According to everymac.com the PPC XServe was $3000 when it came out. Not saying it wasn't good value, but I don't think that was particularly cheap at the time. The big cost-saving though was no software licensing fees: things like Windows Server, Netware and commercial Unix implementations tended to be expensive and based on per-user licensing fees. Plus, XServe supported things like Apple File Sharing Protocol and worked with things like Mac Mail.
Then Linux happened - mostly free software, pay for whatever support you need, from a competitive market, runs much the same server software as MacOS and netatalk for Mac file sharing/Time Machine etc. (although Macs were shifting to SMB by then anyhow). There wasn't really much point to the XServe - nice UIs don't count for a headless server and swish looking hardware doesn't count for a box locked away in a machine room. I couldn't really see any argument for buying an Intel XServe other than habit/familiarity. ARM may have been a bit more interesting, but there were already multiple server-class ARM implementations.
As for Pro XDR - claiming that a fairly coarse mini-LED backlight matrix was somehow equivalent to those $20k reference displays (which had dual-layer LCDs giving pixel-accurate HDR) was always a stretch. However (ignoring the totally shameless pricing of the $1k stand) AFAIK it's still the only 32", 6k, mini-LED illuminated display in town (the cheaper 6k Dell doesn't have local dimming) so it's hard to compare prices.
OTOH, the "pro" Macs haven't always been that expensive if you compare like-with-like: the original Xeon Mac Pro was pretty good value at the time - for a dual Xeon "workstation". The 27" iMacs all had display panels which would probably have sold for the thick end of $1000 as stand-alone displays. Even the stratospherically expensive 2019 Mac Pro wasn't that out-of-kilter with comparable Xeon-W systems - that had similar PCIe bandwidth and >1TB RAM capacity. It was just overkill for most people and the base $6000 model made zero sense unless you were planning to add $10k worth of expansion.
...and I think that is the general industry trend, even with PCs. PC users I know who always used to have massive tower systems are moving to small-form-factor PCs and laptops for anything other than AAA gaming and AI training. At the other end, if you need heavy lifting you can rent it somewhere in the cloud - and for "trending" things like AI that's where your datasets already live - on the other side of the bottleneck of your personal broadband link.
The difference is that it has far more PCIe bandwidth (as well as more slots) than a PCIe expansion box, which is limited to 4 lanes (and only PCIe 3 without TB5). Expansion boxes will get a bit better with TB5 but probably not catch up completely.
Please, no. An M-series trashcan would be a massive form-over-function folly.
The trashcan design assumed that everybody was going to be using multple GPUs for computing, and assumed that the heat would be spread evenly between the CPU and two GPUs. That led to a triangular core & a cylindrical enclosure. Superficially, a great example of form-follows-function - except Apple bet on the wrong functionality, which is part of the reason why the Trashcan never got an upgrade (plus, it barely worked well enough for the original chips).
The M-series is all on one package and most of the heat comes from that package - and even though the ultra is in some respects two chips they're still fused together in a single package. There's no design reason for the Trashcan's triangle-in-a-circle layout. Having a fairly conventional active heatsink sat on top is boring, but it works.
