Abstract: Ocean foundation models offer a new intelligent backbone for environmental forecasting, disaster early warning, smart fisheries, and intelligent shipping, yet their large-scale deployment is constrained by persistent data bottlenecks: sparse and uneven observations, noise and systematic biases, heterogeneous sources that are hard to align and fuse, and limited sharing under compliance requirements. Based on representative practices in China and worldwide, this article examines how these issues translate into uncertainty and decision risk in fisheries management, maritime operations, polar expeditions, and disaster forecasting, and assesses key pathways for a multi-scenario data foundation, including multimodal spatiotemporal alignment, physics-constrained modeling, synthetic/simulation augmentation, quality governance with provenance and traceability, and secure collaboration via federated learning and privacy protection. We further highlight emerging directions, and conclude that a usable, trustworthy, and sustainable data infrastructure is essential for reliable deployment and continued model improvement.