I have a containerized set of headless unity processes that take a map selection from a variety of public and private sources plus user data and compile it into a detailed modeled environments to be later served in WebGL, mobile, desktop, and 3D printed. Doing it in the users client would involve a lot of memory management and storage on sometimes restrictive devices, also unity can hand things off to numpy/scipy, ffmpeg, etc easier on my server than it can your phone or desktop.
There is also a huge push from Nvidia[1] specifically and others to build new tools for remote GPU use taking the load off of thinner and thinner clients doing more and more things. OSX High Sierra offers remote GPU usage in Metal2. Connecting all these vector processors will frequently be for data ETL, but rendering that data is going to fall onto tools like Unity a lot, which will also be sharing in the role of assembly and iteration.
With the new new embedded machine learning chipsets in every new gizmo, tools like Unity and Unreal are about to be ubiquitous doing all sorts of heavy lifting. There isn't quite a written playbook for what's going on, but it's pretty crazy and exciting. Almost like it was 1996 internet again.
[1] https://devblogs.nvidia.com/parallelforall/goai-open-gpu-acc...
Much of the process is explained in this post https://medium.com/@tonywangcn/how-to-build-docker-cluster-w...