What was disclosed
Ai2 described the production architecture behind the OlmoEarth Platform on July 28, 2026. OlmoEarth models are pretrained on roughly 10 TB of multimodal satellite data and support applications such as deforestation monitoring, food security, and wildfire risk.
A three-stage pipeline
- CPU and high-I/O workers find scenes, reproject and align resolutions, and normalize imagery.
- GPUs focus on model forward passes and stream minimally processed outputs to storage.
- CPU workers stitch windows, apply masks or scaling, and export GeoTIFF, GeoJSON, and other formats.
The design keeps expensive GPUs from waiting on downloads and reprojection. Geographic jobs are split into independent partitions and windows, while reentrant, idempotent tasks support retries when providers fail or imagery is missing.
Reported operating result
Ai2 says a North America wildfire-risk map peaked at roughly 19,600 CPUs, 994 GPUs, and more than 168 GB/s of network throughput. An estimated 4,737 hours of serial compute finished in about 30.5 wall-clock hours, a reported 155× speedup.
Limits
The cost and speed are a platform case study shaped by Ai2's cloud, caching, model, output resolution, and parallelism choices. Provider APIs, quotas, data licenses, regional accuracy, clouds, and missing observations require separate validation. Fast inference does not by itself guarantee sound environmental decisions.