I think this is going in the right direction. You take a bunch of tamper-proof hardware devices from different manufacturers and then model attack costs to compromise them as part of a proof-of-stake scheme. Now you can build consensus algorithms on top of them that are highly secure and scalable compared to anything that exists today.
I'm not convinced that Stellar consensus here is the right algorithm for doing this, but I think SGX is promising technology that has been somewhat overlooked in the blockchain space (not by everyone.) SGX has a lot of potential. You can use SGX as a way to expand the consensus rules of any blockchain by using it as a blackbox obfuscation construct. Everything and more that Vitalik wrote in his article about Indistinguishability Obfuscation is possible with SGX today.
Want to create a specialized oracle that only signs certain transaction formats, even on untrusted hosts? Yep - use SGX. Now you can have agents that run in a cluster that will only move assets between blockchains based on a user's prior agreements, allowing for more complex cross-blockchain smart contracts to be written in high-level languages. What about having a nice way to do transaction commitments to scale any blockchain without having GB zero-knowledge proofs? SGX again. It could be used for privacy preserving protocols... It could be used for solving data availability problems in sharding / decentralized storage systems. The list goes on.
Some of the biggest trust problems are solvable with this technology - but like others have already said - you still have to trust the hardware manufacturer. In this case, my thoughts are that you already have to trust the hardware manufacturer anyway (nobody is going to inspect every chip with an electron microscope...) My bet is that a non-trivial portion of full nodes today are already running chips with backdoors like the Intel Management Engine anyway...
The point here is that you can't fully remove trust from any system without introducing vast inefficiencies, but you can at least formalize the risks in a system and design so that a compromise is too expensive to be worthwhile, and for me I think that's where the potential lies with this tech. Cryptoeconomic systems based on tamper-proof hardware where individually a component may be compromised, but where it is simply infeasible to compromise each and every device. You build a network out of these components and you have yourself the first on-chain scalable blockchain bound by physical hardware encumberments instead of computational difficulty.