To all the people banging on about upgrades... exactly what in the name of silica are you working on that this machine can not possibly handle for the next 5 years? What could you possibly want to upgrade? And what are you currently running?
So each time you want to upgrade your GPU, you're also having to purchase a new screen. The difference in my case was between having about 10fps with a lot of debug draw in the UE4 editor on the iMac, and about 90fps for the same thing on a new PC with a 1080.
About the only good thing about the iMac was that it held its value remarkably well, and I sold it for half the purchase price 3 years later.
Just because you cannot imagine what people might need to use a computer for does not give you the right to tell them how much performance they need.
Perhaps they are going to the arctic circle and have to rely on parts lying around if anything breaks!
Perhaps they work with cutting edge computer graphics and need a brand new GPU with a certain new feature or hardware chip to test out!
Perhaps they are simulating something gigantic and it is easier to test out parts of the code on a local machine with a PCIe add in card before pushing it to a supercomputer.
This is the definition of a PRO machine, people who have extraordinary requirements and need power and flexibility. Stop telling them they have "enough", you are not in the position to make that call.
With respect though, you're talking about different sets of constraints.
The iMac pro is inevitably targeted at the group of people who don't need to upgrade, and maybe they will replace their machine every three to five years. Realistically, there is a limited amount of upgrading that will be done for most users, meaning that engineering effort that goes into building an expandable machine is relatively wasted.
The iMac Pro will not be suitable for every professional user, but it will cover a lot of use cases, and that's okay. If you do need an expandable machine, then there are other options there – but most people don't.
Thunderbolt 3 just isn't fast enough for that to be feasible. The bottleneck is such that there's already no point in putting a high end card like the GTX 1080 Ti in an eGPU enclosure (Source: https://arstechnica.com/gadgets/2017/08/laptop-external-grap... ).
The whole point is I have no idea exactly what I'll be working on 5 years from now or what kind of machine it might need. But what I do know for certain is that the things I'm doing today would have been extremely tricky to do on a machine from 5 years ago.
And I can imagine specific hardware improvements for ML or VR in the near future.