What's the downside? Don't they stop when they hit diminishing returns?

You’d think so, but I haven’t seen it explicitly discussed in their papers, and nobody else that I know of trains on that many tokens