Allen Institute for AI launches OlmoEarth Platform for large-scale geospatial inference

The Allen Institute for AI (Ai2) released the OlmoEarth Platform, infrastructure designed to operationalize geospatial inference at planetary scale. The platform supports the full lifecycle from fine-tuning and evaluation to large-scale inference for Earth observation foundation models. Ai2 states these models are pretrained on roughly 10 terabytes of multimodal satellite data and are already being adapted by governments and NGOs for applications including deforestation monitoring, food security, and wildfire risk assessment.

Ai2 highlights that most environmental organizations lack the infrastructure or engineering capacity to manage labeling, fine-tuning, and large-scale inference themselves. The OlmoEarth Platform addresses this gap by providing infrastructure that can run models cost-effectively at the right time and place, monitor performance, and turn raw outputs into actionable insights. Ai2 draws on over a decade of operating platforms such as Skylight and EarthRanger, which it says must work reliably every day.

The platform processes dozens of terabytes of imagery daily at a cost of fractions of a penny per square kilometer, according to Ai2. To achieve this, the system divides inference jobs into three stages matched to distinct hardware profiles: CPU-intensive data acquisition and preprocessing, GPU-based model inference, and CPU-based postprocessing. Ai2 reports that this staged approach keeps expensive GPUs fully utilized while distributing work across many machines.

Ai2 describes the execution layer, OlmoEarth Run, as dividing geographic regions into partitions sized for individual compute instances, then subdividing those partitions into smaller windows processed independently. This design enables thousands of compute instances to operate in parallel. Ai2 cites a recent wildfire risk map covering all of North America that, at peak, used roughly 19,600 CPUs and 994 GPUs in parallel with network throughput exceeding 168 GB/s. Ai2 estimates this reduced serial compute time from 4,737 hours to 30.5 hours—a 155× speedup.

Ai2 notes that parallelism is a per-run knob constrained by cloud quotas, and that output resolution, model size, and caching strategies trade off data volume, compute, storage, and speed depending on the task and budget. The platform relies on public STAC catalogs and open standards where possible, but maintains its own metadata index to avoid overwhelming external services during large inference jobs. For datasets hosted on AWS Open Data, Ai2 receives SNS notifications for new scenes; for other providers, it polls upstream indices every few minutes.

Ai2 emphasizes that satellite inference differs from typical ML workloads due to the scale of inputs and outputs, the need for precise geographic alignment, and the complexity of acquiring and preparing data from multiple providers with varying projections, resolutions, and cloud cover.