The OlmoEarth Platform: Geospatial inference at planetary scale

Source: HuggingFace Date: 2026-07-28 URL: https://huggingface.co/blog/allenai/olmoearth-infrastructure

Summary

Allen Institute for AI (Ai2) details the infrastructure behind OlmoEarth, its geospatial foundation-model platform: continent-scale inference in roughly a day at fractions of a penny per square kilometer. A North America wildfire-risk mapping run peaked at ~19,600 CPUs and 994 GPUs with 168+ GB/s network throughput, compressing an estimated 4,737 serial-processing hours into ~30.5 wall-clock hours (155x speedup) across a three-stage CPU/GPU/CPU pipeline processing ~10TB of multimodal satellite imagery. Target users are governments, NGOs, and mission-driven organizations doing deforestation, food-security, and wildfire monitoring.

Implications

This is the physical-AI / Earth-observation lane the landscape has been tracking as a rising side domain (alongside Cosmos-class robotics models) — non-chat, non-coding applications of foundation models at planetary scale, backed by open infrastructure rather than a closed API. It’s a data point for “open frontier widening beyond chat/coding into hardware- and domain-adjacent verticals,” the same pattern noted for NVIDIA’s Cosmos line. Worth a watch-item tie to whether Ai2 open-sources the orchestration layer itself, not just the OlmoEarth model weights.