Okay reading, it sounds like before Burnham and Anderson (2002) it was only known to be valid for nested models, but they were able to derive it from more fundamental assumptions to remove that restriction.
https://en.wikipedia.org/wiki/Akaike_information_criterion#H...
I wasn't aware of that update until now (despite being 16 years old), thank you!!
(And Wikipedia: https://en.m.wikipedia.org/wiki/Akaike_information_criterion...)
Do you have a reference? Or were you thinking of the Likelihood Ratio Test?