Well, the other big network that used imps was the MILNET. I had the dubious distinction of having the only router (other than the "mail bridges") that was on the MILNET after they split the ARPANET. And indeed, the NSA, BRL, and a few others had IMP based systems (we even had TACs at BRL). PDP-11's running software from Dave Mills, driving 56KB lines. Dave was certainly PI on that project, but a lot of the work was done by Mike Petry and Louis Mamakos at the Unviersity of Maryland. Cornell (Scott Brim) had the operational responsibility to keep the initial NSFNet backbone going involving IBM won, with IBM RT PC's running AIX driving T1 lines. MERIT was the prime contracator. They indeed were backed by IBM (Jakob Recktor primarily). In fact, the computers didn't drive the T1 lines directly. Fearing their commo cards weren't fast enough to drive a T1 at full speed, they used muxes to break the T1 into 500K lines and set up some virtual circuits to provide a bit more interesting networking technogy. I was actually on the committee that evaluated those bids. I just found all that stuff filed away when I moved recently. There were six proposals considered. Of the six, three were thought to be actually implementable: UCAR (with BBN), MERIT (with IBM), and NyserNet (Cisco? I'm having a hard time remembering). There were also three other proposals including one from DEC which really wasn't all that well thought out (I don't recall the other two). The pane was me, Jake Fienler (SRI NIC), Jon Postel, and a few others. My old boss from BRL, Steve Wulff, was the NSF honcho in charge. He went on to Cisco after his stint at NSF. I can remember that after the technical eval, we looked at the bid prices. If I remember it was $45MM for Nyser, $30MM for UCAR, and $15MM for Merit.