Mapping the Planet, One Penny at a Time
Ai2's OlmoEarth Platform can scan a continent in a day for fractions of a cent per square kilometer. Here's how it works.
By The ai-compute Desk
To generate a wildfire-risk map covering all of North America, Ai2's OlmoEarth Platform recently marshaled roughly 19,600 CPUs and 994 GPUs in parallel, pushing network throughput past 168 GB/s. What would have taken an estimated 4,737 hours of serial compute finished in about 30.5 hours of wall-clock time — a 155× speedup — at a cost of fractions of a penny per square kilometer. [1]
That job is not a one-off stress test. It is the routine ambition of OlmoEarth: a family of Earth observation foundation models pretrained on roughly 10 terabytes of multimodal satellite data, paired with an inference platform designed to turn those models into operational tools for governments, NGOs, and conservation groups. [1] Since Ai2 released OlmoEarth v1 in November 2025, partners have applied it across a wide range of tasks, from tracking mangrove change to classifying drivers of forest loss to producing country-scale crop-type maps in days, scaling deployments to national, continental, and global areas. [2]
The distance between a capable model and a useful one is often the story of infrastructure — who built it, who can afford to run it, and who ends up locked out. That gap is particularly acute in the environmental sector, where the stakes are highest and the engineering budgets are thinnest.
Ai2 is explicit about the problem it is trying to solve. Most organizations in the environmental space — the ones best placed to apply these models — don't have the infrastructure or engineering teams that can manage the full lifecycle: labeling data, fine-tuning models, and running large-scale inference. [1] An open model alone does not close that gap. The OlmoEarth Platform is Ai2's answer: infrastructure for taking geospatial models from fine-tuning and evaluation to large-scale inference. [1]
Satellite inference is structurally unlike most machine learning workloads. Most ML models take in a few megabytes of data and produce a result in under a second — think LLMs processing a paragraph of text or computer vision models analyzing a photo from a smartphone. [1] Earth observation is different. A single job fine-tuning a foundation model for maximum performance can move terabytes of data and run for hours, and the inputs may span multiple spectral bands, sensor types, and time steps across a large geographic area. [1] The output is itself a map, so every prediction must remain precisely aligned with the same projection and coordinate grid as the areas around it. [1]
Even acquiring the imagery is non-trivial. Prediction jobs often spend more time downloading and preparing imagery than running the model itself, making efficient data pipelines critical. [1] Scenes come from providers that use different catalogs, projections, and publication delays. For optical imagery such as Sentinel-2, the platform generally wants the least-cloudy scenes available; for synthetic aperture radar, the available polarization channels may matter more. [1]
To keep the system's most expensive hardware on the right work, Ai2 splits each job into three stages matched to distinct hardware profiles. Data acquisition and preprocessing runs on CPUs with high I/O: fetch, reproject, align, and normalize imagery, then write it in a format optimized for fast loading. Inference runs on GPUs for the model's forward pass. Postprocessing — stitching outputs, applying masks, exporting to Zarr, GeoTIFF, or GeoJSON — runs on CPUs again. [1] Multiprocess data loaders continuously feed each GPU while completed outputs stream directly to blob storage. [1]
Parallelism is the engine of the platform's speed. OlmoEarth Run, the execution layer, divides a geographic region into partitions sized for individual compute instances, then subdivides those into smaller windows the model processes independently. A state-sized area might become a hundred or so partitions; a continent-scale run can become thousands. [1] Adjacent partitions overlap slightly, and the platform reconciles that overlap during assembly so no seam appears in the final raster. [1]
Fan-out is not unbounded, though. More workers push against cloud quotas, so parallelism is a per-run knob, one of several that Ai2 can adjust on individual jobs. Output resolution trades data volume and compute for detail; model size trades GPU time for accuracy; caching raw imagery trades storage for speed across repeated runs over the same area. [1] The right setting depends on the task — and budget — at hand. [1]
Cost is not an abstract concern in this domain. Every dollar saved on compute is a dollar a conservation group can spend on fieldwork, policy advocacy, or the next monitoring cycle. Efficiency gains at the model level compound across thousands of inference jobs.
That logic drove OlmoEarth v1.1, released to cut compute costs by up to 3x while maintaining the performance of the original models on a mix of research benchmarks and partner-constructed tasks. [2] The lever is token sequence length. In transformer-based models, compute costs scale quadratically with sequence length, so even small reductions meaningfully cut the cost of running the model. [2]
Collapsing those resolution tokens into one produces three times fewer tokens and material savings across pretraining, fine-tuning, and inference. [2] It is not straightforward, though. Naively combining the tokens leads to significant performance drops, including a 10 ppt drop on m-eurosat kNN, a common benchmark task for remote sensing models. [2] Ai2 hypothesizes that separating Sentinel-2 bands into different tokens makes it easier for OlmoEarth to model important cross-band relationships. [2] Merging tokens without impacting performance required modifications to the pre-training regimen. [2]
The newest iteration, OlmoEarth v1.2, adds rotary positional embeddings — RoPE — which reduce artifacts in the embeddings and give a small performance boost. [2] Older versions marked patch positions with a fixed signal, absolute positional embeddings, and that signal showed up as extraneous artifacts in the embeddings. RoPE fixes this: unlike APE, RoPE rotates vectors the model compares according to their distance, encoding patch location without introducing artifacts in the embeddings. [2] The update came directly from partners asking for cleaner embeddings. [2] OlmoEarth v1.2 comes in Nano, Tiny, Small, and Base — all open source. [2]
On the platform side, the roadmap extends toward automated model runs — scheduling inference jobs in advance or triggering them whenever the imagery index registers a new scene — and change-detection alerts that notify users when the landscapes they monitor change, so events such as deforestation or flooding surface as alerts rather than rasters someone has to find and inspect manually. [1] Further out, Ai2 is working on agentic tools and interfaces designed to lower the barrier to using geospatial models, so that users at any technical level can do work that previously required an experienced ML researcher. [1]
A dedicated embedding model is also on the roadmap. Precomputing embeddings at global scale would mean that for many tasks, running inference against those embeddings could replace a full forward pass over raw imagery, making workloads substantially faster and less expensive. [1] Fine-tuning and direct inference would remain important for maximum performance on challenging tasks, but embeddings could open up a much broader range of efficient applications. [1]
The real test of any infrastructure is not what it can do at peak, but what it enables for teams that could never have built it themselves — whether a small conservation NGO can finally afford to run the same analysis that a well-funded government agency takes for granted.
Ai2 frames the gap between what environmental teams need to understand about the planet and what their budgets and technical resources allow them to do as the defining problem it is building toward. [1] Whether open models and a managed platform can bridge that divide at the scale the climate crisis demands is still an open question — but the North America wildfire map, processed in 30.5 hours on a cloud that can be spun back down to zero, is a concrete data point in the argument. [1]
Most organizations in the environmental space – the ones best placed to apply these models – don't have the infrastructure or engineering teams that can manage the full lifecycle: labeling data, fine-tuning models, and running large-scale inference.Ai2