The OlmoEarth Platform: Geospatial inference at planetary scale

Training a good open-weight model is one thing. Getting an NGO with no ML engineers to actually run it across a continent is another problem entirely, and it's the one AI2 decided to solve with the OlmoEarth Platform, infrastructure built to take their Earth observation models from fine-tuning straight through to planetary-scale inference.

The scale here is genuinely different from typical ML workloads. A chatbot processes a paragraph in under a second; a geospatial inference job can chew through terabytes of multi-band, multi-sensor satellite data over hours, pulling from providers like Sentinel-2, Sentinel-1, Landsat, and NISAR, each with its own projection, resolution, and cloud-cover quirks. AI2 says data acquisition and prep often eat more time than the model's actual forward pass, so the platform splits jobs into three stages matched to the right hardware: CPUs handle fetching and reprojecting imagery, GPUs run inference and dump raw outputs, and CPUs again stitch everything into clean GeoTIFFs or Zarr files. Nobody's wasting expensive GPU cycles waiting on I/O.

The execution layer, called OlmoEarth Run, chops a region into worker-sized partitions and then into smaller windows the model can process independently — a state might become a hundred partitions, a continent several thousand. Because the pieces don't depend on each other, AI2 can throw enormous parallelism at a job. Their recent North America wildfire-risk map used roughly 19,600 CPUs and 994 GPUs simultaneously, pushing over 168 GB/s of network traffic, and turned what would have been 4,737 hours of serial compute into about 30.5 hours — a 155x speedup. That's the kind of number that makes continent-scale monitoring feel almost routine rather than heroic.

A less flashy but arguably more important piece is the metadata problem. Public STAC catalogs from ESA or Microsoft's Planetary Computer aren't built to survive thousands of simultaneous queries from one inference run, so AI2 built its own index, fed by SNS notifications on AWS Open Data and periodic polling elsewhere, so their traffic trickles in with new publications instead of hammering providers all at once. Failures are treated as normal, not exceptional — every task is idempotent and retryable, with automatic fallback to alternate imagery sources when one provider stalls.

What's next matters as much as what's shipped. AI2 wants automated runs triggered by new imagery, change-detection alerts so deforestation or flooding surfaces without a human staring at rasters, and a dedicated embeddings model that could let many tasks skip full inference entirely. They're also aiming for multi-cloud portability, since the platform currently runs on Google Cloud but only needs Docker-capable VMs and blob storage. None of this is subtle marketing about model quality — it's plumbing, and plumbing is usually what decides whether conservation groups and food-security agencies actually get to use any of this.