Nobody relies on it. But if we can make an attack 10 times more expensive (because we can't prevent it completely), why not do it?
Secure remote computation is better than nothing. I'd start having problem with it when people rely on it when there are better solutions, but that hasn't been the case so far, AFAICT.
If I understand correctly, Intel could (if compromised or untrustworthy):
* revoke any node's SGX keys - knocking any number of consensus nodes out at zero cost * create false attestations of what code was run (which would presumably be detectd eventually) * dominate the consensus mechanism with its hand-picked SGX nodes
Among other things, this means Intel could turn off the MobileCoin chain, if commanded to do so.
(I could be somewhat off on some of this – if so, corrections with explanations very welcome.)
So, if your threat model is "Intel Inc is always trustworthy & secure", Signal & MobileCoin are wonderful. If you have doubts about Intel's trustworthiness/security, the designed privacy properties start to degrade.