A Pure Infrastructure, Power and Cooling Portfolio for the Artificial-Intelligence Electricity Build-Out

The International Energy Agency projects global data-centre electricity consumption will reach approximately 945 TW h by 2030, more than double the 2024 level, with roughly 55 % of the incremental load attributable to servers used for artificial-intelligence (AI) workloads. Equity strategies built around this thesis have to date been dominated by semiconductor and AI-accelerator names whose returns are driven primarily by the global chip cycle. We construct and test a public-equity portfolio that deliberately excludes every semiconductor, foundry, wafer-fab-equipment, memory and AI-server manufacturer in order to isolate the electricity-supply, grid-hardware, power-management, thermal-management and digital-real-estate leg of the AI value chain. From an investable universe of 25 listed companies, a rules-based composite score (40 % thesis exposure, 35 % fundamental quality, 25 % valuation) selects 15 holdings with a 10 % weight cap. Using daily total-return data from 2 January 2020 to 24 April 2026, the strategy delivers an annualised return of 44.2 %, volatility of 25.2 %, a Sharpe ratio of 1.76, a maximum drawdown of −31.2 %, and a cumulative multiple of 8.7× the initial capital. A Capital-Asset-Pricing-Model regression against the S&P 500 yields an annualised Jensen’s alpha of 22.6 % (t = 3.42, β = 0.95, R2 = 0.58). Alpha remains positive and statistically significant against the Utilities sector ETF (XLU, α = 30.2 %, t = 3.62) and against a broad U.S. infrastructure benchmark (IFRA, α = 24.2 %, t = 3.45). The strategy exhibits a Herfindahl concentration well below the unconcentrated threshold of 1 500. The results are consistent with the hypothesis that a meaningful fraction of the AI-energy thesis can be harvested through electrons, grid hardware and cooling alone, without exposure to the compute side of the capital stack.