That’s the problem of reference. Your individual queries might not consume much - especially when compared to the training - but the more people use it, the more the whole consumption is. At some point running those models will consume more than training them