If the leading AI lab can't train models that are smarter for a given level of compute, then it gets commoditized the way Ziplock bag producers are commoditized.
My point is that if there are such intelligence-efficiency differences, compute getting more expensive allows the leading model company to charge higher margins.
If one person could make a Ziploc bag that was 2x better than the next guy's Ziplog bag at preserving leftovers, they could, in fact, capture a large part of that surplus.