In my experience in telecom industry, MTBF calculations founded on base temperature failure rate plus expected average operating temperature did NOT correlate (at all) to actual field failure rates.
In my test labs, thermal cycling of equipment was dramatically more likely to precipitate defects over time than was soak at high temp.
Based on that, I would expect always on is probably slightly better.
EDIT: After that, I joined a startup offering connectivity, timing analysis and signal integrity analysis to OEM’s. I would say from that experience that timing marginalities caused by the lack of both a timing analysis and a signal integrity expert constraining layout are far more likely to cause temperature related problems than outright component failures.
A more interesting statement would be that I believe the best performing signal processing devices measured by Amir likely did have these analyses done, in order to hit the performance level they achieved.