OlmoEarth Platform Scans an Entire Continent in 30 Hours Using Nearly 1,000 GPUs

Key points

• The Allen Institute for AI (Ai2) launched the OlmoEarth Platform to run large-scale AI analysis on satellite imagery for environmental organisations.
• A single North America wildfire-risk map run used 19,600 CPUs and 994 GPUs in parallel, finishing in roughly 30.5 hours instead of an estimated 4,737 hours.
• The platform processes dozens of terabytes of imagery at a cost of fractions of a penny per square kilometre.
• Governments and NGOs are already using the underlying OlmoEarth models for deforestation monitoring, food security assessments and wildfire risk mapping.
• OlmoEarth was pretrained on roughly 10 terabytes of satellite data and is open for organisations to build on.

Satellites photograph every inch of Earth constantly. The hard part is making sense of all those pictures fast enough to act on them.

The Allen Institute for AI, known as Ai2, has built a piece of infrastructure called the OlmoEarth Platform to do exactly that. It takes its family of OlmoEarth models, which are AI systems trained specifically on roughly 10 terabytes of satellite imagery, and runs them at a scale most organisations could never manage on their own.

What did this system actually accomplish?

To map wildfire risk across all of North America, the platform split the continent into thousands of small patches, analysed each one separately, then stitched the results back into a single map. At peak load, it ran on 19,600 CPUs (general-purpose computer processors) and 994 GPUs (the specialised chips that do the heavy number-crunching AI needs) at the same time. Network traffic between machines hit 168 gigabytes per second. The whole job finished in about 30.5 hours. Without that parallelism (running many tasks at once rather than one after another), the same work would have taken an estimated 4,737 hours, roughly 197 days.

The cost sits at fractions of a penny per square kilometre, which makes continent-wide scans financially realistic for NGOs and government agencies.

Why is satellite AI so much harder than ordinary AI?

Most AI models handle a few megabytes of data and return an answer in under a second. Satellite inference is a different category of problem entirely. A single analysis job can move terabytes of data and run for hours.

The images come from multiple providers, each using different coordinate systems and resolutions. Clouds block optical sensors. Some data arrives minutes after capture; other datasets lag by days. Every prediction must line up precisely on a geographic grid so the final output is an accurate map, not a patchwork.

Downloading and preparing imagery often takes longer than running the AI model itself.

How does the platform handle all of that?

Ai2 splits each job into three stages matched to different hardware. CPUs handle the fetch, alignment and formatting of raw imagery. GPUs run the actual AI analysis. CPUs again handle the final stitching and export into formats like GeoTIFF, a standard file type used in mapping software.

Keeping GPUs busy only on the AI step (rather than wasting them on file downloads) is what makes the economics work.

To avoid flooding satellite data providers with thousands of simultaneous requests, the platform keeps its own internal index of available imagery, updated as new scenes are published. That means requests to external services stay at a steady trickle rather than a sudden surge.

Who is this for, and what does it mean for ordinary people?

Ai2 notes that most environmental organisations, the ones doing deforestation monitoring or tracking food supply risks in drought-prone regions, lack the engineering teams to run this kind of infrastructure themselves. The OlmoEarth Platform is designed to close that gap.

For people living in wildfire-prone areas or regions at risk of food insecurity, faster and cheaper satellite analysis means the organisations watching for those threats can update their maps more frequently and act earlier.

The OlmoEarth models themselves are open, meaning any organisation with a capable technical team can take them and build directly, as first reported by Hugging Face. The platform is the layer built on top for everyone else.

Common questions

Does this mean AI can now predict wildfires?

Not exactly. The platform produces risk maps based on current satellite data, showing which areas have conditions associated with higher fire danger. It is a monitoring and early-warning tool, not a precise forecast.

Is the OlmoEarth model free to use?

The underlying OlmoEarth models are open, meaning organisations can download and adapt them. The OlmoEarth Platform, the managed infrastructure for running them at scale, is a separate product from Ai2 aimed at organisations without large engineering teams.