Ai2 launches OlmoEarth Platform, open infrastructure for continent-scale geospatial AI inference

The Allen Institute for AI (Ai2) has launched the OlmoEarth Platform, open infrastructure designed to take its OlmoEarth Earth-observation models from research fine-tuning to running inference across entire continents. Ai2 published the announcement on its official Hugging Face blog on July 28, 2026.

What's new

The problem Ai2 is solving is explicit in its own framing: "That's why we built the OlmoEarth Platform: infrastructure for taking geospatial models from fine-tuning and evaluation to large-scale inference." Having a strong foundation model, in other words, isn't the same as being able to run it cheaply and quickly across an entire country or continent — that operational gap is what the platform targets.

The scale numbers are the headline. Ai2 says "the platform can run inference across continent-scale areas in roughly a day, processing dozens of terabytes of imagery," and it does so "at a cost of fractions of a penny per square kilometer." Under the hood, the system distributes work across "roughly 19,600 CPUs and 994 GPUs in parallel," which Ai2 reports delivers "a 155× speedup" over running the same workload without that orchestration layer. The platform exposes this through APIs meant to plug into existing workflows rather than requiring users to build their own distributed inference pipeline from scratch.

Ai2 is explicit about who this is for: organizations working on conservation, food security, and disaster response that have real Earth-observation problems but not the engineering headcount to operate distributed ML infrastructure — as the company puts it, "most organizations in the environmental space...don't have the infrastructure or engineering teams that can manage the full lifecycle."

Context

This follows Ai2's original OlmoEarth model family, a multimodal Earth-observation foundation model line the organization released in November 2025, trained on roughly ten terabytes of satellite and sensor data. Partners have already applied that model family to tasks like tracking mangrove change, classifying the drivers of forest loss, and producing country-scale crop-type maps. The Platform announced now is the operational layer that sits on top of those models — the difference between having a capable model available for download and being able to point it at a continent and get results back in a day. It's a common second act for foundation-model releases: the model ships first, then the infrastructure needed to actually deploy it at the scale the use case demands follows once real users hit that wall.

Why it matters

Geospatial AI has a particular version of the deployment problem: the interesting use cases — tracking deforestation, mapping crop types, monitoring disaster damage — are inherently large-area and time-sensitive, but the organizations best positioned to act on that information (conservation groups, food-security agencies, disaster-response teams) are rarely the ones with in-house distributed-computing expertise. By packaging a 155x speedup and sub-penny-per-square-kilometer economics behind an API, Ai2 is trying to close that gap directly rather than leaving it to each downstream organization to solve independently. Because both the underlying OlmoEarth models and this platform are open, it also sets a public benchmark for what continent-scale geospatial inference should cost and how fast it should run — a reference point other Earth-observation efforts, open or commercial, will now be measured against.