Accurate segmentation of fluid regions in optical coherence tomography (OCT) images is crucial for ophthalmologic diagnosis and treatment monitoring. However, automated segmentation models face two key challenges: high annotation costs and limited generalization across OCT devices. We investigate foundation models to address these challenges, evaluating a domain-specific OCT foundation model trained using the SimCLR method and an adapted Segment Anything Model 2 (SAM2). Through experiments with datasets ranging from 50 to 2,500 images and cross-device validation, we show that foundation models outperform ImageNet-pretrained models in small data regimes. In cross-device evaluation, both foundation models demonstrated superior generalization. Our findings indicate that foundation models significantly reduce annotation requirements and enhance cross-device adaptability, lowering development costs and accelerating deployment of OCT fluid segmentation solutions.