What the GP talked about is there for a reason. If you used UDP in a congested network, you'd probably get lots of packet loss. If you tried implementing your own retransmision algorithm, you'd probably get the same results as TCP with no congestion control algorithm: extremely low throughput, even though you have enough bandwidth.
Just don't do one thing at once and you won't suffer the penalty as hard as you would one connection at a time.
(and I'm not sure about that 40% figure, TCP RENO is probably more efficient, and you're not supposed to connect to the net with something worse than that)
And since the penalty is per connection, the ISP doesn't suffer the way you would. Their pipes probably fill to capacity (or close), provided there's demand.