Dunbar is an anthropologist at the University College of London who hypothesized that there is a cognitive limit to the number of individuals with whom any one person can maintain stable relationships, and predicted that 150 is the "mean group size" for humans.
Ross Mayfield wrote a very interesting post last year starting from there on Ecosystem of Networks and came out with this nice summarized graph, which I have been struggling a little with to be honest since I saw it for the first time (are axes consistent from one raph to the other?..) but still it is intriguing.
Essentially, as we increase group sizes beyond 80, to 150, 200, or even 350-500, we typically do so by breaking larger groups down into smaller ones, and continually reducing community sizes down to the point where they can be understood and managed by people -- and so efficiency reasserts itself.
In my experience and vision of communities of practice, I tend to find similar numbers floating around. Typically vibrant communities of practice have around 100 - 150 members. As social structures, they are "onion-shaped", with layers of membership behaviors.
At the center, the "core group" of the community of practice is typically composed of 5 to 7 people. These are the guys who are willing to spend some time together, typically 15%-20% of their time (not much more, because they are busy on their projects anyway), reflecting on past experience and planning ahead for the community's learning activities in a peer mode. Then you have a second layer of 20-30 active contributors, typically those who follow the community ritual: they come regularly at meetings, they often contribute, and they also complain when something goes wrong in the planned schedule. These are the ones, whose attention is grabbed by other topics but have made some time for the community activities in their calendar, typically 2% to 5% of their time. And finally you have the "lurkers", who actually don't follow the community ritual, but participate just enough to be aware of what is going on. They also contribute a minima to maintain a feeling of social belonging, typically a few hours twice a year.
You might add, to try and connect even more to Ray Mayfield's graph, that a community of practice is often surrounded by a larger community of interest: people who are somewhat interested in what the community of practice does, and who will read its publications, and even give feedback in a point to point mode following something like the power law distribution.