Title: Do Renewable Certificates Lack Digital Footprints? Theoretical Frameworks and Empirical Findings on the Grid Effects of AI-Driven Power Consumption

Abstract

Although data centers currently consume 4.4% of the electricity generated in the United States, the actual impact on grid stability of the renewable energy certificates (RECs) and power purchase agreements (PPAs) utilized by hyperscalers to assert carbon neutrality is not well understood. This study constructs a game-theoretic model where data center operators select between RECs, PPAs, and behind-the-meter colocation, while power generators decide on market entry based on endogenous financing costs. The analysis reveals a "timing wedge"—a discrepancy between the moments of consumption and the crediting of renewable output—as the primary driver through which artificial intelligence (AI) demand undermines reliability, drives up prices, and boosts emissions, even in scenarios where RECs fully offset annual usage. By eliminating revenue risk for generators, colocation paired with storage directly resolves this timing mismatch and fosters the highest level of renewable market entry.

To validate these theoretical predictions, we leverage the phased rollout of large language models as a natural experiment. Employing a difference-in-differences approach on a new dataset that connects AI activity with local grid metrics, we find that AI demand substantially increases fossil fuel generation, wholesale electricity prices (rising by up to 25% in specific PJM zones), and outage frequency (adding 0.5 to 1 additional outages annually) in the vicinity of data centers. These effects intensify as model size grows. Notably, data centers equipped with on-site generation show a reversal in power quality trends, aligning with the model’s forecast that behind-the-meter assets can absorb demand surges. Counterfactual simulations indicate that strategies such as edge inference, geographic redistribution, and colocated storage significantly reduce grid stress, whereas relying solely on RECs fails to do so. Ultimately, our findings highlight that the grid externalities associated with AI are deeply influenced by procurement methods and the physical layout of data center infrastructure.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC