The first "workstations" I used were SPARCstations, diskless and noiseless Suns SLC that only held a CPU and RAM inside a monitor's backside, booted and operated off the network. I don't know if that was even 100Mbit/s or still Thick-Ethernet at 10Mbit/s, I believe CPU clocks were around 33-40 MHz but at least they were true 32-bit machines and even paged over Ethernet. You could swap the keyboard and the mouse, but there were no user manageable parts on these things otherwise.
It gave you the local compute power you needed for a (monochrome) graphical user interface and a (for the time) super responsive system, yet it also returned the power of the mainframe terminals, where it didn't matter which one you used, because all your code and data was in one central space. It also gave you a huge number of these relatively small yet powerful systems, as well as some faster and bigger but 100% compatible SPARCservers to hold your data, which you could use from your own workstation simply by adding an "on <system>" clause to any command or part of a pipe.
If your single workstation was too slow for a serial set of compiles or computues, you just had to invest a bit of brain power to field the work out to all the others and it became a bit of a game of who could use all disposable compute at once...
That completely stateless mode of operation with the perfect local responsiveness of local compute and "cluster/cloud" expandability has forged some significant paths in my brain and to me still represents an ideal environment.
I've always faulted Microsoft for tying their OS to a disk for copy protection and crippling network boot and operations as a result. Even more importantly I believe it held back the evolution of networks, which have lagged far behind the bandwidths available for storage, even if that was also distributed as in the case of SANs.
So for me a real workstation is stateless and thus ready to be swapped as a whole, without the user losing much if anything in the mean-time, least of all time or productivity. Now if that workstation can be fixed from spare parts locally afterwards or shipped whole is logistics details, which are likely to vary on where and to whom you deploy workstations and their range of capabilities.
If we used Thunderbolt/PCIe based networks instead of Ethernet, this stateless mode of operation would be much more natural. 10Gbit networks are much better than all their ancestors, but no match for NVMe class bandwidth and PCIe class latencies.