Let me confirm the difference between your case 2 and 3:
As I first understood it, in both cases there are two machines. Let's call them A and B. Machine A is running a Manifold Server instance (Server) and machine B is running a Manifold Release 9 desktop installation (Desktop). In both case 2 and case 3, you launch Desktop on machine B and create and use a Database: manifoldserver data source that's connected to Server on machine A, right? But in case 2 everything is fast while in case 3 there are some delays, is that it? If that's it, I'd focus on the difference between the machines and the software stack used in cases 2 and 3.
It could be I misunderstood your description, and the difference between cases 2 and 3 is that in case 2 both the Server instance and the Desktop instance are running on machine A, whereas in case 3 the Server instance is running on machine A with the Desktop instance running on machine B with some kind of network in between the two. Is that it?
If so, then I'd guess the issue has something to do with the network software stack in use. Could it be something about how the VM virtual network connects to physical networks or something in the network stack has some sort of cache or latency? Those things are all very tricky, highly specific to how the VM works, and are not guaranteed to match how Windows does it. You end up debugging how the VM or the network stack is different from how it is in native Windows. I note that's one reason Manifold tech support doesn't support virtual machines.
I use Server a lot, and always when I run it with a network between the Server machine and the Desktop machines, it seems instantaneous regardless of the machines or network involved. I run with Windows 10 or 11.
You might want to try setting up Server on a real (that is, non-VM) machine, setting up Desktop on a different machine, and verifying that on the Desktop instance you can indeed create and use a Database: manifoldserver data source to the Server machine and it runs fast even with the network in between.
To eliminate strange effects from "smart" switches or other gear that may be trying to "help" the speed of the network with tricky caching or whatever, it might be useful to make the network connection as simple as possible, such as having the two machines next to each other and connecting an ethernet cable between the two: no router required. Modern ethernet ports will automatically flip to handle direct connections like that and you can connect using the IP address of the destination machine.
Once you see that runs fast you can start adding whatever else is in the network, like "smart" switches or whatever, to see if adding something like that suddenly causes a reduction in performance. If it's nothing in the network when Server and Desktop are running on real, physical machines, then maybe you can look into how the VM is configured for a connection to a physical network.
I realize the above is a kind of hit or miss approach to figuring out what's going on, but it's probably something to try anyway.