Sorry but your argument doesn't seem coherent: How is the cost of RL relevant here?

It would also help if you could substantiate your initial claim (i.e. "internet training data is not where frontier capabilities come from")