COBOL seems to be very translatable to any modern language, e.g. Pascal or BASIC. Real world reasons against "why not just translate it then" are understandable, but at some point that could become economical.
COBOL's numeric stack doesn't directly translate to many modern languages, which makes the translation introduce all sorts of fun edge cases that are extremely difficult to track down.
IEEE 754, which most of our modern languages are mostly compatible with, came about in '87. Long after COBOL had standardised itself on how numbers were expected to behave.
Beside AOT compilation to native code, Fujistsu and Micro Focus have Cobol compilers with .NET and Java backends, and I wouldn't be surprised if they eventually come up with WebAssembly support as well.
Also don't forget that languages like C and UNIX are just 10 years younger than COBOL.
yeah, I wrote a neural network a while back to run both in front and back end. to the front end I used JavaScript and for the back python. apart from matrix multiplications (python I used numpy, JavaScript I wrote my own class) the code is almost identical.
that got me wondering that I could write a program to translate a language to another. a large company could commission a (bunch of) programmer(s) to do this
