Geosynchronous satellites, such as Geostationary Operational Environmental Satellite, are incredibly rich with hydrological information, given their high temporal resolution and complete spatial coverage. While most hydrological applications of remote sensing data focus on cloud-free pixels covering the specific point of interest, we exploit the full spatial context of the complete images themselves. We will present a highly efficient and scalable, yet simple, AI-based hydrological modeling methodology that can make large-domain forecasts directly from satellite imagery capturing atmospheric dynamics. This skips data processing steps that can lead to misleading data artifacts and biases. We will analyze this modeling approach through the lens of computational reducibility, and discuss the hydrological process that can and cannot be reduced to learnable functions by AI.