Aerodynamic interactions within wind farms significantly degrade downstream turbine performance, with wake-induced power losses typically ranging from 10–20% and reaching up to 40% under adverse conditions along with amplified fatigue loading on structural components. Active wake steering through intentional yaw misalignment of upstream turbines offers a means to mitigate these losses, with energy recovery demonstrated in both numerical studies and full-scale field campaigns. However, wake steering introduces its own trade-offs: yaw-induced asymmetric loading can exacerbate fatigue in bearings and blade roots depending on offset magnitude and inflow conditions. Balancing short-term power gains against long-term structural integrity, while accounting for both wake-induced and control-induced fatigue, necessitates control and decision support frameworks that explicitly incorporate fatigue constraints. Surrogate models for structural load prediction provide a foundation to evaluate fatigue metrics across diverse operating conditions and control actions in a computationally efficient setting. Established approaches in this domain include Gaussian process regression, feed-forward neural networks, and polynomial chaos expansions, which map mean inflow statistics to aggregated fatigue indicators such as damage-equivalent loads. A common limitation, however, is the reliance on fixed-dimensional inputs, which restricts applicability to the specific farm layout on which the model was trained. In this work, we propose Gaussian Processes over Graphs (GPG) to embed topological features for accurate inference of wind farm power and loads with uncertainty estimates. A Gaussian Process (GP) encoded global inflow field provides graph-level conditioning without redundant node-wise concatenation, while the graph-structured GP captures non-linear wake dependencies that generalize across arbitrary layouts and wind conditions. Trained jointly on power and structural load data, the surrogate enables rapid layout-agnostic farm-wide inference of both power yield and damage-equivalent loads, forming the computational foundation for load-aware wake steering and system-level optimization. The uncertainty estimates provide a principled basis for risk-aware decision support, enabling conservative decisions under high uncertainty.