Zupil, a Spanish AI startup building an earth system of record for physical-world data, shares how they use Anyscale on Azure (built on Azure Kubernetes Service with the Ray distributed compute framework) to scale Python AI workloads. The company processes satellite imagery with multiple spectral bands across trillions of pixels, combining CPU-based data preparation with GPU-based inference using foundation models like Prithvi (IBM/NASA). Key outcomes include GPU utilization in the high 80s–90%. Engineering advice includes decomposing problems into the smallest parallelizable unit, starting from the problem rather than model benchmarks, and favoring managed platforms over DIY infrastructure for small teams. Anyscale on Azure is now available in public preview.