Indeed, this is even documented in the code of compute_requiredby function :
* A depends on B through n depends <=> A listed in B's requiredby n times
* n == 0 or 1 in almost all cases */

iirc, this was introduced with a patch from Nagy. It should be possible to find the reference on pacman-dev ML about this.
I agree it looks odd, but I still have two minor arguments for the current behavior
1) the implementation is natural. multiple depends satisfied <-> multiple entries in the requiredby list. No additional checks needed.
2) having this equivalence gives more information. in this case we know kernel26 satisfies 2 deps of atl2 and 2 deps of nvidia.

The problem in that in the case of multiple versioned depends, it looks rather stupid.
We were rather thinking about the case of multiple provisions where this info might be more interesting.
In both cases, we don't know the version or the provision satisfied though, just the package name :/

To sum up, here are the possibilities :
1) keep the current behavior, this situation is not that common anyway. and it is an additional info for users who know what it means
2) filter the list. downsides : very minor additional cost, and very minor loss of information, advantages : less confusing for most users
3) store more than just the package name in requiredby list. but that would complicate the code

(sorry for that long comment for such a minor issue, I don't really care. I just thought I could share what I remembered)

I sent a patch to ML (http://www.archlinux.org/pipermail/pacman-dev/2008-April/011609.html), which removes compute_requiredby usage from back-end. So it is possible to make it front-end friendly. But I also think that this is a minor issue ;-)