Here are a few reasons:
1. Continuum analytics doesn't have a real business model and someday it will go away.
2. In the bad old days, before wheels, there really was a problem integrating C libraries with Python. Wheels work basically the same way conda packages do, so there is no problem. You can get an "official intel MKL" numpy, the only trouble is that (officially) the python packaging system has no way of saying "these N packages all satisfy the same dependency and you can install the one that is best for you".
3. "Testing together" is a wild goose chase because the combinatorical explosion of different libraries you could test together is practically unlimited.
4. Anaconda for a long time has been unable to package a "tensorflow-with-gpu-that-just-works" because NVIDIA won't let them. Without that, the window is broken and Anaconda is not giving you a problem free experience.
5. Numerous things about the way Anaconda is implemented make it slower than it has to be by a lot.
6. Since Anaconda doesn't confront all of the problems I've mentioned above you will still need to implement stabilizers to stabilize it, and if you're going to do that you can just build on wheels and have fewer things to understand, have go wrong, etc.