I believe this was discussed in the Kusama Direction channel: you can find the beginning of the conversation HERE. I'd like to highlight some comments during the discussion:

• "BABE and GRANDPA are running smoothly with up to 1000 validators on Kusama, but the parachains consensus logic is far heavier and more complex and will need optimization to reach that point. As such, the parachains team is planning to only use some limited subset of the validators initially (perhaps 50, 100, or 200) randomly selected each session for parachain consensus. Rewards are still fair in expectation. They are just higher variance due to this. We do intend to scale up the active para-validation set"
• "We need to use the same set across all parachains because of availability and approval voting which aren't apparent from the UI but are integral to consensus: In a way each of those validators is securing all of the parachains, Not just the one they're assigned as a backer for at the moment"
• "...Technically, we will be able to achieve a situation where every validator is a parachain validator. I think reducing variance is a nice-to-have but at the moment there is no trade-off that seems sensible in order to reduce variance in the short term e.g.. reducing rewards"
• "There is probably some fiddling with parameters that could improve the short-term variance somewhat but the long term plan is to make all validators para-validators anyway"

Thanks Raul, just found it (29 Jun).

So the short answer is that there is no plans to rebalance in a short-term.

"There is probably some fiddling with parameters that could improve the short-term variance somewhat but the long term plan is to make all validators para-validators anyway".

I don't understand the implementation so maybe it's not doable but would it be possible to reduce the point from Para-Validating from 20 to 15 or 10 points.

This could reduce the extremes but still incentivise people to setup for Para Validator.