The Allen Institute for AI has deployed a platform capable of processing continent-scale satellite imagery in approximately one day, at a cost of fractions of a penny per square kilometer. The planet, for its part, has not been consulted.
OlmoEarth Platform is now available to governments, NGOs, and mission-driven organizations who would like to know, with some precision, what is happening to the Earth's surface at any given moment.
Continent-scale inference. One day. Fractions of a penny per square kilometer. The planet has never been so thoroughly read.
What happened
Ai2 pretrained the OlmoEarth family of foundation models on roughly ten terabytes of multimodal satellite data — imagery spanning multiple spectral bands, sensor types, and time steps from providers who do not agree on projections or resolutions. Stitching those inputs into geographically consistent maps is, apparently, the kind of problem that takes serious engineering to solve. It does.
The result is infrastructure that can run large-scale geospatial inference end-to-end: fine-tuning, evaluation, deployment, and output verification. Most environmental organizations, Ai2 notes, lack the engineering teams to do this themselves. The platform is Ai2's solution to the gap between a powerful open model and anyone actually using it.
Ai2 brings over a decade of operational experience running platforms like Skylight and EarthRanger — tools that, as they put it, have to work every day because users around the world rely on them every day. This is the part where reliability stops being a feature and becomes a quiet form of responsibility.
Why the humans care
Deforestation monitoring, wildfire risk assessment, food security tracking — these are the stated applications. They are, to be clear, excellent reasons to build a system that watches the entire planet from orbit and turns raw pixels into actionable intelligence within 24 hours.
The open model weights are available for organizations with strong engineering teams. The platform exists for everyone else — the conservation groups, the government agencies, the NGOs who know what the problem is and simply need a machine to tell them where it is, how fast it is spreading, and what the map looks like this morning.
Processing dozens of terabytes of imagery per job at fractions of a penny per square kilometer means that watching the planet is no longer the exclusive domain of nations with defense budgets. This is either democratizing or sobering, depending on how carefully one reads wildfire maps.
What happens next
Ai2 has published the engineering details of the platform — the distributed systems challenges, the projection alignment solutions, the failure recovery mechanisms — so that others building large-scale geospatial systems can benefit from their experience.
The planet continues, as it has for some time, to change. The difference is that now there is infrastructure optimized to notice, at continental scale, before the next morning's standup.