FWIW, a few years back, I was averaging <5ms for a benchmark of live traffic (many thousands of queries per second) on Pinterest queries against their full dataset on a single EC2 box with a weeks worth of customization on top of Lucene.
The existing stuff isn't slow.