I used PyTorch since it came out, and at the time it outperformed TF1 on computer vision workloads quite handily, in addition to being much easier to work with. I literally watched the entire departments in one of the major research labs where I worked at the time switch to PyTorch in the span of a couple of months.
I'm going to go out on a limb here and suggest you didn't use PyTorch when it came out in 2016 then.
> Google had to add eager mode in TF2, ruining performance in the process
Eager mode PyTorch has worse performance than torch.compile too.