Alexander Crosier, Kyle Onghai, Ronnie Sircar
Under plausible ERCOT-calibrated growth rates, data-center electricity demand will raise wholesale prices materially unless new generation is built faster than decentralized profit incentives induce, because each new plant cannibalizes existing capacity revenue.
A stylized market-clearing model of the ERCOT grid (reference price $30/MWh, 55 GW traditional and 8 GW data-center demand) with price-responsive demand curves, deterministic and Poisson-jump growth scenarios, and a stochastic-control (HJB) formulation of a revenue-maximizing generation investor, calibrated to published Texas/PJM figures over a 6-year horizon.
With zero supply growth, modeled prices double from $30 to $60/MWh within 10 years and would take over 22 years to reach the traditional consumer choke price of $70/MWh; supply growth of roughly 3 to 12 GW/year determines whether prices rise or fall, with cannibalization able to stop optimal investment even while demand keeps growing.
The model is stylized and single-region: it ignores retirements, intermittency, outages, transmission constraints, and short-run price dynamics, treats demand as two aggregate groups with hand-chosen elasticity and choke prices ($70 and $150/MWh), and assumes increasing supply; parameters are inferred from public forecasts rather than estimated, and the demand forecasts themselves (e.g., ERCOT's 33 GW by 2031) are highly uncertain.
The paper is a theoretical exercise fitted to a single grid with unvalidated elasticity assumptions; the explosive load-forecast numbers may double-count speculative interconnection requests, and results are sensitive to chosen choke prices and growth rates, so the claimed underinvestment result could largely be an artifact of parameterization rather than a real market prediction.
Little or nothing yet for AI economics directly; the paper's value is framing a real possibility that power scarcity and slow generation entry become binding cost constraints on data-center expansion within the next 12 to 24 months, raising both AI compute costs and retail electricity prices in affected regions.
Market & Valuation: Investors in hyperscalers and utilities should note that power availability and capacity-price cannibalization, not chip supply, may be the binding constraint on near-term AI capacity deployment.
Regulatory & Legal: Policymakers face a mechanism where high prices fail to induce timely generation entry, implying that market prices alone may not prevent consumer bill increases or reliability risks from data-center load.