The OlmoEarth Platform: Geospatial inference at planetary scale

The Allen Institute for AI (Ai2) has officially unveiled the OlmoEarth Platform, a sophisticated infrastructure designed to bridge the gap between cutting-edge Earth observation foundation models and the mission-driven organizations that need them most. While Ai2 has a proven track record of releasing powerful, open-source models, the reality for many NGOs, government agencies, and environmental researchers is that they lack the specialized engineering teams required to manage the complex lifecycle of geospatial data—from labeling and fine-tuning to large-scale, cost-effective inference.

The OlmoEarth Platform is the culmination of over a decade of experience operating mission-critical software like Skylight and EarthRanger. By providing a robust framework for running models at scale, Ai2 aims to turn raw satellite imagery into actionable insights for critical global challenges, including wildfire risk assessment, food security monitoring, and the fight against deforestation.

The Engineering Challenge of Planetary-Scale Inference

Most machine learning applications, such as Large Language Models (LLMs) or standard computer vision tasks, operate on relatively small data footprints. Earth observation, however, exists in a different dimension. A single inference job can involve terabytes of multimodal satellite data, spanning various spectral bands, sensor types, and temporal resolutions.

The primary hurdles in this domain are not just computational, but logistical:

• Data Heterogeneity: Imagery arrives from multiple providers with varying projections, resolutions, and quality levels.
• Data Availability: Observations are frequently obscured by cloud cover or missing entirely, requiring complex reconciliation.
• Geographic Consistency: Because the output is a map, every individual prediction must align perfectly with the coordinate grid of its neighbors to avoid seams or artifacts.

The OlmoEarth Platform addresses these challenges by automating the entire pipeline, allowing users to process continent-scale areas in roughly 24 hours at a cost of mere fractions of a penny per square kilometer.

Optimizing Hardware for Geospatial Workflows

A critical insight behind the platform’s efficiency is the strategic separation of compute resources. Because data acquisition and preprocessing often consume more time than the actual model inference, the platform avoids wasting expensive GPU cycles on tasks better suited for CPUs.

The architecture is divided into three distinct stages:

1. Data Acquisition & Preprocessing (CPU-intensive): This stage handles fetching, reprojecting, aligning, and normalizing imagery. It prepares data in formats optimized for rapid loading. 2. Inference (GPU-intensive): The model performs its forward pass, writing minimally processed outputs directly to storage. 3. Postprocessing (CPU-intensive): The final stage stitches individual windows into a cohesive map, applies necessary rescaling or masking, and exports the data into user-friendly formats like Zarr, GeoTIFF, or GeoJSON.

By utilizing multiprocess data loaders, the platform ensures that GPUs remain fully saturated, streaming outputs directly to blob storage to minimize latency.

Scaling Through Parallelism

The platform’s execution layer, known as OlmoEarth Run, utilizes a "fan-out" strategy to handle massive geographic regions. It partitions a target area into segments sized for individual compute instances, which are then further subdivided into smaller windows for model processing.

"At peak, the run used roughly 19,600 CPUs and 994 GPUs in parallel, with network throughput exceeding 168 GB/s. That level of parallelism reduced an estimated 4,737 hours of serial compute to about 30.5 hours of wall-clock time—a 155× speedup."

This modular approach allows for massive horizontal scaling. Because each window is processed independently, the platform can reconcile overlapping partitions to ensure a seamless final raster. However, this parallelism is a tunable knob; users can adjust settings based on their specific budget and accuracy requirements, balancing output resolution, model size, and caching strategies.

Intelligent Data Indexing

Finding the right pixels is often the most difficult part of the process. To avoid overwhelming public APIs like those from the European Space Agency (ESA) or Microsoft Planetary Computer, the OlmoEarth Platform maintains its own internal metadata index.

This index is updated in real-time via SNS notifications from AWS Open Data or through periodic polling of upstream catalogs. By storing pointers to the best available imagery, the platform performs "windowed reads" against cloud-optimized formats (like COG or Zarr). This means the system only downloads the specific bytes required for a task, rather than entire satellite scenes, drastically reducing I/O overhead.

Best Practices for Earth Observation Data

Ai2 emphasizes that the most efficient providers share three key characteristics:

• Queue-based notifications for new imagery.
• Cloud-native storage without restrictive rate limits.
• Cloud-optimized formats that support ranged reads.

Resilience and Fault Tolerance

At the scale of thousands of concurrent processes, failure is not an exception—it is a guarantee. The OlmoEarth Platform is built with an "idempotent" design, meaning every task can be safely retried without side effects.

The system dynamically provisions virtual machines via Docker containers for each task. If a provider is slow, a band is missing, or a container crashes, the platform’s monitoring layer detects the stall and automatically triggers a retry or falls back to an alternate data source. This self-healing architecture ensures that long-running jobs can complete without human intervention.

The Future Roadmap: What’s Next for OlmoEarth

Ai2 is not stopping at current capabilities. The roadmap for the OlmoEarth Platform is heavily influenced by the needs of its partners, focusing on lowering the barrier to entry for non-technical users.

Key Upcoming Features:

• Automated Model Runs: Scheduling inference jobs to trigger automatically as soon as new imagery is indexed for a specific region.
• Change Detection & Alerts: Moving from static maps to dynamic alerts that notify users of events like flooding or deforestation in near-real-time.
• Agentic Tools: Implementing AI agents that can assist with data curation, feature engineering, and model fine-tuning, allowing users without deep ML expertise to leverage the platform.
• Embedding Models: Precomputing global embeddings to replace full forward passes for specific tasks, which would make workloads significantly faster and cheaper.
• Expanded Modalities: Integrating new sensors and environmental data, such as ERA-5 weather data, to provide a more holistic view of the planet.
• Multi-Cloud Portability: While currently operating on Google Cloud, the architecture is designed to be cloud-agnostic, eventually allowing partners to run the platform within their own compute environments.

Closing the Gap

The gap between the data we have about our planet and the insights we need to protect it remains significant. Many organizations working in climate, disaster response, and conservation have the mission but lack the technical infrastructure to operationalize modern geospatial foundation models.

The OlmoEarth Platform represents a significant step toward democratizing this technology. By providing a reliable, scalable, and cost-effective pipeline, Ai2 is ensuring that the most powerful tools in AI are available to those working on the front lines of global environmental challenges. As the platform continues to evolve, it promises to transform how we monitor, understand, and respond to the changing state of the Earth.