A serious bear case on AI infrastructure should survive in every world where models keep improving. Merryn’s dies the moment GPT-6 ships something visibly better than GPT-5. If that happens this year, every legitimate concern about the capital stack underneath this buildout will get tarred with the same brush as “the LLMs are stupid and it was all a bubble.” That is the cost of letting the wrong bear case become the consensus bear case, and it is why this column is worth picking apart.