But there are a lot of disadvantages. The process only really works for final renders. All the creative work needs to be fast and local, but thumbnail renders don't always give you enough detail to make good decisions.
It doesn't work for audio, because audio isn't great with low-quality previews and often needs hands-on changes.
It doesn't work for many kinds of ML - not unless the render machine has the latest graphics cards.
It doesn't work for renders with huge asset libraries spread around a LAN.
And so on. Basically there is a list of use cases, but it's not even half of all possible use cases.
Meanwhile dual-Xeon servers have really started to kick ass. These are workstation grade machines with high four and low five figure prices. You can have up to 48 independent cores on a dual-CPU board, which is excellent for video rendering.
Or you can get a blade rack. Depending the project, they can be better value than an off-prem cloud account. (Sometimes).
Apple has turned its back on this space. I don't entirely understand why. Xserve had real potential, but they decided to go dumbed-down consumer instead.
use proxy videos in your project, transfer the split instructions and composites, and let the server do the rendering
no latency at all
> no latency at all
Um...
Not cheap, but...