Whilst there are some good arguments here against small caps (which are often there to milk customers), the technical understanding of these analysts leaves something to be desired, and they clearly don't really understand what happens in real networks.
For a start, it's not about averages at all, but about the instantaneous traffic presented to the network at peak times. Operators build networks to handle these peaks, and it is certainly true that a small number of users can dramatically change the peak level of traffic presented.
Contention at the access layer has more or less nothing to do with TCP/IP, but is a function of the layer 2 architecture (where multiple wireline users share a single access switch and its backhaul capacity - ATM in old DSL implementations, and Ethernet in newer ones). In wireless networks, it's much simpler, because the real constraint is the available bandwidth on a single sector (something that cannot easily be adjusted, unlike wireline) and one or two users with sustained high demand can do a lot more damage, since the multiple access protocols (e.g. in 3G networks) are designed to give high instantaneous bandwidth to an individual user, and hence reasonable performance for every user. What makes any access network practical at all is that a small proportion of users want lots of bandwidth at any one instant.
In the core network, it's true that the issue is how routers deal with TCP/IP when there is more traffic offered than the available bandwidth, and packets must be delayed or discarded.
However, the reason for individual users being able to "hog" the network is precisely because of the egalitarian nature of TCP/IP. Typical router queuing mechanisms do quite a good job of giving each TCP stream roughly the same priority and performance. However, since TCP has no inherent mechanism to address true "fairness" between users (rather than TCP ports / sessions), applications that want to grab more resources can employ various clever tricks, notably opening multiple parallel sessions.
Some interesting theoretical work has been done to "fix" TCP to enable really fair resource sharing between users, and to prevent these "hogging" strategies from being effective. Obviously, one can apply separate QoS mechanisms (e.g. MPLS etc) to try to improve matters, but typically not in regular broadband services, which are generally pure, best effort IP.
It's just another tactic from the supplier to milk/rape the consumer 
