I've seen this chart too. But research != Production. I bet that TensorFlow is still more commonly used in serving than Pytorch because of great tooling like TensorFlow Serving. I could be wrong though, as I'm not up to date with the latest in the Pytorch ecosystem.