs for data acquisition and preprocessing, GPUs for model inference, and CPUs for postprocessing and stitching results.
• Custom Metadata Index: To avoid overwhelming public satellite imagery APIs like those from the ESA or Microsoft Planetary Computer, the platform maintains its own index that is continuously updated, enabling efficient querying for large jobs.
• Fault Tolerance: The system is designed for failure, using reentrant and idempotent tasks that can be safely retried. A monitoring process automatically detects and restarts stalled workers to ensure job completion.

Lowering Barriers for Environmental AI Applications

By providing a managed infrastructure, AI2 is effectively abstracting away the immense engineering complexity of operationalizing foundation models for Earth observation. This allows organizations in conservation and climate science to focus on applying insights rather than managing distributed compute clusters. The platform's roadmap signals a move toward even greater accessibility, with plans for automated model runs, change detection alerts, and agentic interfaces. The development of a dedicated embedding model could further reduce costs and speed up many workloads, broadening the range of viable applications for organizations with limited budgets.

By building the OlmoEarth Platform, the Allen Institute for AI is addressing the critical 'last mile' problem of AI deployment, recognizing that the value of foundation models is often limited by the operational capability of end-users to run them at scale.