Texas froze every new data center license, and the queue explains why. In 2024, the state had 63 gigawatts of large-load requests in the grid queue. By June, 474, about ninety percent of that data centers, more than five times the state’s peak consumption record.
The new rules, in effect since October 8, turned access into a deposit: a 100,000-dollar study fee plus 50,000 per megawatt. A 1-gigawatt campus deposits 50 million. Losing capacity forfeits 20% of the deposit, at minimum 10 million.
The number that dismantles comfortable planning
And connecting a large data center takes five to ten years.
Seventeen projects of 75 megawatts or more had completed every other step and remain stuck until December’s audit report. These are projects that did everything right and keep waiting: the queue does not judge merit, it judges arrival order against a physical bottleneck.
The engineering figure that closes the math: keeping a 1 GW campus running through a 14-hour winter night requires 14 gigawatt-hours of batteries, about half of the ERCOT grid’s total storage. Energy autonomy at campus scale is a multi-year project.
The FinOps thesis at grid scale
Some operators have started designing for 99% availability, about 88 hours of downtime a year, because idle hardware costs more than the energy to run it.
This is the thesis I have repeated for months, elevated to the scale of a state’s grid: the bottleneck stopped being the kilowatt-hour price and became time-to-power. Whoever waits five years for interconnection is not saving: they are paying the project’s most expensive bill, the bought-and-idle asset. It is the same arithmetic as the GPU commitment that became a liability: an asset without measured use is contracted waste.
The variable that changed in the location spreadsheet
For anyone designing capacity, the question stopped being how much the megawatt-hour costs. It is how many megawatts are in the queue ahead of you.
And that is where the third-power thesis gains its regulatory chapter: while Texas freezes and charges deposits, Brazil created a competitive transmission-access auction, with price and a winner, the same week. An administrative queue on one side, a bidding process on the other. Workloads serving Brazilians already had a reason to live here; global load just gained one more.
Conclusion
Texas’s brake is not anti-AI: it is grid physics arriving before AI’s marketing. Forty-seven gigawatts of promises do not fit in a cable, and the regulator’s answer was turning waiting into a price.
For designers, the new ruler is the queue: how many megawatts ahead, how many years to power, what the idle asset costs. And for anyone hosting Brazilian workloads, the same ruler exists in a favorable version: capacity that is bought, energy that is spare, and a distance nobody waits for.