The OlmoEarth Platform: Geospatial inference at planetary scale
Hugging Face ● Covered by 2 sources
Ai2 launched a platform to run its OlmoEarth satellite AI models at massive scale, not just release them as open weights. It can scan a whole continent in about a day for fractions of a penny per square kilometer.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ai2 has spent years building open Earth-observation models, but open weights alone don't help a small NGO trying to track wildfire risk or deforestation without a dedicated engineering team. So the company built something less flashy but arguably more useful: the OlmoEarth Platform, an infrastructure layer meant to take a foundation model from fine-tuning all the way to inference across an entire continent.
The scale here is the real story. A recent wildfire-risk map covering all of North America used roughly 19,600 CPUs and 994 GPUs running simultaneously, pushing over 168 GB/s of network traffic. That turned an estimated 4,737 hours of serial compute into about 30.5 hours of wall-clock time — a 155x speedup. Ai2 says the platform can now chew through continent-sized regions in roughly a day, processing dozens of terabytes of imagery for fractions of a penny per square kilometer.
Getting there meant solving problems that have nothing to do with model architecture. Satellite data arrives from providers like Sentinel-2, Landsat, and NISAR in different projections, resolutions, and formats, often with clouds obscuring the pixels you actually need. Ai2's answer was to split every job into three stages matched to the right hardware: CPUs handle the I/O-heavy work of fetching and aligning imagery, GPUs run the actual model inference, and CPUs again stitch the outputs into seamless maps. Keeping GPUs busy while CPUs do the grunt work sounds obvious, but it's the kind of detail that determines whether a job finishes in a day or a week.
Ai2 also built its own metadata index rather than hammering public STAC catalogs like ESA's or Microsoft's Planetary Computer, which weren't designed for thousands of simultaneous queries from a single inference run. The system listens for new-scene notifications via AWS SNS where possible and polls everywhere else, so it pulls data at a steady drip instead of a flood. Combine that with automatic retries, idempotent tasks, and fallback providers, and the platform is built to shrug off the routine chaos of distributed computing rather than pretend it won't happen.
What comes next is arguably more interesting than what's already shipped. Ai2 wants to add scheduled and event-triggered inference runs, automatic change-detection alerts for things like flooding, and agentic tools that let non-experts fine-tune models without an ML researcher on staff. There's also a push toward precomputed global embeddings, which could let many tasks skip full model forward passes entirely — a shortcut that would matter a lot to organizations already priced out of running large-scale inference themselves.
My take — AI-written commentary, not fact-checked reporting
This is Ai2 quietly admitting that open weights were never the hard part — the real bottleneck for conservation groups and disaster-response teams was always the plumbing, and nobody wanted to build boring infrastructure for them. I like that they're leaning into unsexy engineering (metadata indexes, retry logic, CPU/GPU scheduling) instead of chasing another benchmark headline. If this platform actually stays usable by teams without ML researchers on staff, it'll matter more for climate outcomes than another point of accuracy on a leaderboard.
Read more about this at: Hugging Face