There might be the following issues with this model:
1. Wiring will be too complex. QFabric essentially externalizes the guts of
a switch. It is like telling the network admins to design and assemble a
switch themselves. Is that an attractive proposition? And I don't see what
advantage this has over normal high-density modular switches.
2. Latency and reliability will be a problem. Switch latency is approaching
zero. More wiring means slower switch. This is against the trend. The more
wiring also the less reliability.
4. If the director is just a policy management software, then it is just a
hype. Otherwise, if it functions as a centralized control plane, convergence
could be messy. For example, when link-flap happens, or peering instability
happens, all will be waiting for the director to resolve the routing table.
Significant delays may happen across the data center, and hence traffic
black holes. The central director is a single point of failure.
It seems that the purpose of QFabric is to simplify leaf-spine topologies by
replacing TOR leaf switches with TOR line cards. It's an interesting idea
though.