I can't be the only one doing ML development this way.

I do have a Windows machine with an nVidia GPU (RTX 2070 Super), but I don't use it anymore (bought it years ago).

ML on the cloud is way more convenient because you can trivially adjust your cost based on what you're doing: Training? spin up something big/expensive. Inference? cheaper (less VRAM) is usually fine.

I also like that I can run multiple instances simultaneously, something that would be prohibitively expensive if I had to have multiple machines sitting around waiting for me to use them.

The M1/M2 is brilliant for the form factors that consist of about 90% of PC demand. Macs are better as a device than almost any PC on a few axes.

But… an architecture using an SoC to scale from an iPhone to a large personal server, which is what a modern era Workstation really is will hit limits.

IMO, the failure was they were boxed into dropping MacOS Intel support in the near future, and for whatever reason were unable to ship the Apple Silicon they intended to. So they shipped a compromised Mac Pro that meets the needs of a few key use cases that can sorta work with the limited top line memory.

Intel segmented their products because they had the similar constraints. There’s a mondo-expensive 8 core 4 GHz+ part, and a part with lots of cores at half the speed. In Intels case, the memory is of course not part of the SoC.