Ai2 details the infrastructure behind planetary-scale OlmoEarth inference
Ai2 published the engineering behind its OlmoEarth Earth observation models — pretrained on roughly 10TB of Sentinel-1, Sentinel-2, Landsat, and NISAR data — and the numbers are concrete rather than promotional: continent-scale inference in about a day at fractions of a penny per square kilometer. A North America wildfire-risk run peaked at roughly 19,600 CPUs and 994 GPUs in parallel, cutting an estimated 4,737 hours of serial compute to 30.5 hours of wall clock, a 155x speedup. The copyable part is mundane: CPU for data acquisition and reprojection, GPU only for the forward pass, CPU again for stitching, plus a self-maintained STAC metadata index so large jobs never overwhelm external catalogs. No licensing terms were stated.
Source: huggingface.co ↗
Inference at this scale presents its own set of challenges... the platform can run inference across continent-scale areas in roughly a day, processing dozens of terabytes of imagery at a cost of fractions of a penny per square kilometer.
Why this matters
- → Planetary-scale geospatial inference now costs fractions of a penny per km², making climate and disaster monit
- → 155× wall-clock speedup on North America wildfire mapping proves fine-grained hardware allocation (CPU for I/O
- → Open-source OlmoEarth models + reference infrastructure remove engineering gatekeeping, letting NGOs and gover