To be precise, the cassette port samples at 1000000 (more or less) ticks per second, allowing a theoretical throughput of (1000000/8) bytes/s = 125KB/s.
That's a completely theoretical limit, of course - no loader gets even close (the ROM loader being 50 bytes/s) but in theory if you could sample reliably that to work out if the flux had changed or not you could get that rate. If you wrote a routine to check for changes in flux that often and connected it to a digital input, you could get that rate if your tape routine consisted of one instruction - which of course it won't.
In reality, I believe the fastest loaders were probably running at about 300 B/s (2400 baud) but that's just a guess from what I've observed.