He worked for a company that did industrial control systems and one plant (a steel mill I think) had a very old system controlling it that booted from tape. However, as it was very old the paper tape had long since worn out and had been replaced with a sturdy length of leather. He needed to make an alteration to the boot process so had to resort to the hand-drill.
Oops.
So you worked out some set of instructions with total execution time comparable to storage I/O. And put that into ~firmware. Then you read data from one device, and wrote data to another device. And kept repeating that, with other instruction sets, until you were done.
That's not very different from current usage. Except that most intermediate results go into various sorts of memory. Or /tmp devices, if there's too much data for memory. And at the CPU level, instruction sets are just about as small.
But the point is that one could process ~huge amounts of data. Many GiB, as I recall. Just very slowly, with lots of thrashing tape.
And now one can use the same approach, except with clusters. And process humongous amounts of data.