Power and permission, not chips, now limit the US AI build

For three years the question about the AI build was whether there would be enough chips. The question now is whether there will be enough permission. In Texas, the largest market for new datacentre load, the outgoing governor's August pause on grid approvals has already pushed ERCOT's large-load study past April 2027, and state environmental permits for datacentres are on hold until an audit is done. On 3 November Texas elected a governor who campaigned on a moratorium 'until a legislative session that puts Texans first', and governors-elect in Ohio, Georgia and Wisconsin back pauses of their own. New York already holds permits for large datacentres for up to a year; Pennsylvania now requires them to bring their own power and pay all costs. In PJM, the largest US grid, the last capacity auction cleared at its price cap and still came in 6.8 GW short of the reliability requirement. The claim is that over the next one to two years these restraints, not demand and not silicon, set the pace of US datacentre capacity, and that electricity already energised and contracted becomes the scarcest input in the stack. The strongest objections are stated in the falsifier: most of the Texas queue is speculative paper (9.5 GW approved to energise against 446 GW applied for), builders can route around grids with on-site gas, federal policy pushes for speed, and most of the new pledges are conditional and need legislatures to act. This thesis cuts against parts of MenFem's own book, and says so: IREN and Core Scientific are bullish calls on Texas capacity, and CoreWeave's bearish call is reinforced by it. Owners of plants that already run, such as Vistra and Constellation, can gain from the same scarcity. BULLISH on the thesis, conviction MEDIUM, one year. The direction is Connor's; the conviction is on MenFem's research recommendation.