One of the primary virtues of b-trees is that they make efficient use of disk. Even if the full data structure is too large to fit in memory, because of the large branching factor you can usually fit all the interior nodes in memory. So, in a lookup, only the final access of the leaf node requires a disk access. Of course, traversal may require additional seeks and reads.
You're definitely right that the cache should be hot though, so I am now also a skeptical hue.