CUDA is dying other than hobbyist ML. The big cloud players are all moving to ASIC (e.g. Google TPU). CUDA was a half-decent interim step but making a GPU do ML/AI tasks won't compete with purpose built silicon.
That's true for bulk Google-scale production, not for developmental and research work. You can't afford to spin an new ASIC, or even learn to code for a proprietary ASIC, to explore new ML methods.
You wouldn't do production ML on a desktop computer either. You do development work. Enough stuff to then deploy it on a cluster.
There's also a ton of other scientific computing on CUDA that's not ML.