Stress-Contingent Risk Transmission in the AI–Energy Nexus: Evidence from TVP-VAR Connectedness and Nonlinear Predictive Models

Artificial intelligence is creating closer links between computing infrastructure, electricity supply, critical minerals, nuclear capacity, and clean-energy investment. This study examines daily market-price transmission among U.S. natural gas (NG), copper futures (HG), the Global X Uranium ETF (URA), the iShares Global Clean Energy ETF (ICLN), and the PHLX Semiconductor Sector Index (SOX). Source prices cover 4 November 2010 to 31 December 2025; exact-date alignment produces 3810 common prices and 3809 daily returns. The analysis combines a discount-factor time-varying parameter vector autoregression with generalized forecast-error variance decomposition, directional connectedness, VIX-defined stress regimes, a prespecified post-November-2022 calendar comparison, and leakage-safe expanding-window forecasts. Average total connectedness is approximately 30.6%. Stationary block-bootstrap confidence intervals show that TCI is 10.01 percentage points higher in high-VIX states and 2.94 points lower after 30 November 2022. In a joint Newey–West regression, the corresponding conditional differences are 9.84 and −2.08 points. Weekly aggregation preserves the baseline transmitter–receiver ordering, and pairwise lead-lag checks show that the largest absolute correlation for every market pair is contemporaneous. Cross-market models do not produce systematic forecasting gains. Overall, the results point to an AI-energy risk network that becomes more connected under market stress, rather than to uniformly higher integration after the dated calendar cutoff or consistently exploitable predictability.