The explosive growth in power demand projections from AI data centers has triggered a massive wave of generation and transmission investment, but the industry has largely failed to model the downside scenario where AI compute demand plateaus or shifts toward efficiency, leaving utilities and ratepayers exposed to stranded assets and underutilized infrastructure. If the current build-out assumes a hockey-stick demand curve that flattens — whether from algorithmic efficiency, economic headwinds, or regulatory pushback — the resulting overcapacity could lock in higher electricity costs for decades while carbon-reduction goals stall on idle gas plants and underused renewables.
Grid planners and regulators are accustomed to forecasting errors, but the AI-driven load forecasts now circulating carry a unique asymmetry: they are being used to justify long-lived capital commitments — combined-cycle gas turbines, high-voltage transmission corridors, dedicated renewable PPAs — on the basis of demand signals from a handful of hyperscalers whose own capital expenditure plans remain volatile. The PJM and ERCOT interconnection queues already reflect gigawatts of projects anchored to speculative data-center loads, and history suggests that when industrial demand forecasts overshoot, the costs socialize across the rate base while the benefits privatize to the developers.
Efficiency gains in model architecture, inference optimization, and hardware specialization could dramatically reduce compute-per-watt ratios faster than new generation comes online. Meanwhile, geopolitical constraints on chip supply, rising capital costs, and potential saturation of near-term AI applications all introduce demand-side risk that current resource plans treat as negligible. A rational planning framework would require probabilistic scenarios, not single-trajectory assumptions, and would price the option value of modular, shorter-lead-time resources — batteries, demand response, distributed generation — over monolithic baseload commitments.
Regulators in Virginia, Texas, and Ohio are beginning to ask harder questions about cost allocation and exit fees, but the institutional inertia favors approval. The industry needs a structured “stress test” for AI load forecasts, similar to banking capital adequacy exercises, to quantify the ratepayer impact of a 30–50 percent demand shortfall. Without it, the energy transition risks building the wrong infrastructure at the wrong scale, financed by customers who never asked for it.
Read the full report at CleanTechnica.