One of the problems with ANN/ML is that compared to other algorithms, their results can be improved by adding more electronics and electricity. Since those things are undercosted due to the broken economics of market capitalism, that means that it’s “cheaper” to throw more robots at it than to actually think of a more efficient solution.
Docker and containerization has the same problem.
“Just throw another machine at it” becomes the mantra because, thanks to the weird memory-leaky processes of economic transactions and labor, and how bad market capitalism is at valuating the true costs of things, it’s “cheaper” to run something on 40 servers that with optimization work could be run off of one Raspberry Pi.
Yeah, yeah, we want some way to handle requests that spill over beyond what one machine can handle, but we shouldn’t neglect what local optimization can do. We went to the moon with the computing equivalent of one Game Boy.