The Physical Limits of Compute: Dynamic Risk Contagion Networks Between Semiconductor and Energy Infrastructure Assets Under the AI Super Cycle

Abstract

Artificial intelligence depends on a production chain connecting semiconductors, data centres, electrical equipment, generation, nuclear supply, cooling and gas infrastructure. This paper shows how that physical chain becomes a financial risk network. We combine daily adjusted equity prices through June 2026 with published electricity-capacity outcomes, power agreements and a regulatory access decision. A seven-layer system uses 28 continuously represented stocks and an unbalanced extension containing five additional infrastructure firms. Factor exposures are estimated using strictly lagged rolling regressions. Between the compute-expansion and infrastructure phases, return connectedness rises from 6.34% to 21.72%, and connectedness in transformed standardised downside innovations rises from 5.27% to 22.12%. Both changes survive the three-outcome multiplicity correction and remain significant across beta windows, control sets, block lengths, layer deletions, extreme-date deletions and firm membership. Boundary analysis locates the transition within 2024: the downside increase holds for every boundary from January 2024 through January 2025, and the return-network increase is strongest around mid-2024. The observed 15.38-point return-network increase exceeds 97.3% of 5,000 size-preserving random layer assignments, while the seven-layer diversification ratio falls by 17.8%. The network peaks during the 2025 infrastructure build-out and remains significantly above the compute-expansion benchmark in the final 252 sessions. Grid equipment has the largest outgoing variance attribution. Competitive generators outperform regulated utilities by 12.18 percentage points around the Crane agreement, a contrast that remains significant against its pre-event placebo distribution after Holm correction across ten source-dated events. Most measured connectedness is already present at the one-step horizon. The evidence establishes a stronger, uneven financial risk network spanning the physical inputs to AI, with rapid joint repricing as its dominant daily manifestation.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.