I guess that's solvable by requiring payment for verification runs and just splitting up your work into much smaller pieces.
The authors already thought about how to incentivize verification runs. It's really cool. Check out this informative podcast with some of the team members: https://youtu.be/QY0OUTaIWIc
There's an active downside of random people being able to see your data, not to mention randomly sized boxes and shoddy network connections (which is a big deal for the microservices which they are touting as a use case, and also for big data. Data transfer is one of the slowest bits of data science). Running a web service on random devices where keys, sessions, whatever end up in memory on some box would be a security nightmare. The government doesn't need a back door if all they need to do is host some compute nodes.
Is it just that for some tasks that are very parallel and not sensitive it might end up cheaper?
Perhaps a digitally signed NDA could be verified by a smart contract? Then, to prove the NDA was breached, a quick court case (standardized NDA) leads to a published ruling that the digital contract can check.
Combine the clear guidelines of a standard NDA with a proper court adjudicating the more complex disputes, and you have the foundation for a strong reputation system for Eth-cloud services. Is reputation already a thing in the Ethereum network?
The network bandwidth will be a real problem regardless. Latency too. Single digit millisecond scale database roundtrips are awesome. 50-100ms would be just terrible. If that's a fundamental limitation then applications can only get worse, not really better.
Overall, what fundamental problem are these solving? If they aren't solving a real pain point with a worthwhile cost, uptake is going to be nonexistent.