Not a way to prevent it but they can make damn sure its difficult to find.
Example: Where is Newtonsoft's Physics Library v1.x? It was awesome, easy, fast, and worked with my engine (or rather, my engine worked with it?). Gone. Nowhere to be found, not even on the internet archives way-back-machine.
It's rather trivial for a juggernaut like NVidia to wipe the earth of older cuda compilers by tweaking a driver and making cuda compilation cloud-based.
Wouldn't C# be the most reasonable option if so? Maybe they did the math and the .NET ecosystem is much more expensive than Java? Even ignoring MS cost, most 3rd party libraries on .NET usually require a license.
I dont know their needs, but last time I tried to do a computer-vision project on C#, all libraries in Nugets where shitty wrappers for cloud solutions or just paid libraries (last update to OpenCV is 7 years old).
Using bindings to C++ libraries is likely to yield better experience. C# has really good interop API (P/Invoke).
I don't have experience with machine vision but here's an example of a speech2text library that integrates whisper.cpp in an idiomatic way: https://github.com/sandrohanea/whisper.net
There are many good community projects with code quality way higher than your average enterprise SDK with layers upon layers of abstractions and allocations. Finding them is the same as with most other languages like Rust or TS.
Personally, it always run fine at the beginning until it isnt and then you have to spend the rest of the day trying to figure out what is wrong. For desktop that is it, servers can just run 24/7 without problems (mine just run containers).
Plasma looks pretty good but is not nowhere as good as window's.
I heard that they fixed power management but haven't tried myself.
My last point is that hyper-v got soo good with GPU partitioning and can be fully managed using type-safe Powershell.
