"""For a Google project, this was somewhere between "space aliens" and "narnia" on the Scale of Surprising Things."""
What you need to understand is that the people who are doing Jax and TFRC are a group of excited brilliant engineers on the cusp of junior/senior status who have a complete understanding of the current state of ML hardware, and have built a collection of powerful tools that are far simpler than TensorFlow. It's still in the rapid expansion phase, new ideas can be tried out quickly, and engaging with the developer community enables a virtuous cycle that improves the system rapidly.
TensorFlow on the other hand has turned into a complex software behemoth where nearly everything requires extraordinarily long software development efforts to make minor changes, and any changes in the core impact thousands of users.
What I learned from interacting with TFRC and the Jax community is that the real secret value is in XLA and the runtimes that wrap around it. Once you've mastered those, you're well on the way to using large resources like a TPU with a CPU (I think that's just the host CPU and RAM, BTW) and harnessing all the compute elements simultaneously to get peak performance.
I expect that as more people see it's easier to program a simple layer on top of TFRC and XLA, tensorflow will not be as popular.