About a decade ago, one of our customers had a 25Gbit (IIRC) line that went cross country and had something like 150ms of latency. Then after hours they would basically cross replicate their entire data center through a couple machines they purchased from us (running some of our software). While their packet loss was fairly trivial, it was enough to cause some serious pain. Explaining why, if they insisted on TCP they were going to need an enormous amount of ram (for the time) in order to keep the thing running at line rate took about 6 months... Eventually someone ponied up the $$$$ for the ram, and the problem went away, and it sat there happily doing its thing, with only the occasional problems caused by other departments trying their own experiments and killing its throughput.