Ai2's OlmoEarth runs satellite AI at continent scale

Ai2 has described the infrastructure behind OlmoEarth, its platform for running foundation models over satellite imagery at very large scale. The models are trained on about 10 terabytes of Earth observation data, and the engineering story is how you actually apply them across a continent without the job falling over. The pipeline runs in three stages: CPUs handle data acquisition and preprocessing, GPUs run inference, and CPUs reassemble the output. Regions are split into machine-sized sections and then into model-sized windows, so thousands of partitions can run independently.

The numbers give a sense of the scale. A wildfire-risk map of North America used roughly 19,600 CPUs and 994 GPUs at once, moving data at more than 168 GB per second. That turned an estimated 4,737 hours of serial computation into about 30.5 hours, a 155 times speedup, at a cost of fractions of a penny per square kilometer. The platform keeps its own metadata index across providers like Sentinel, Landsat, and NISAR, and reads only the windows it needs from cloud-optimized files. Every task is built to be idempotent and reentrant, so a run recovers on its own from missing imagery, cloud cover, or hardware failures.

Ai2 is aiming this at groups working on conservation, food security, and disaster response that rarely have this kind of engineering in-house.

Why it matters

If your team wants to run geospatial models but cannot build a distributed pipeline, an open platform that maps a continent in a day for pennies per square kilometer changes what is realistic on a small budget.