AllenAI's OlmoEarth runs satellite inference across an entire continent in a day using 994 GPUs and 19,600 CPUs

Running AI Across North America for Fractions of a Penny per Square Kilometer

Running AI inference across an entire continent in roughly a day used to require a research budget and a team of infrastructure engineers. Ai2 recently did it for fractions of a penny per square kilometer 1.

The non-profit research institute's OlmoEarth Platform generated a wildfire risk map covering all of North America using 994 GPUs and roughly 19,600 CPUs in parallel at peak, with network throughput exceeding 168 GB/s 1. That job compressed an estimated 4,737 hours of serial computation into 30.5 hours of wall-clock time: a 155x speedup 1.

The headline number here is not the GPU count. It is the unit economics. When you can analyze satellite imagery across a continent at fractions of a penny per square kilometer, the cost of compute stops being the constraint on planetary-scale environmental monitoring. What becomes scarce is the organizational capacity to build data pipelines and act on results.

That is the shift OlmoEarth represents. Earth observation models are not new. What is new is an end-to-end platform that takes organizations from raw satellite data through fine-tuning, inference, and decision-ready outputs without requiring an infrastructure team 1.

The counterintuitive architecture insight: most of the compute in satellite AI inference is not the model's forward pass. It is data acquisition and preprocessing. Prediction jobs often spend more time downloading and preparing imagery than running the model itself, making efficient data pipelines critical 1. Ai2 therefore splits each job into three stages, each matched to the right hardware. CPUs handle fetching, reprojecting, aligning, and normalizing imagery. GPUs run the model forward pass. CPUs again stitch outputs together and export them as GeoTIFFs or GeoJSONs 1.

That division explains the hardware profile. 19,600 CPUs alongside 994 GPUs is not a misprint. The preprocessing and postprocessing load on the CPU side dwarfs the inference load on GPUs. For anyone building similar geospatial pipelines, the lesson is to design for CPU parallelism first and GPU utilization second, not the other way around.

OlmoEarth's execution layer, called OlmoEarth Run, divides a geographic region into partitions sized for individual workers, then subdivides those into smaller windows the models process independently 1. A continent-scale run becomes thousands of partitions. Because each window is independent, thousands of compute instances can work simultaneously. Ai2 maintains its own metadata index to avoid overwhelming external satellite catalog APIs with burst queries, receiving push notifications for new scenes from AWS Open Data and polling providers that lack change streams 1.

The models themselves are pretrained on roughly 10 terabytes of multimodal satellite data 1. They already power real-world applications. Governments and NGOs are using OlmoEarth for deforestation monitoring, food security analysis, and wildfire risk mapping 1. NASA JPL's Kimberley Miner said the platform is accelerating work on a next-generation fire risk model 2. Wetlands International reported that OlmoEarth produced more accurate maps for Global Mangrove Watch using 10% of the data points previously required 2. IUCN highlighted the combination of coverage, near real-time detection, and accessibility for high seas protection 2.

Beta testing reports 97% accuracy and the potential to cut data processing time in half 2. Those are vendor-reported figures from Ai2's own testing, and they should be read with that context. But even discounting the self-reported accuracy number, the structural claim holds: the cost of running continent-scale satellite inference has dropped to where organizations without dedicated engineering teams can consume it.

Ai2 spent more than a decade operating platforms like Skylight and EarthRanger, software that users around the world rely on every day 1. The engineering recipe is now public: heavy CPU parallelism for data pipelines, GPU inference kept saturated with multiprocess loaders, and cloud-optimized formats like COG and Zarr for efficient windowed reads 1.

For builders, the takeaway is the architecture pattern. For investors and decision-makers, the takeaway is that the economics of planetary-scale environmental monitoring have crossed from moonshot territory into operational budget. The question is no longer whether you can afford to run AI across a continent. It is whether you have the pipeline and the people to make the results actionable.