The Allen Institute for Artificial Intelligence detailed its OlmoEarth Platform on July 28, 2026, outlining a cloud infrastructure designed to run geospatial AI models across continent-sized areas in roughly 24 hours. In technical documentation published on Hugging Face, the organization explained that the system processes dozens of terabytes of satellite imagery at a cost of fractions of a penny per square kilometer.
The platform supports Ai2’s OlmoEarth foundation models, which are pretrained on approximately 10 terabytes of multimodal satellite imagery. Governments and non-governmental organizations currently adapt these models to track deforestation, evaluate food security, and measure wildfire risk. Ai2 noted that while engineering teams can manage open foundation models independently, many environmental groups lack the infrastructure to handle data labeling, fine-tuning, and large-scale computing.
Processing Pipeline
To optimize hardware spending, the platform splits processing into three distinct stages instead of running all tasks on GPUs. Data acquisition and preprocessing run on high-I/O CPU instances to fetch, reproject, align, and normalize imagery. GPUs then execute only the forward pass of the model, after which CPU instances handle postprocessing tasks such as stitching outputs, applying masks, and exporting files into Zarr, GeoTIFF, or GeoJSON formats.
The execution layer, called OlmoEarth Run, divides target geographic areas into partitions assigned to individual compute workers. During a test run generating a wildfire-risk map for all of North America, the system deployed approximately 19,600 CPUs and 994 GPUs simultaneously. Peak network throughput exceeded 168 gigabytes per second, completing the workload in 30.5 hours compared to an estimated 4,737 hours of serial processing.
Data Indexing and Fault Recovery
To avoid overwhelming external public STAC APIs with thousands of concurrent metadata queries, the platform maintains an internal satellite imagery index. The system receives scene updates via Amazon Web Services Simple Notification Service alerts or polls upstream provider indexes every few minutes. It performs windowed reads against cloud-optimized formats, pulling data from Sentinel-1, Sentinel-2, Landsat, and NISAR satellites.
Every task within a partition runs inside a dedicated Docker runner container provisioned dynamically on virtual machines. Because tasks are designed to be reentrant and idempotent, the platform handles worker crashes, missing spectral bands, or cloud cover by automatically restarting failed processes. A separate monitoring system tracks stalled instances and reissues their assignments.
Ai2 currently operates the OlmoEarth Platform on Google Cloud, though the container architecture is built to deploy across multiple clouds or within a partner's private account. Development plans include adding ERA-5 weather data, generating precomputed global embeddings, and creating automated triggers that launch inference runs whenever new satellite imagery becomes available.
