Your answer is good, but the main question here is: “Can you do better than Nvidia for your particular use case?” The answer depends entirely on what your use case is.
If the physics are reasonably simple, and the main goal is to make it deterministic across client/server and across CPU architectures, then it’s very possible to do better than Nvidia.
If the goal is “do exactly the same thing as PhysX does, but faster/better, on your own” then, well, no.
Btw, I remember when Nvidia bought NovodeX to make PhysX. That code is more like 20 years old by now 