Is this just the cost to train the final model, or does it include the cost of all the R&D leading up to that model (e.g. the cost of training smaller versions for hyperparameter tuning)?

In either case, how will the cost be determined, since these numbers are never made public, and the public estimates can vary significantly?