Nvidia's AI Supercomputer Costs Reveal High GPU Dependency
A recent analysis shows that approximately $4 million of the bill of materials for Nvidia’s VR200 NVL72 rack-scale AI supercomputer is allocated to its Vera Rubin GPUs, which account for roughly 70% of total costs. This data comes from an HSBC estimate and highlights how much value remains concentrated in GPU silicon despite broader industry efforts to reduce AI infrastructure spending. The remaining cost breakdown includes DRAM at 13%, HBM at 9%, NAND at 2%, NVLink, cooling, power supply, and CPU each contributing less than 5%. This insight aligns with CEO Jensen Huang’s assertion that Nvidia is selling a full-stack AI platform rather than just chips. Nvidia's data center revenue reached $89.02 billion in Q2 FY2027, up sharply from the previous year, and management expects Vera Rubin to account for about 20% of this segment in Q3. Huang also noted that supply constraints are limiting growth potential despite strong demand, with purchase orders already secured across major hyperscalers. The company anticipates a roughly 70% revenue increase in fiscal 2028 but acknowledges it may fall short due to production limitations.
Topics: AI supercomputers bill of materials (BOM) hyperscalers NVIDIA Vera Rubin GPU