Because when Pytorch came out, Tensorflow API was cumbersome and hard to work with. In comparison, Pytorch was much more intuitive. Being beginner friendly, it quickly gained traction. It was like Java vs Python :)

At the time, except for advanced use cases or prodution usage, most people used Keras, the simpler higher level API on top of Tensorflow. Until Pytorch came out, which supported both simple and complex use cases, with a simple API.

I heard that tensorflow made considerable changes to its API after that, but I didn't have a chance to work with either of them in the past two years. If I had to choose now, I'd choose Pytorch as I know it got its design right from the beginning.