According to Morgan Stanley, on July 27, the firm quantified generative AI investment returns for the first time, constructing three valuation frameworks. Self-operated infrastructure model API business achieved return on invested capital (ROIC) of 46%, hyperscaler GPU rental services returned 31%, and third-party compute API services returned 25%. Under base-case assumptions, a single 1-gigawatt data center requires approximately 410,000 Nvidia GB300 GPUs at 75% utilization with hourly rental prices of $8.50.
Combined capital expenditure from three major cloud providers is projected to exceed $1.4 trillion. Morgan Stanley maintained overweight ratings on Microsoft, Amazon, Meta, and Google, citing sustained GPU demand growth as AI transitions from training to inference phases, positioning self-built compute operators to capture substantial margins through pricing power in the GPU-constrained ecosystem.