From what I understand, the RAM shortage is more about AI inference than AI training. Yes, training created much of the early HBM crunch because frontier-model training clusters need tons of HBM near GPUs, but inference is what is keeping the pressure on now and into the future.