What does Xeon have to do with training? You need a fast GPU (or apparently TPU). The CPU is relatively unimportant. Furthermore, you can roll your own algorithms (i.e. architecture) in TensorFlow. You don't have to use "theirs".
Now, you may think I'm crazy for designing something from scratch, I may be indeed, but how else can we discover/invent something totally new. I think I'm actually onto something given that the early results look quite promising.
I actually created a video demonstration of my AI Engine which you can find a link to it in my profile but I have done a crappy job of explaining the strengths of it [1]. Too much work, so little time.
Where I get stuck is figuring out new challenges to throw at the system. I find it funny that there is lots of discussion here about performance and tuning, and not so much about practice applications and valuable problems to solve.