Tensor cores don't make a GPU to be a "machine learning card".
You are absolutely correct.
The "real" machine learn cards are Google's TPU's, an ASIC.
Tensor cores are just the best we have on consumer hardware to run neural net code.
This links explains a bit how neural nets work, which in turn makes one understand why tensor cores are much better than traditional GPU compute units to run such code:
https://cloud.google.com/blog/produ...k-at-googles-first-tensor-processing-unit-tpu
