As long as you don't have "realtime" workloads, owning the GPUs quickly becomes the economical option. The main cost problems is in e.g. chat applications where the workload is spikey, and users expect an near-instant response, for which you need to scale the GPUs to the highest spikes of the workload.