Programmers of today have replaced customly designed algorithms with off-the-shelf heuristics. They wander through packages and libraries and end up making something that is in no way viable to run in production.
(Paraphrasing Tesla)

Seriously, how many dependencies do a data science stack for some internal process needs? It would be fine of it was just an image compilation issue, but every other library nowadays needs a fucking subscription.
It makes sense when vendors want to bill you for using their neat library. It takes a lot of effort to develop those things.
But you better deliver something great and have a fucking good reason to make it an API instead of a self hosted or locally installed library.
Because it doesn't matter how much you've optimized your hardware to run your library, the network latency makes it much more expensive (in processor time alone!) to call an API and await a gRPC response. Oh, I can do async? No shit, Sherlock. Your fucking server cannot handle my load. Not without significant investment on a dedicated host and then we're back to square one.
Shit, its cheaper to run the thing on my side. "Oh, you can do self-hosted! We just need to make sure we have our digital rights management bloatware ruining the performance of everything!"

Just to make things worse, every other data science library nowadays is just a fucking distilled LLM. Those pieces of crap manage to be more unreliable than politicians.

Fuuuuck, the world is burning already and we're pouring fuel all over it with those fucking get-rich-quick "advanced API tools". Can't we just hire talented people instead of licensing a dozen vendors? But noooooo, everybody gotta be their own fucking boss because companies hate when their employees are happy and sabotage it all with RTO mandates.

Fuck, I need a drink. The enshittyfication has come for us, at last. May our end be quick as `sudo shutdown 0`.