Add to that, you need a lot of overhead like administrative assistants, grant managers, technology transfer officers.
The former Dean of Georgia Tech's College of Computing reached a similar conclusion: http://innovate-edu.com/2011/05/18/if-you-have-to-ask-ten-su...
DOE is slightly more parsimonious with its overhead, NSF is the most conservative. DARPA gives overhead in all sorts of ranges. Or so I hear.
And TSRI doesn't give most of its professors salary, unless they are doing "administrative duties" (Pres, VP, dept chair, etc) If you run out of grants, you basically don't get paid anymore. At one point ~90% of TSRI operating costs were run off of NIH overhead, all coming in from a very dynamic patchwork of grants from all of their PIs. I think then the pressure to keep getting grants made the place a bit of a boiler room, there was an uptick in shady science (Peter Schultz's retractions, Geoffrey Chang's retractions, and others I'm not ready to talk about yet)... Eventually professors started leaving because they didn't want to take it anymore.
That's because an agency like the NIH, will have a program project and decide, "ok. We want project X to be done and we think it will take ~150,000 in JUST science costs" - in theory, that should be more or less invariant across the institutions that get funded. THEN an added cost is what it takes for it to get done in X institution. If overhead were calculated as % of the full grant, then the math gets trickier to isolate the "invariant science part".
The reason why separating the "invariant science part" is important, is because when the agencies send out an program call, they say, "you get $2M to do the science", and each institute that is bidding to get the grant is not supposed to know how much of that gets taken up by overhead by other institutes. The net result is each institute is competing for the same pool of "we can do science with this" money without the PIs having to worry about how much of the cookie gets eaten up by their own institute. In general, I don't really like much of anything to do with the federal government's involvement with science, but: The way this is structured is probably a good thing, since scientists are really bad at thinking about much besides their own science, and if you make them do this mental gymnastic, something's going to come short, and there will probably be more failure all around from lack of funds. The responsibility to figure out how to get that administrative cost then falls on the program manager - and the grant reviewers are instructed to take overhead into account.