Scaling inference with remote-sensing foundation models on large datacubes

High-resolution satellite constellations have grown Earth Observation (EO) archives to petabyte scale. While remote-sensing foundation models offer powerful capabilities for analysing this data, scaling inference across a variety of models to inputs far larger than memory, particularly on modest hardware, remains a challenge. We present GSIP, the GeoSpatial Inference Pipeline, a high-performance inference framework that decouples data processing from the specific ML model, facilitating extensibility and evolution with the rapid advances in remote-sensing models. To maximize hardware utilization in the face of compute and memory constraints, we employ a dynamic data processing strategy built on an adaptive Zone of Responsibility (ZoR) within a producer-consumer pipeline, entirely decoupling input preparation, GPU inference, and output consolidation. The framework automatically adjusts chunk and batch sizes based on available system and GPU memory, ensuring both high performance and robustness even when datasets exceed available memory. Across eight model configurations and two hardware tiers, GSIP runs both conventional classifiers and remote-sensing foundation models, including Prithvi-EO-1.0 and 2.0, SatlasPretrain, and the generative TerraMind model, on Sentinel-1/2 and Harmonized Landsat Sentinel-2 (HLS) inputs. It processes a full Sentinel-2 granule on a 4 GB-VRAM consumer laptop within an 8 GB memory budget (peak process RAM ≤7.00 GB), and classifies a 17.8-gigapixel, 170-scene mosaic covering 1.8 million km2 on the same laptop, with peak memory held at 10.51 GB against the 11.3 GB budget it was given, demonstrating that the analysable area is bounded by storage rather than by RAM. Under a matched 8 GB budget it is also faster and more accurate than equivalent pipelines assembled from standard geospatial and deep-learning tooling.