The current wave of AI hype demonstrates of the problem of capacity creep. In some measurable regards, hardware keeps getting better, right? More powerful, more efficient, more affordable. But rather than letting those gains ameliorate the problems created by tech, tech companies see them as increased capacity, and look to fill that capacity. So while hardware improvements should be reducing the emissions caused by computing, their growing as companies look for ways to turn capacity into profit.
It follows that assurances to the effect that the AI industry will eventually find ways to shrink the carbon footprint of their technology are entirely untrustworthy. Any improvements that might lower the emissions caused by AI processing are bound to be treated by the industry as excess capacity, which they will promptly find ways to fill, excusing the increasing size of the footprint the same way they excuse it now: by promising that they'll find ways to shrink it in the future.