The real AI risk for energy may be moving too slowly, not too fast
The anxiety around artificial intelligence and power has, until now, run in one direction: how many gigawatts the data-centre build-out will demand, and whether ageing grids can carry it. A recent argument relayed via OilPrice.com turns that framing on its head. Drawing on a note from the law firm Duane Morris, it contends that for an industry responsible for critical infrastructure, the more consequential long-term risk is not deploying AI too aggressively but failing to use it enough, in forecasting, grid balancing, predictive maintenance and materials discovery.
The counter-case is equally pointed. A 2025 MIT review cautioned that the efficiency gains so often promised have yet to materialise, even as approvals for new data centres accelerate. Put plainly, the optimistic claim that AI will ultimately save more energy than it consumes remains a hypothesis, not a result. The proven applications are narrower but real: better supply-and-demand forecasting for renewables, faster screening of fusion and battery chemistries, and tighter industrial process control.
The broader takeaway is to separate the two stories when planning. AI-driven efficiency is a credible medium-term lever for sites with complex loads, cold chains in agriculture, HVAC in hospitality, process heat in industry, continuous demand in healthcare, but it is not yet a reason to assume system-wide demand will flatten. The prudent base case treats incremental data-centre load as a firming pressure on forward power prices, while treating site-level AI optimisation as an opportunity to be tested on its own merits rather than a guaranteed offset. Crucially, anyone modelling long-term decarbonisation must account for the AI gold rush diverting venture capital away from next-generation clean energy research. Furthermore, as Big Tech leans on natural gas to power its immediate data-centre expansion while funding longer-term clean tech ambitions, base fossil fuel prices will likely remain firmer than standard transition models suggest.