Morgan Stanley: NVIDIA Racks Deliver Huge Profits, AMD Stuck in the Red

Morgan Stanley has dropped a bombshell analysis on the economics of next-gen GPUs for AI data centers.

According to the bank’s latest research, NVIDIA’s GB200 NVL72 racks are not only expensive up front but also the only option that makes sense for large-scale AI factories when it comes to profitability.

Each NVL72 rack, priced around $3.1 million, houses 72 B200 GPUs and 36 Grace CPUs, all linked with NVIDIA’s ultra-fast NVLink 5. While the initial investment dwarfs the roughly $190,000 cost of an H100 rack, the long-term payoff appears unmatched. Morgan Stanley estimates that a 100MW AI facility built with NVL72 racks can deliver a staggering 77.6% profit margin. For comparison, Google’s TPU v6e pods come close at 74.9%, though rental costs for those pods are typically 40–50% lower. The pricing for Google’s pods remains private, but the economics still look competitive.

The real shocker lies in AMD’s position. Data centers running on AMD’s MI300 or MI355X platforms, according to the report, actually operate at a loss – posting margins of -28.2% and -64% respectively. This paints a grim picture for AMD’s AI hardware strategy, especially given the billions poured into catching up with NVIDIA.

Morgan Stanley’s model assumes $660 million in infrastructure costs for a 100MW data center (depreciated over 10 years), with GPU hardware costs ranging from $367 million to $2.27 billion depending on configuration. Add in four years of GPU depreciation and power/cooling operating costs, and NVIDIA’s NVL72 racks still stand tall despite having the highest total cost of ownership ($806.6M versus AMD’s $774.1M). The sheer revenue potential wipes out the higher TCO, leaving NVIDIA alone in the profit zone.

The numbers highlight a widening gap: NVIDIA is pulling in record profits while AMD bleeds cash trying to keep pace. For investors and industry watchers, this isn’t just about hardware – it’s about survival in the AI arms race.