I'm thinking many of the linear algebra algorithms used in R and Numpy (Python) began life in Fortran decades earlier. R might have the edge on Python as to speed but it is also a domain-specific language. You aren't likely to run a server or design games in R, for example, while these (and much more) are easy in Python.

A lot has changed since you ran SAS on a mainframe -- notably, the fact that no one runs mainframes anymore. Back then, RAM was very expensive and limited (8MB, in your case). Today, you can have nearly a million times as much RAM on a single high-end data server, compared to your IBM 370. And if one node is not enough, you can use ten or a hundred. Or a thousand. Both R and Python support massively parallel computing.

I take your point about SAS verification and all those PhDs who work there...yet I cannot recall any cases of R or Python giving spurious results. For that matter, I would hazard that many (perhaps most?) "real money" corporate decisions hinge on results from Microsoft Excel, which has known issues in its stats. (I am thinking of a non-standard formula for skewness, as a case in point.)

If paying nine grand a year for the absolute bare-bones version of SAS helps you sleep at night, then go for it. But I say R and Python will give the rest of us the same results for $0.00. A least-squares regression is still a least-squares regression, no matter how you get there.