“"Vertical" inference time scaling, which involves generating longer chains of thought, requires systems with large amounts of high-bandwidth memory (HBM) and high interconnectivity, such as NVIDIA's NVL72.”