From the point of a rational organisation it doesn't make sense either. Do it half-way and you end up with everyone on video calls to people who aren't there; do it fully and you're overpaying for the space because, yes, really, it's
colossally inefficient to have everyone on-site when they don't need to be.
Weirdly it also doesn't make sense to give people full flexibility to pick when to be on-site, though - at least, not from the point of view of their role within the larger organisation. Yes, you want them to be able to choose to be on-site when someone they need to work with is there, but also there needs to be a mixing function: as their boss you need to be ensuring that they are physically proximate to functions they haven't previously interacted with. You don't need much of this, but the main benefit of on-site working is discovery, not collaboration within an existing framework. And without a certain amount of curation that's unlikely to happen by accident. When two people get chatting in the kitchen, you need one of them, at least, to have that spark of recognition that says "Ah ha! Your thing is relevant to my thing!" because that's the fundamental innovation and discovery driver.
What's happening is that the bosses are leaning on random chance to do that curation job. I think that for most organisations it's very, very unlikely to be doing a good enough job to compensate for the downsides, but it also explains why the 37Signals approach of getting everyone in the building together at once works for them. That's still relying on randomness, but it's stacking the deck. Or rather, un-stacking it. If you allow people to pick which day a week, say, they come into the office, you'll get the Monday people, the Thursday people, and ne'er the twain shall meet. With everyone on-site at once, there's at least a chance a natural Monday will meet with a natural Thursday and produce an outcome that wouldn't have happened otherwise.