If the simulator becomes more dynamic, only being required when necessary (an SL sim is always running, regardless), it's conceivable the cost could drop dramatically. Concurrency, of course, remains a tricky subject - WoW goes with shards, with its own issues, whereas an SL location is shared by everyone, making for very modest numbers - from 40 up to maybe 100 - possible before the lagmonster grips everyone.
Funny coincidence — I just had a conversation with some friends about this a few days ago. Now, I don't know much about SL's grid architecture, but from what I understand, the grid consists of a number of (physical) servers, each running up to a certain fixed number of sims, right?
We were talking about lag (there's a lot of that to be found in the SL11B sims), and I observed that it might be better to have larger systems (physical or virtual) with some sort of dynamic resource allocation so that resource-intensive sims could have more CPU power, RAM etc. thrown at them, while little-used sims would be put on the backburner until they needed more power again. (Virtualization would further allow the Labs to migrate running sims to different hardware without interruptions in service, and other nice things.)
I imagine that might help with the lag. 100 people? I think you're going to get a "sorry, the sim is full" error long before that. :) In practice, ~70 seems to be about the maximum, and even then the lag will be unbearable. It's possible to support that sort of load, but you'd need to drastically cut down on scripts, both in your builds and (crucially) on your visitors. And really, that's a stopgap solution, isn't it? Scripting is one of the things that makes SL nice, and it should remain possible to make use of SL's facilities.
Scripts, BTW, are also the reason you'll never be able to power down sims entirely, even when there's noone in there. (Unless perhaps you had a sim that had literally no running scripts in it, but that's extraordinarily unlikely in practice.) I think it was in one of the Labs' own presentations that SL was once described as a world of interacting scripts occasionally visited by avatars: a spot-on observation.
Sharding (for public spaces, rather than databases) is evil and needs to go. It's also a public declaration that your company has been unable to design/code a proper solution, and instead decided to capitulate. :) It may work for games like WoW, GW etc. where having everyone in the same location isn't that important to begin with, but for SL it would be deadly — not to mention that if you have, say, five instances of a region each comfortably holding 20 people, there really is no reason why you couldn't have one comfortably supporting 100. And dynamic resource allocation and deallocation would ensure that you're still making good use of your iron even without spawning/despawning instances of regions as people come and go.
I do wish the Lindens well for their new endeavor, in any case. But I also hope that this won't spell the end of SL, because let's face it, a) there's nothing like it, b) the new project likely won't replace it, c) people have invested a lot into SL (financially, creatively, socially, emotionally etc.), and d) even if all that were no concern, if LL were willing to shut down SL in favor of a new project, who's to say they wouldn't shut down the new project the same eventually? LL would be well-advised to build a reputation as honoring commitments they made, explicitely or implicitely.
I just thought this was worth preserving.