AI2's OlmoEarth ran a continental wildfire job on 19,600 CPUs and 994 GPUs.

The infrastructure writeup details the platform behind AI2's Earth-observation foundation models, pretrained on roughly 10 TB of multimodal satellite data. The wildfire-risk mapping job compressed an estimated 4,737 hours of serial compute into roughly 30.5 hours at fractions of a penny per square kilometer, pulling Sentinel-1/2, Landsat, and NISAR from AWS Open Data, GCP, USGS, Copernicus, NASA/ASF, and Microsoft Planetary Computer. Tiling, scheduling, and serving strategy for continuously-updating streaming data is rarely documented this openly.