Like Glide, CUDA took off because getting started is so much easier than getting started with OpenCL,etc, so a ton of researches could do stupid prototypes without getting bogged down in implementation details and once those prototypes took off a ton of code was tied to Cuda (seriously, look at OpenCL vs Cuda hello world or simple 1d kernel examples).
Now with higher level Python bindings being popular and US export restrictions, people has put effort into stuff like PyTorch supporting other API's, because there is enough industry pressure to have alternatives, the question is how long the CUDA moat can keep giving NVidia a hardware premium.
So, CUDA really wasn't anything they embraced or extended (even if they did so with GL,etc), it's their own API thus GP's parallell to 3dfx/Glide and the question is how long they can keep milking that moat (especially, since iroincally, AI's probably makes porting to other frameworks far easier).
CUDA stayed relevant because they stood behind it and supported it for 20 years rather than changing to a new framework every few years and only half-heartedly supporting, and because it worked on consumer cards so anyone could learn how to use it.
- Nvidia only started CUDA when the Nvidia/AMD (initially Nvidia/ATI) duopoly was already well established.
- The applications for CUDA are less consumer-oriented than the applications for 3dfx's Glide: if you want to publish a game, you care about it running on all graphics cards on the market; while the stuff CUDA was and is used for is usually done by bigger players, who don't have to care about this type of compatibility.