Quite a few others like Gevent and supervisor, for example, have py3k ports nearly done, and mostly await merging.
I will try to be a good boy and use it going forward (I already find it more useful for CSV and other file I/O type tasks thanks to the unicode support, but am reliant on mysql for some routine drudgery as well).
edit: changed grammar for clarity
Your average interactive webapp is reasonably query-heavy, some of which is alleviated by reasonable caching.
The performance of the connector comes into play when someone hits an uncached page which requires tens of queries. This matters the most when it drastically affects load time. It's really a shitty experience when a user comes to a page, waits more than a second for the page to load. I do find, though, that a large portion of my query overhead comes from round trip time to the db server, and not intermediate processing.
So I figure for low to medium traffic sites with reasonable caching (and cache priming when needed), this probably won't be a bit hit, like you say.
There is of course the interesting benefit that if you're using a JIT implementation like pypy, then everything within a pure python db adapter is up for auto-optimization. This is interesting because it could effectively provide something close to application-specific hand optimization of the connector, which tends to be a pain in the ass that not many people would bother with.