Ai2's new platform runs continent-scale satellite AI in a day

Running AI at planetary scale is a wildly different beast than chatting with a chatbot.

While an LLM needs just a few megabytes to generate a paragraph, processing raw Earth observation data across a continent moves dozens of terabytes. In fact, most satellite machine learning pipelines spend far more time finding, downloading, and aligning imagery than actually running models.

To solve this, the Allen Institute for AI (Ai2) just launched the OlmoEarth Platform. It is an end-to-end infrastructure system designed to run continental-scale geospatial inference using its OlmoEarth foundation models, which are pretrained on roughly 10 terabytes of multimodal satellite data.

Turning 4,700 hours into a single day

Ai2 has trained great open models, but they realized environmental NGOs and public agencies rarely have the engineering firepower to handle the data plumbing. Most of these teams just need actionable insights on food security, deforestation, and wildfire risk.

To avoid wasting pricey GPU hours on routine file downloads, OlmoEarth splits every job into three distinct stages. CPU instances fetch, reproject, and normalize imagery; GPUs handle only model forward passes; and secondary CPU workers stitch windows back together into formats like GeoTIFF and Zarr.

In a recent test generating a wildfire-risk map across all of North America, the platform showed what it can do. It carved the continent into thousands of independent windows, throwing 19,600 CPUs and 994 GPUs at the job simultaneously.

The run hit network speeds over 168 GB/s and slashed 4,737 hours of serial processing down to just 30.5 hours. That is a 155x speedup, bringing inference costs down to fractions of a penny per square kilometer.

Smarter reads and fault tolerance

Scaling geospatial workloads also breaks standard APIs. Firing thousands of concurrent queries at public satellite catalogs can easily overwhelm external services.

Ai2 bypassed this by running its own internal metadata index. When new scenes drop on AWS Open Data, an SNS alert updates the index; otherwise, the system polls upstream sources. Instead of downloading full satellite images, workers make ranged reads directly into cloud-optimized GeoTIFFs to grab only the exact pixels required.

Because computing at this scale breaks constantly, every task runner is idempotent and reentrant. If a node stalls or cloud coverage ruins an observation, the platform quietly retries or fails over to alternative providers.

Why it matters

For environmental analysts and researchers, this changes the economics of earth monitoring. Foundation models are useless if running an update over an entire crop belt bankrupts your grant. Cracking continent-wide inference down to about a day for pennies per square kilometer means regular, large-scale monitoring actually becomes viable for underfunded NGOs.

Across the AI landscape, we think Ai2's infrastructure-first approach highlights where real-world robotics and physical-world AI are actually heading. The real moat is rarely the model weights alone; it is the messy data pipeline connecting raw planetary sensors to compute.

That said, the platform is still primarily an orchestrated batch runner. Ai2's next test is shipping fully automated triggers and instant change alerts for floods and deforestation. Until real-time triggers arrive, users are still driving the pipeline rather than letting the planet alert them automatically.