In practice, single desktop class machines with 6000+ concurrent users are not typical use cases. Instead, high performance applications are likely to look more like 3D rendering engines.
In practice, when you have 16ms to produce a frame, it really matters that the garbage collector doesn't kick in for 2ms once every second, because that'll push you past your frame window and lead to stuttering and dropped frames.
In practice, it really matters that you can structure your code to make sure no allocations are happening during certain critical operations, because an allocation will potentially need a new page, and the kernel barrier and/or hit locks resulting again, in 1-2ms of unexpected delay, and a dropped frame.
In practice, it really matters too that you have enough control over memory layout to guarantee that certain structures are all going to end up in cache at the same time, and that you're not going to be doing a bunch of pointer indirection fetching memory during time critical rendering code.
In practice, modern garbage collection doesn't allow you to solve any of these problems. That is why real time rendering engines are still written in C++, and will continue to be, and why everyone writing them will continue to be glad that C++ is not garbage collected.