I think it also might have been more successful if instead of porting it to POWER they'd ported it to x86 – they might have needed to add some tagged memory extension to x86 to do that, but I'm sure Intel would have been happy to talk. Indeed, the same idea has a history in Intel, if you look at the iAPX-432 and the i960, so Intel would certainly have understood the request–they might have even been excited about it. (Of course, given IBM's commitment to RISC and POWER, wasn't going to happen.)
Like a lot of other IBM systems, it became more closed over time. The System 38 had huge amounts of very detailed technical documentation publicly available. With the AS/400, much of those details moved under NDA, and it has only got worse over time. Now, you can't even write a compiler for (non-PASE) AS/400 without signing an IBM NDA, the executable format is under NDA. How closed is that!
I even wonder, what if instead they'd ported 400 to 370/390/z, or tried to make 38/400 and 370 more compatible from the start? Maybe some of these ideas might have cross-pollinated to MVS/VM/etc. A lot of what is most original about the 38 and 400 (single-level store, capability-based addressing) actually descends from the Future Systems project, a cancelled 1970s effort to design a next-generation successor to the 360/370. Maybe those technologies might have been more influential if they had a future on the mainframe? A merger of the AS/400 and 9370 (low-end mainframe) lines could have led to that.
Different tools designed to do different things.
I'd actually like to learn PL/I (or COBOL beyond a cursory knowledge), I'm someone who has a great deal of knowledge of the thought patterns of programming (as well as how you architect software, systems design, etc), without knowing any one tool well, I tend to learn a tool for a task, then forget the tool when I no longer require it anymore. My job is.. hard to explain, but its quasi programming, but also very much not aside from helper scripts I write now and again (I use all of the thought patterns of programming, without any of the tools). I enjoy writing software, but only have so many itches to scratch in my personal time.
PL/I is a general purpose applications programming language. Its primary design goal was to unify COBOL, FORTRAN and ALGOL into a single language, and as such was intended to meet the needs of both business and scientific application programmers. IBM invested heavily in trying to convince business software developers to move from COBOL to PL/I - see for example [0] - but, while there was a decent amount of business software development done using PL/I in the 1970s, it never had the success in that market which IBM had hoped for, PL/I ended up being an also-ran to COBOL instead of what it was meant to be, the successor.
PL/I proper was generally not used as a systems programming language. On the one hand, it had a number of features (decimal arithmetic, formatted and record IO and multithreading support built-in to the language syntax) which while useful in applications programming were unavailable or unsupportable or not useful in systems programming. On the other hand, a lot of features which are valuable in systems programming - such as inline assembly, the ability to manually control the allocation of variables to machine registers - are not in PL/I proper.
Given this, it was common to create systems programming dialects of PL/I which removed those application-centric features and added systems programming specific features. But those dialects were not properly speaking PL/I and were often called by other names - IBM’s was called BSL, later renamed to PL/S, which spawned a bunch of descendants with similar names (PL/AS, PL/X, PL/DS, PL/MI, PL/MP, PL.8). Honeywell had PL/6 which was used as the systems programming language for its CP-6 operating system (a backward compatible rewrite of Xerox CP-V). Prime Computer had PL/P as the systems programming language for PRIMOS. Gary Kildall developed PL/M in which he wrote the first versions of CP/M (later versions were rewritten in assembly to get better performance, but even in later versions PL/M was still used to implement some of the utilities.)
However, none of these systems programming languages were in the strictest sense PL/I; the original unmodified PL/I (without subsetting or supersetting) was an applications programming language not a systems programming language.
[0] http://bitsavers.org/pdf/ibm/360/pli/SC20-1651-1_A_Guide_to_...
Hear, hear. I started out in the block and record-oriented I/O world, and quickly found that being unable to make a quick use of highly advanced I/O system calls to be frustrating – for simple tasks. One had to define the record structure, typically in MACRO-11/MACRO-32 (we did not use COBOL, anyway), and only then commence the I/O. Record-oriented I/O also comes with its own wealth of quirks, e.g. if an application crashes / the OS restarts, the system administrator has to determine which files had not closed and manually close each of them off after the crash / system restart. Applications would refuse to start up if the files they operated on had not been properly closed. Blimey. Direct device I/O was even more complex.
Even the block I/O wasn't as trivial to use, especially for the simple text processing as the system did not keep the used bytes counter for the last file's disk block; the app would have to read all the blocks in and then either search for a special control character that indicated the end of the text or at least one text editor reserved the very last word of the last file's disk block as the total file character counter. Wash, rinse and repeat. All in all, as soon as I could get a hold of access to a UNIX box, I ran away screaming.
UNIX shifts the disk I/O complexity into the application layer (user space) and raw devices (user space + VMM bypass in the kernel space). Want to implement a record-oriented I/O? Write a user space application for it. Want to implement a high performant, record-oriented I/O? Use a raw device (Oracle absolutely wanted to give exclusive access to such one, and it was very fast for its time). That has yielded a highly successful cottage industry of databases, relational and record-oriented ones (e.g. BerkelyDB).
Looking back, I have come to appreciate the record-oriented I/O, and I would love to be able to define my application's record structure in the COBOL-like way in C/Rust/whatever and use that for simple and slightly more complex things. I certainly would not want to have to deal with the complex legacy that comes with it, i.e. seeking out unclosed files after the restart to close them and stuff. For simple one-off things and for talking to simple device drivers, I will take my bag of bytes any moment.
In the early 2000's I was involved in porting Subversion to OS/400. I had to build a replacement for make and the build tools but it almost all compiled on the first try. Then spent the next year dealing with EBCDIC issues. By the time we were done, IBM had updated the C compiler so we could make it UTF8-native and that let us remove our entire patch.
There was/is a lot of cool stuff about the AS/400
As for handling "the abstraction between memory and storage", that's what memory mapped files are essentially. It's being done routinely in modern systems.