I mean that the researchers submit jobs to the data owners who run the jobs, possibly in the cloud, and return the results to the researchers. The researchers don't need direct access to the data. I, along with millions of others, have opted out of every NHS digital initiative because I don't trust them. And my instincts were right because the data was given to Palantir. But if I was given the choice I would allow my data to be used in the privacy-respecting system described above.

I honestly don't think that model works. What if you want to join data between two datasets from two providers? Who pays for the cloud? What if the researchers want to do more featurization, making snapshot views for performance?

Having looked at similar contracts (specifically UofC's CrimeLab running analysis of police data against UofC hospital xray data), the pairing is required to be done on a host such that the xray data never leaves the UofC imaging infrastructure. Contractually, at least. Not exactly sure where your questions are coming from.

yeah, that happens. the researchers hate it (the service provided is often inferior to their own in-house setup) and it's mainly to retain some level of funding for the lab, rather than benefiting the most users. Also, university infrastructure is far less secure than a cloud account run by a cloud provider.