Spatial reasoning is foundational for Vision-Language Models deployed as Vision-Language-Action agents in physical environments. This paper introduces MultihopSpatial, a benchmark designed for multi-hop and compositional spatial reasoning with complex queries across diverse spatial perspectives. The work proposes Acc@50IoU, a metric that simultaneously evaluates reasoning and visual grounding by requiring both answer selection and precise bounding box prediction. MultihopSpatial-Train provides a large-scale training corpus to foster spatial intelligence. Extensive evaluation of 37 state-of-the-art VLMs yields key insights into the capabilities and limitations of current models for robust VLA deployment in embodied AI scenarios.