What you see and what you comprehend are 2 different things.  What is being done with Crackdown is not the same as running a MP game on dedicated server.  

First understand that each portion of the game and its environment are processed by multiple different servers.  To understand why it’s called cloud tech is to understand what the term actually means. 

It’s the main reason why you are getting confused with what a dedicated server is compared to what a cloud does.  A cloud uses multiple dedicated servers and distribute the work among those servers.  When load increase, the cloud has the ability to add more servers or reduce the number of servers used for processing.  Servers can be spun up or down based on load and the switch of this process happens in Nano seconds.  So when MS talks about their cloud this is what they are talking about.  The ability to distribute work among multiple servers at the drop of a dime and scale back those servers also at a given moment.  

 MP games that use dedicated servers cannot do anything that I just mentioned.  MP dedicated servers allow users to be matched on that one server and the server handles sending data to all parties but it does not distribute the work among multiple servers nor does it have the capability to spin up and down servers for load balance.

This tech is not new, it’s been used for years in the enterprise.  The new part is that it has not been used for compute work within a game world.  MMOs would be the closes you get and even those do not do what Crackdown is doing with its destructible environments.  Companies use cloud tech like this and send way more data than what Crackdown is doing.  I doubt we will see any difference in real world compared to the demo since the technology is mature.

It can't look as amazing because rendering resources are still limited. The physics calculations only off-load the CPU, all the extra persistent rubble and falling pieces still need to be rendered with their own lighting, texturing and shadowing.  More pieces, more memory demands, more geometry to handle, more smoke effects, fire effects.

Targeting 2-4 mbps sounds low, it's not though when you share your internet connection with your household. Plus how far will it spike during big collapses. A lot of new geometry gets created that needs to be send to the clients. (unless things breaking up is scripted and only the falling and collision physics are computed in the cloud) It takes a lot more bandwidth to send geometry than a simple lossless compressed video stream.
For example in their early build they had over 40,000 chunks being tracked by the server. Just updating the position and orientation of those chunks at 30fps is close to 220 mbps. In the new build it looks like there are far less pieces and after stuff hits the ground it stops moving. Yet buildings collapsing into eachother, creating new geometry and everything flying through the air colliding into eachother will make a big spike.

Those bandwidth requirements are why we still don't see any of the various lighting and other rendering enhancements that can be done in the cloud. It simply takes too much bandwidth to do anything that looks substantially better. Physics is the least expensive in bandwidth yet I'm curious how far they can push it and how much will still have to be done by the client. (For example tracking the pieces locally according to simple gravity rules while the server only updates collisions and provides new speed vectors next to introducing new pieces)