Agree with almost everything.

> if you want to run physics serverside, it’ll probably need to run on things that aren’t windows

On the other hand, you’re in control of hardware. For some kinds of projects, CUDA saves lots of time compared to DirectCompute or OpenCL: better libraries (these hand-optimized BLAS, FFT, etc.), better runtime i.e. simpler CPU-GPU interop, and better language, CUDA is relatively modern C++, with templates and constexpr. Unlike graphics, CUDA runs fine on Linux.

I’m not an expert in CUDA but I’ve completed a couple of projects where my clients needed to compute some GPGPU stuff on their servers, or on other people’s servers they’ve rented.