Rather a few thousands at least. Most larger companies have that amount of data (if they can use the model is another question).
Yep, but running inference on it at any reasonable performance requires you to have all of it in GPU RAM - Ie. you need a cluster of ~100 high performance GPU's.
The largest version of GPT-3 is 175B parameters, which is ~350GB. I frequently use 8x and 10x RTX8000 boxes (and can access a 16x JBOG system as well), and the 8x system would have 384GB of VRAM. These sort of systems start off at only ~$60k.