OlmoEarth Platform turns Earth observation models into planetary-scale inference systems

Introduction

Earth observation AI is not simply a modeling problem. In a new Hugging Face post, Ai2 describes the OlmoEarth Platform as the infrastructure layer needed to take geospatial foundation models from experiments to real-world, large-area deployments. The target users are not only machine learning teams, but also governments, NGOs, and mission-driven organizations that may understand the environmental problem deeply while lacking the engineering capacity to manage satellite data pipelines at scale.

Key points

• Beyond open weights: The OlmoEarth model family is pretrained on roughly 10TB of multimodal satellite data. The platform extends that work into fine-tuning, evaluation, and large-scale inference so model outputs can become usable maps and decision-support products.
• Satellite inference is data-heavy: A single job may involve multiple spectral bands, sensor types, time steps, and geographic regions. Imagery can come from several providers, each with different projections, resolutions, formats, and publication schedules.
• Hardware is split by task: Ai2 separates the workflow into CPU-heavy data acquisition and preprocessing, GPU-based model inference, and CPU-based postprocessing. This keeps expensive GPUs focused on forward passes instead of downloads, reprojection, and resampling.
• Partitioning enables scale: OlmoEarth Run divides a requested region into worker-sized partitions and then into smaller model windows. Slight overlaps between adjacent partitions are reconciled later to avoid seams in the final raster.
• Metadata indexing matters: Large jobs could overwhelm external STAC APIs if every worker queried them directly. Instead, the platform maintains its own metadata index, updated through provider notifications where available or frequent polling where needed.

Why it matters

The most interesting takeaway is that geospatial AI is becoming an inference-serving challenge. For environmental applications, the bottleneck is often not whether a model can recognize a pattern, but whether an organization can reliably fetch the right imagery, align it, process it cheaply, and deliver outputs in formats users can trust.

Ai2 reports that the platform can run continent-scale inference in roughly a day, processing dozens of terabytes of imagery at fractions of a penny per square kilometer. Those details highlight a shift: open models may create the possibility of broad access, but robust infrastructure determines whether that access becomes operational impact. OlmoEarth Platform is therefore less a standalone model announcement than a blueprint for making Earth observation AI usable in the field.

Source: Hugging Face Blog