Ai2's OlmoEarth maps continents in a day for fractions of a penny per sq km

Ai2 just pulled back the curtain on the OlmoEarth Platform, and the underlying message is blunt: a great open model is useless if the people who need it can’t run it. The non-profit has been training its OlmoEarth family of geospatial foundation models on a massive 10-terabyte diet of multimodal satellite data. Governments and NGOs are already kicking the tires for deforestation monitoring, food security analysis, and wildfire risk mapping. But as Ai2 points out, most environmental organizations don’t have the engineering muscle to manage the labeling, fine-tuning, and continent-scale inference these models demand.

The platform is designed to bridge exactly that gap. It’s infrastructure that handles the full lifecycle, from fine-tuning to evaluation to the kind of massive inference jobs that would crush a typical setup. We’re talking about processing dozens of terabytes of imagery to map areas the size of a continent in roughly a day. The cost? Ai2 claims fractions of a penny per square kilometer. That’s not a typo.

Getting there meant solving some deeply unsexy distributed computing problems. The raw challenge is that satellite data is a mess. It lives across different providers, uses mismatched projections and resolutions, and is often riddled with clouds. Ai2’s solution splits the work across three hardware profiles to stop expensive GPUs from twiddling their thumbs during data prep. CPUs with high I/O handle the heavy lifting of fetching, reprojecting, and aligning imagery. GPUs are kept perfectly fed for pure inference. Then CPUs stitch the resulting map tiles back together. The platform’s execution layer, called OlmoEarth Run, breaks a single job into potentially thousands of independent windows that can be processed in parallel, reconciling overlaps at the end.

It’s a playbook born from Ai2’s decade of running always-on tools like Skylight and EarthRanger, where downtime isn’t an option. They’re openly sharing these engineering solutions, betting that other teams building large-scale geospatial systems will hit the same walls. The bigger picture here is about access. By turning planetary-scale inference into a commodity, Ai2 is trying to ensure that the organizations best positioned to act on environmental insights aren’t locked out by infrastructure costs.