In the United States, AI systems currently generate around 125 tokens per person per day against an energy ceiling closer to 44,500. This gap acccording to Alec Litowitz, Nick Polson and Vadim Sokolov read as evidence that, at 2024 infrastructure levels its hardware deployment that binds American AI output, not electricity (Litowitz et al., 2026). Their model projects that by 2028, US AI electricity allocation of 326 TWh could support closer to 225,000 tokens per person per day, at which point the constraint stops being how many chips exist and starts being the kind power that feeds them.
This process of reallocation of energy towards AI deployment and training isn not however without consquences. Electricity costs at vertically integrated AI companies already doubled as a share of expenses between 2019 and 2023, going from 0.8% to 1.6%, and pure data centre operators now carry a share of around 15% when it comes to energy costs. Christian Bogmans findings contribute to this trend, showing that AI-driven demand produces manageable but uneven effects on prices and emissions. The outcomes also hinges on whether renewable capacity and grid investment expand at historical rates or are held below them (Bogmans et al., 2026). Adding renewable capacity would however not settle the problem. Luyi Gui and Tinglong Dai's equilibrium model of AI development, finds that when the market rewards capability scaling, developers push towards the frontier regardless of whether the marginal megawatt-hour is clean or fossil (Gui & Dai, 2026), an outcome they term an adaptation trap. Renewable expansion seems thus to mainly relaxes the scaling constraint rather than truly displacing fossil generation. Mohammad Hemmati, Gbemi Oluleye and Vassilis Charitopoulos show the geographical nature of the trap. Their spatially explicit modelling of European data centres projects 73 to 723 TWh of additional demand and 67 to 181 megatonnes of cumulative emissions overshoot by 2050 across 21 growth scenarios (Hemmati et al., 2026). Past 2030, they find that where AI infrastructure gets built will hinge less on how much power a region has than on whether its grid can deliver firm, flexible capacity on demand.
Eshta Bhardwaj, Rohan Alexander and Christoph Becker's systems-dynamics account of AI scaling names the pattern these findings fall into. Efforts to overcome one limit to growth tend to displace the problem rather than resolve it (Bhardwaj et al., 2025). Read together, these five studies describe something more specific than an energy crunch. AI's expansion is being sorted not by how much power exists but by its reliability and location, and access to frontier AI will increasingly depend on access to reliable capacity rather than the other way around.