Cobol is an extremely verbose programming language, and it was used in an era when the practice of programming was much less developed. Calls into libraries were often not used, and instead any re-used code was copied, essentially inlined by hand. (With all the obvious problems that caused.)
The combination of automating complex processes, requiring embarrassing amounts of code to do simple things, re-use by copy and the fact that it was dominant in it's field for such a long time (4 decades!), the amount of COBOL code that exists out there is just staggering.
I know of 1 smaller Cobol company. They alone had 10 to 100 million lines, after a decennium of trying to move to Java. Source files containing 1 routine were easily a few KLoc. The number in the article is probably too low.
The main problem with these code bases are they predate modern coding practices, so the sheer size, incomprehensability and untestability will crush you. You can easily spend your whole carreer reading the source code and reach pension age before you finish. Also, the organisational difference between 2 code bases is much bigger than modern code bases, as every company invented their own practices, and people rarely switched companies so didn't know what others were doing.
That number sounds enormous. If the same code runs on 10,000 ATMs, are they counting that 10,000 times?