Implied here is the emergence of a new business model: developer powerful custom hardware that you do not sell, but only make available as a service in the cloud. This way you get multiple layers of lock-in.
I think the main reason to push TensorFlow as open source was to be able sell the TPU cloud without a total lock-in.
I suspect it's the opposite. TensorFlow in source will never be as fast as the TPU cloud with custom processors, so it won't be as suitable for serious high volume production stuff. It's more like a hook into the TPU ecosystem.
If TensorFlow is open source then there's nothing stopping AMD, NVIDIA, Intel, etc. from making chips that can integrate with TensorFlow. Other than the fact that it's hard, of course.
Is it though? Facebook has been pushing hard for open hardware and open datacenter designs. I'm glad they remained an independent company if only to act as a counterweight to Google's dominance in developing customized hardware.
Agreed. NVDA is probably very concerned right now.