Fair graph node classification aims to reduce the dependence of predictions on sensitive attributes while preserving predictive utility. However, many existing methods rely on a shared representation path for fairness and utility, which may couple potentially competing objectives. In addition, fairness-oriented modeling is often centered on local neighborhoods, which may overlook fairness-related information from broader structural contexts. To address these issues, we propose FairMoE, a mixture-of-experts framework for fair graph node classification. FairMoE functionally separates the two objectives through a Fairness Expert Group and a Utility Expert Group, which respectively perform fairness-oriented and utility-oriented learning. Within the Fairness Expert Group, multiple Fairness Experts operate over different receptive fields to learn fairness-oriented representations from local and higher-order structural contexts. In parallel, the Utility Expert Group employs complementary experts to preserve task-discriminative information. The two outputs are then coordinated by a Dynamic Gate, which produces node-specific weights to control how much utility information supplements the fairness-oriented prediction. Experiments on multiple real-world datasets demonstrate that FairMoE achieves a competitive trade-off between fairness and classification utility, with particularly strong results on fairness metrics. Further ablation and structural analyses support the effectiveness of the core designs. The code will be made publicly available upon acceptance.