TFA says the snowflake is a good model for networks that are inexpensive to repair, not necessarily robust. Considering that most repairs will happen at level 2 or level 3, that may be true ON AVERAGE. As the number of total nodes grows, I bet there is a point where the central node, which supports the most connections, becomes the expected common failure mode of this kind of network. Not only is the central node, by necessity, the most complex and by far the most expensive to repair (every level 1 function is down at this point), because of its complexity it may also have the shortest mean time between failures.
As soon as you see this and try to go back to a redundant central node, the next level nodes become vulnerable. And so on. Vulnerability propagates down the levels. The snowflake, er, melts down.
Maybe there needs to be a limit to the number of branches per node....but then you will have more than 3 levels.