Ah, this is a bit of a false one. (A)GPL's pretty great for things that look like applications (or alternatively, pretty much anything GNU's ever made), but for anything looking like a library that is designed to be both free as in speech and free as in beer (Qt is the former but not the latter - want commercial usage? Pay us for a differently licensed edition!) it's a complete non-starter because it says that any usage of a library is a derivative work and infects. This is always going to be poisonous to any commercial use, and PyTorch and Tensorflow only gained the traction they have because they could be used commercially - they wouldn't have been successful if GPL'd.

Even LGPL is super challenging here due to requiring dynamic linking. This is fine in Java (unless you use Graal AOT compilation) and Python, impossible in Go, and at best discouraged in other compiled languages like Rust and C++.

I think there's likely a copyleft middle ground - something like the Mozilla public license seems about right, where it copylefts improvements without infecting users. I don't know why Google chose Apache 2.0 over MPL or the like - for something like Android it makes sense, for something like Guava it's not clearcut. But, when people talk copyleft, they're rarely, rarely talking MPL.

I think it's unlikely that Google opensourced Tensorflow with a view to getting some free dev hours, they're doing it to push their OSS agenda, to wield influence over how everyone else does work.