The first paper in this series described the AI buildout as a war bond market: capital allocated not on expected return but on the cost of losing. This paper asks the counterfactual. If there were no race — no existential framing, no strategic rivalry, no fear of becoming the next Digital Equipment Corporation — how much AI infrastructure would a rational capital allocator build?
Six parameters decide the answer: the rental yield on the asset (R₀), the cost of capital (r), the rate of economic obsolescence (δ), the elasticity of demand for compute (η), whether the buyer is captive or merchant, and the correlation of the bet across the portfolio (ρ). All of them can be bounded with market data.
The Yield on a GPU
A 100-GPU H100 cluster costs roughly $3 to 4 million at 2026 hardware prices — $25,000 to $40,000 per GPU plus networking fabric. Rented at one-year contract rates of about $2.35 per GPU-hour (SemiAnalysis, March 2026) at 75 percent utilization, it generates about $1.5 million a year: a gross yield of roughly 40 to 50 percent, depending on hardware cost. At on-demand rates near $3.40 per GPU-hour, the gross yield runs 55 to 75 percent.
Discussion