Bayesian Optimization (BO) is a popular approach to solve optimization problems using as few function evaluations as possible. In particular, Efficient Global Optimization (EGO) based on Kriging surrogate models has been successfully applied to many real-world applications in low dimensions (less than 30 design parameters). However, in high dimension, building an accurate Kriging model is difficult, especially when the number of samples is limited as is the case when dealing with numerical simulators. This is due to the inner optimization of the Kriging length-scale hyperparameters which can lead to inaccurate models and impacts the performances of the optimization. In this paper, we introduce a new method for high-dimensional BO which bypasses the length-scales optimization by combining sub-models with random length-scales, and whose expression, obtained in closed-form, avoids any inner optimization. We also describe how to sample suitable length-scales for the sub-models using an entropy-based criterion, in order to avoid degenerated sub-models having either too large or too small length-scales. Finally, the variance of the combination being not directly available, we present a method to compute the prediction variance for any weighting method. We apply our combined Kriging model to high-dimensional BO for analytical test functions and for the design of an electric machine. We show that our method builds more accurate surrogate models than ordinary Kriging when the number of samples is small. This results in faster convergence for BO using the combination.