Decades ago we had (switch mode) power supply problems on the new fangled Apple II computers.
Turns out there was an electrolytic coupling capacitor that provided PWM feedback to the power stage through a transformer (rather than an optocoupler).
The capacitor was forward biassed under light load, but we were a research lab and we filled all the available slots with cards. But under this heavy load condition, the coupling capacitor was reverse biassed and would eventually depolarise.
Swapped the cap for a bipolar type and it fixed the problem.