Another drawback of ensembles is the computational cost they impose. Training a single neural network is computationally expensive. Training a 1000 of them requires a 1000 times more computational resources. Furthermore, some methods are sequential by nature. This means that it is not possible to harness the power of distributed computing. Instead, each new model must be trained when the previous model is completed. This imposes time penalties on the model's development process, on top of the increased computational cost.

Computational costs do not only hinder the development process; when the ensemble is put into production, the inference time will suffer as well. If the ensemble consists of 1,000 models, then all of those models must be fed with new data, produce predictions, and then those predictions must be combined in order to produce the ensemble output. In latency-sensitive settings (financial exchanges, real-time systems, and so on), sub-millisecond execution times are expected, thus a few microseconds of added latency can make a huge difference.