Data centers constitute a foundational element of modern digital infrastructure, supporting the rapid expansion of cloud services and Artificial Intelligence (AI) workloads. AI-oriented data centers, in particular, exhibit high energy consumption and highly variable power demand, creating significant challenges for operational planning, cooling management, and integration with sustainable or variable energy resources. Consequently, accurate load forecasting is vital for optimizing resource allocation, improving energy efficiency, and maintaining reliable electrical supply. In this study, we apply an ensemble model to perform very-short-term load forecasting for an AI data center, generating minute-ahead predictions at a 15-s sampling resolution. In particular, 4 kernel-based Gaussian Process Regression (GPR) models are developed and trained using historical measurements of AI-data-center power usage. The models are subsequently fused and evaluated over a 72-day testing period to assess their predictive performance and subsequently fused into a single ensemble forecaster. Results demonstrate that the ensemble achieves strong predictive accuracy, yielding approximately 2.8% MAPE on the test set. Notably, weekly weekday and weekend MAPEs reveal low initial errors that increase gradually over time, reflecting natural distributional drift in the High-Performance Cluster (HPC) workload. Despite this drift, the model consistently tracks fast load fluctuations and maintains stable performance across both weekday and weekend segments. These findings highlight the effectiveness of the proposed ensemble for high-frequency data center forecasting, while also motivating future work on adaptive and drift-aware learning strategies to sustain accuracy in continuously evolving computational environments.