I have software on my Quad at work that I use for writing exams and things like that.
One specific software package(Chemdraw) is available for Intel computers, but it's freakishly expensive. On the other hand, I've used "scavenged from the trash" versions for a while now on PPC computers.
I have an SL Server virtual machine on my Macbook Pro with Chemdraw(along with Office 2011) installed and it's workable, but nothing beats actually using it on real hardware. The end results are the same as if I'd used something newer, but I can get it done faster and more efficiently running natively on a PPC computer.
I have some other specialized scientific software that either never made the jump to Intel(sometimes not even to OS X) or if it did is freakishly expensive.
Finally, I maintain a G4 tower at work that's operating a Nuclear Magnetic Resonance spectrometer. This is an older instrument and is highly specialized(as NMRs go). For anyone who cares, it's a solid state NMR with a 500mhz(proton) wide bore magnet to support magic angle spinning, and all the probes for it are custom-built by the PI who operates the instruments or his grad students. There's really no modern replacement that would allow him to continue his research(and even if there was, it would be a quarter million dollar or more investment).
In any case, the software for this instrument runs only on OS 9. In fact, he was told that it wouldn't run on 9.2.2, but it ended up working fine. It's enough of a customized install that I actually kept an image of it after the last hard drive crash so I wouldn't have to go through the song and dance of getting it working again. The software has to "talk" to a National Instruments PCI card, so classic mode isn't really even an option.
We've also been pushing the speed barriers on it. The 466mhz Digital Audio he was using for lost its power supply, and to just try and get things online again, I grabbed a 733mhz Quicksilver that was sitting in the room. It caused a somewhat bizarre situation-the software would send the pulse signals to the instrument(we could observe them on an oscilloscope), the instrument would respond with its signal-called a free induction decay(FID) that we could also observe on an oscilloscope, but the computer wouldn't "see" the FID. "Downgrading" the Quicksilver to 466mhz(from the dead DA) actually fixed it. I've been working on putting together a few surplus computers for him so that we can switch with minimal downtime if necessary. I've also been promised dinner at any restaurant in town of my choosing sometime soon