Assuming you have a high-speed private network as your interconnect, the cache to cache block exchanges of Cache Fusion should never saturate the channel, even with the high frequency of DML common to OLTP systems.
Of course it does not scale 100% linear, but I have seen published benchmarks of over 80% scalability when adding a sixth, eighth, tenth node and beyond to an E-Business application database. I personally have enjoyed results of near-linear scalability going from four nodes to eight with very acceptable latency numbers within the interconnect in support of a highly utilized OLTP database.
When scaling your database horizontally, just as you do vertically, you need to do testing to create a cost-based analysis in determining the benefits of how many additional nodes you need or want.
Obviously this is all generality. I am sure there are databases that exist that do not scale well with additional RAC nodes (though I could argue they probably wouldn't scale vertically with much benefit either), and I understand the scenario where deadlocks can occur with application tiers that create multiple connection pools, etc. But as a rule for consideration there is no reason you cannot continue to enjoy the scalability benefits of RAC well beyond four to five nodes for fear of bottlenecks within the interconnect.
Hope this helps,
This was first published in October 2006