Sure, but if the model takes multiple seconds to execute, then even 100 milliseconds of network latency seems more or less irrelevant