Saving the replay costs data though. It’s not only the reviewing.
It does slightly come down to what sort of multiplayer experience you want. Budget in the temporary archiving of ranking multiplayer games. Sell it as a feature. It'll keep the whole thing ticking over a lot longer.
The review cost is nothing. Really. Communities love to self-police so giving them the tools to do it completely, or to a point where they can identify flagrant cheaters for paid mods to verify... It's the Stack Overflow model. It works.
Logging a rolling week's activity at that higher rate requires 53GB but you could dramatically reduce the burden by only keeping things with in-game reports and then deleting recordings after the last issue was dealt with.
My point is it's not nothing, but it's nothing by modern standards if you've got more than two brain cells to rub together.
If you only record player input: mouse movement, clicks and keystrokes, it should be much lower.
As I said, it's not nothing, but it's not bad.
Technically you could record only the players' inputs but to play back such a recording, you would then have to deterministically reconstruct the entire match from all players' inputs. That's a bit more complicated than just replaying the local part of the simulation that your client already does with what data the server sends to players during normal course of gameplay.
oliwarner's numbers amount to about 45 bytes per frame per player, which is not unreasonable.