No their claim is that there are dimishing returns for a fixed compute budget (in training) to scaling up data past that threshold vs. scaling up params.

This doesn't take inference into account either, obviously.