Ai2's OlmoEarth platform turns satellite inference into a distributed systems problem

When the Allen Institute for AI (Ai2) set out to build a platform for geospatial inference, it hit a wall other ML systems rarely encounter: satellite imagery is not neat. A single fine-tuning job can move terabytes of data and run for hours. Inputs span multiple spectral bands, sensor types, time steps, and providers, each with different projections and resolutions. Clouds block pixels. Missing bands are routine. And the output is a map, which means every prediction must remain locked to the same geographic grid as its neighbors.

The OlmoEarth Platform, detailed in a recent post on Hugging Face, is Ai2's answer to those constraints. Designed for governments and NGOs that need to monitor deforestation, crop health, or wildfire risk at national to continental scales, the platform treats inference as a distributed systems problem first and a modeling problem second. This framing is gaining traction beyond geospatial AI; the idea that orchestration can matter more than raw model size is one that a growing number of AI teams are exploring.

Three stages, matched to hardware

The core insight is that GPU time is expensive and most of an inference job's clock is spent on data acquisition and preparation, not the model itself. The platform splits each job into three stages, each assigned to the hardware best suited to it.

Data acquisition and preprocessing run on CPU instances with high I/O. They fetch, reproject, align, and normalize imagery from sources like Sentinel-2 and Landsat, then write it in a format optimized for fast GPU loading. The inference stage runs on GPUs, executes the model's forward pass, and streams minimally processed outputs straight to blob storage. Postprocessing, which stitches per-window outputs into seamless rasters in formats like Zarr or GeoTIFF, goes back to CPU.

Multiprocess data loaders keep each GPU fed continuously. The division prevents the system's most expensive hardware from idling while slower operations finish. This kind of careful pipeline design echoes the approach taken by GeoLibre's recent work on browser-based geoprocessing, which also prioritizes latency over raw throughput.

Data acquisition as a bottleneck

Acquiring the right pixels is harder than processing them. The platform must query multiple providers' STAC catalogs to find which scenes cover a given region and time range. For optical imagery the goal is the least-cloudy scene; for synthetic aperture radar, the relevant polarization channels matter. A large job can generate thousands of metadata queries, far more than external services from ESA or Microsoft Planetary Computer are designed to handle concurrently.

Ai2 sidesteps this by maintaining its own satellite imagery index, updated via SNS notifications from AWS Open Data sources and by polling upstream catalogs every few minutes. At runtime, the platform selects the best available source and performs windowed reads against cloud-optimized formats like COG or Zarr, retrieving only the bytes needed for a given partition rather than downloading entire scenes. The same index powers annotation tools, serving tiles from any indexed scene through the same windowed-read system. Automating this data pipeline is exactly the sort of grunt work that Alibaba's DataBridge Agent aims to simplify for AI workflows in other domains.

Resilience at planetary scales

At the scale this platform targets, failures are not exceptions; they are design inputs. Providers go slow or offline. Metadata points to imagery that turns out to be missing a band. Cloud cover leaves too few usable observations. Tasks crash.

The OlmoEarth Platform treats every task as idempotent and reentrant. For each geographic partition and stage, it provisions a virtual machine running a Docker container with the task parameters. The runner executes the work, returns the result, and shuts down. If it fails, the system retries automatically, falling back to alternate providers when available. A separate monitoring process detects runners that have stalled and restarts their tasks.

Concrete results: 155x speedup at fractions of a cent

The platform's most dramatic demonstration so far is a wildfire risk map covering all of North America. 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 155x speedup.

The cost: fractions of a penny per square kilometer. The platform can already run inference across continent-scale areas in roughly a day, processing dozens of terabytes of imagery. Those economics are further improved by the OlmoEarth v1.1 model release from mid-2026, which cuts compute costs by up to 3x over the original v1 while maintaining similar performance through a smarter token merging technique. The broader trend of driving down inference costs is why Together AI's $800 million bet on open inference infrastructure is attracting so much attention.

Parallelism is not unbounded. The platform exposes several knobs per job: fan-out is limited by cloud quotas; output resolution trades data volume for detail; model size trades GPU time for accuracy; caching raw imagery trades storage for speed on repeated runs over the same area. The right setting depends on the task and budget.

What comes next

Ai2's roadmap, shaped by partner feedback, includes automated model runs triggered by new imagery in the index, change detection and alerts that surface deforestation or flooding as push notifications rather than rasters someone has to inspect manually, and agentic tools that lower the barrier for non-specialist users to do work that previously required an experienced ML researcher. The team is also working on a dedicated embedding model; precomputed embeddings at global scale could replace full forward passes over raw imagery for many tasks, making workloads substantially faster and cheaper.

The platform currently runs on Google Cloud, but the architecture is designed to support multiple clouds and deployments within a partner's own compute environment. For organizations working in conservation, food security, disaster response, and climate, that flexibility matters.

Geospatial foundation models are still an emerging technology, and the gap between what these teams need to understand about the planet and what their budgets and technical resources allow remains wide. Ai2 is betting that treating inference as an infrastructure problem first can help close it.