Allen Institute Launches OlmoEarth Platform for Continent-Scale Satellite Inference

Ai2 (Allen Institute for AI) has released the OlmoEarth Platform, a piece of infrastructure designed to take its OlmoEarth Earth observation foundation models from fine-tuning through large-scale inference. According to Ai2, the platform can process a continent-scale region in roughly a day, handling dozens of terabytes of satellite imagery at a cost of fractions of a penny per square kilometer.

The OlmoEarth models are pretrained on roughly 10 terabytes of multimodal satellite data and are already being adapted by governments and NGOs for deforestation monitoring, food security, and wildfire risk assessment, Ai2 says. The new platform targets a gap the company identified: most environmental organizations lack the engineering teams needed to manage data labeling, fine-tuning, and inference at planetary scale, even when open models are freely available.

Architecture: three stages, three hardware profiles

Ai2 splits each inference job into three stages matched to different hardware:

• Data acquisition and preprocessing (CPU, high I/O) — fetching, reprojecting, aligning, and normalizing imagery into a format optimized for fast loading.
• Inference (GPU) — running the model's forward pass and writing minimally processed outputs to storage.
• Postprocessing (CPU) — stitching per-window outputs into georeferenced maps and exporting to formats such as Zarr, GeoTIFF, or GeoJSON.

This separation keeps GPUs, the most expensive hardware in the pipeline, dedicated to model computation rather than I/O-bound data preparation.

Benchmark: a North America wildfire-risk map

Ai2's execution layer, called OlmoEarth Run, partitions a geographic region into worker-sized chunks, then further subdivides those into windows the model processes independently. A state-sized area typically becomes about 100 partitions; a continent-scale run can span thousands.

According to Ai2, a recent wildfire-risk map covering all of North America used roughly 19,600 CPUs and 994 GPUs in parallel at peak, with network throughput exceeding 168 GB/s. The company claims this reduced an estimated 4,737 hours of serial compute to about 30.5 hours of wall-clock time — a claimed 155x speedup. These figures come from Ai2's own blog post and have not been independently verified.

Data acquisition at scale

A single large inference job can generate thousands of metadata queries simultaneously, more than external catalogs like ESA's or Microsoft's Planetary Computer STAC APIs are built to handle in bursts. To avoid overwhelming those services, Ai2 built its own metadata index that updates continuously — via AWS SNS notifications for datasets on AWS Open Data, and periodic polling for sources without change streams. The index tracks Sentinel-1, Sentinel-2, Landsat, and NISAR imagery and performs windowed reads against cloud-optimized formats (COG, Zarr) to pull only the bytes needed for each partition rather than downloading full scenes.

Ai2 notes several tunable tradeoffs available per job: output resolution (data volume/compute vs. detail), model size (GPU time vs. accuracy), and raw-imagery caching (storage vs. speed on repeated runs). Parallelism itself is bounded by cloud provider quotas.

What this means

The OlmoEarth Platform is Ai2's attempt to close the gap between releasing open Earth observation models and making them usable by organizations without dedicated ML infrastructure teams — a persistent barrier in the environmental and humanitarian sectors that often have mission-critical needs but thin engineering budgets. The reported cost figures (fractions of a penny per square kilometer) and the 155x speedup claim are notable if they hold up under independent scrutiny, since satellite imagery pipelines are typically bottlenecked by data acquisition rather than model inference, exactly the problem this architecture targets. Because pricing for platform access itself was not disclosed in the source material, it remains unclear what this capability will cost end users beyond the raw compute economics Ai2 describes. The broader significance is architectural: separating CPU-bound I/O from GPU-bound inference, and building a custom metadata index to avoid rate-limiting on public STAC catalogs, are patterns likely to recur across any organization attempting geospatial ML at continental scale.