“The primary bottleneck in the AI hardware supply chain has shifted from chip packaging a year or two ago to memory currently.”