The useful question is what changes for users, developers or buyers, and whether the announcement stays industry context or becomes something people can actually use.

The OlmoEarth Platform, developed by AI2, supports large-scale geospatial inference using foundation models pretrained on approximately 10 terabytes of multimodal satellite data. It enables organizations such as governments and NGOs to apply these models for tasks like deforestation monitoring, food security, and wildfire risk assessment without requiring extensive internal infrastructure or engineering teams.

Satellite imagery inference operates at a vastly different scale than typical machine learning tasks, often involving terabytes of data across multiple spectral bands, sensor types, and geographic regions. The platform addresses these challenges by dividing inference jobs into three hardware-matched stages—data acquisition, preprocessing, and model execution—optimizing resource use and reducing costs to fractions of a penny per square kilometer.

To efficiently locate and fetch satellite data, the platform maintains its own metadata index, updated in real time via notifications or polling, to avoid overwhelming external services like ESA’s or Microsoft Planetary Computer’s STAC APIs. It selects the best imagery sources and retrieves only the necessary data using cloud-optimized formats such as COG or Zarr, enabling windowed reads without full scene downloads.

The platform is designed to handle failures automatically by dynamically provisioning tasks as reentrant and idempotent virtual machine runners. It supports parallel processing across thousands of CPUs and GPUs, as demonstrated by a recent wildfire-risk map of North America generated in 30.5 hours using 19,600 CPUs and 994 GPUs, achieving a 155× speedup over serial compute.