Five researchers and one admin is cheaper than two researchers, two postdocs, a grad student, two admins, a compliance officer, an HR office, a grants office, a budget office, a legal office, some Deans, Vice Deans, Vice Chairs, Vice Chancellors, etc. You get my point.
I'm sure there are particular examples of places with high overheads, even higher than the 70% at top universities, especially in the biosciences, but historically pure soft money places have been relatively cheap to operate.
A small independent group might have low overhead but from what I see in the ML research space those types of groups are very limited in what they can do. These small groups also tend to work as subcontractors for large groups anyway, so the overhead advantage becomes somewhat moot. Just submitting research proposals to funding organizations regularly in order to get consistent funding requires a large group, especially these days with lower funding per project. Basically, five researchers and an admin won't survive very long without latching onto bigger groups. Also it's pretty well known that universities are cheaper than contractors by a wide margin.
I'm sure there are particular examples of places with high overheads, even higher than the 70% at top universities, especially in the biosciences, but historically pure soft money places have been relatively cheap to operate.