How Ai2 Runs OlmoEarth Across Continent-Scale Satellite Data
The platform separates data preparation, model inference and map assembly across CPU and GPU workers built to recover from failures.
Ai2 has published an engineering account of the OlmoEarth platform it uses to run Earth-observation models across very large areas. The July 28 disclosure describes how the system moves from satellite-image preparation through GPU inference to geographically aligned map outputs, while handling the failures expected in distributed jobs.
What the platform does
OlmoEarth models were pretrained on roughly 10 terabytes of multimodal satellite data, according to Ai2. The organization says governments, nonprofits and other mission-driven groups are adapting the models for deforestation monitoring, food security and wildfire risk. Its platform is designed to cover the operational steps around those models: fine-tuning, evaluation and large-scale inference.
The core pipeline separates work by hardware. CPU workers find, download, reproject and normalize imagery. GPU workers run the model’s forward pass. A second CPU stage stitches windows together and exports formats such as Zarr, GeoTIFF and GeoJSON. This division is intended to keep costly accelerators focused on inference rather than image preparation.
The scale Ai2 reports
Ai2 says the platform can process dozens of terabytes across a continent-scale area in roughly a day, at fractions of a penny per square kilometre. For one North America wildfire-risk run, the organization reports peak use of roughly 19,600 CPUs and 994 GPUs, with network throughput above 168 gigabytes per second. It estimates that parallelism reduced 4,737 hours of serial compute to about 30.5 wall-clock hours, a 155-times speedup.
Those measurements come from the platform operator, not an independent audit. They are useful as an architecture case study, but they should not be assumed to describe the cost or throughput another organization would achieve with different data, models, cloud quotas or output resolution.
Reliability and what comes next
Each geographic partition can be processed independently. Ai2 says tasks are reentrant and idempotent, allowing automatic retries, alternate-provider fallbacks and restarts when a runner stalls. The platform also maintains its own metadata index so large jobs do not overwhelm upstream satellite catalogues with bursts of requests.
Automated model runs, change alerts and agentic interfaces remain roadmap items in the post. Readers should distinguish those planned additions from the production pipeline described today.
Status
Learning. Internal confidence is medium because the architecture is documented by Ai2, while the operational scale and cost figures remain first-party measurements.
Sources
Update note: Last reviewed 2026-08-02. We will revise this explainer if Ai2 publishes independent measurements or materially changes the platform architecture.
Sources
- Ai2 engineering article on Hugging Face — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.