Zstd belongs to the class of speed-optimized compressors providing “tolerable” compression ratios. Their intended use case is wrapping some easily compressible data with negligible (in the grand scale) performance impact. So when you have a server which sends gigabits of text per second, or caches gigabytes of text, or processes a queue with millions of text protocol messages, you can add compression on one side and decompression on the other to shrink them without worrying too much about CPU usage.
Xz is an implant of 7zip's LZMA(2) compression into traditional Unix archiver skeleton. It trades long compression times and giant dictionaries (that need lots of memory) for better (“much-better-than-deflate”) compression ratios. Therefore, zstd, no matter how fashionable that name might be in some circles, is not a replacement for xz.
It should also be noted that those LZMA-based archive formats might not be considered state-of-the-art today. If you worry about data density, there are options for both faster compression at the same size, and better compression in the same amount of time (provided that data is generally compressible). 7zip and xz are widespread and well tested, though, and allow decompression to be fast, which might be important in some cases. Alternatives often decompress much slowly. This is also a trade-off between total time spent on X nodes compressing data, and Y nodes decompressing data. When X is 1, and Y is in the millions (say, software distribution), you can spend A LOT of time compressing even for relatively minuscule gains without affecting the scales.
It should also be noted that many (or most) decoders of top compressing archivers are implemented as virtual machines executing chains of transform and unpack operations defined in archive file over pieces of data also saved there. Or, looking from a different angle, complex state machines initializing their state using complex data in the archive. Compressor tries to find most suitable combination of basic steps based on input data, and stores the result in the archive. (This is logically completed in neural network compression tools which learn what to do with data from data itself.) As some people may know, implementing all that byte juggling safely and effectively is a herculean task, and compression tools had exploits in the past because of that. Switching to a better solution might introduce a lot more potentially exploited bugs.
