Without the actual training corpus, you can't know how much it takes to train. They could have trained on twice as many tokens for example (not saying they did!)