474 GW queued, 90 GW verified, and a Decade of Grid Capex riding on this Gap
ERCOT’s interconnection queue for data centers sits at roughly 474 GW. Texas’s own record peak demand is under 90 GW. That’s not a rounding error. It’s a queue more than 5x larger than the grid it claims to forecast.
The governor now paused new hookups and ordered an audit (due December 10th), analyzing the ~300 largest requests. Seems like nobody seriously believes all 474 GW converts to real load. The open question is how much doesn’t.
It’s not just a big queue that should worry you. What really should worry you is an unverified queue.
Interconnection Requests == Free Options
I like to think about interconnection requests are free options: developers file early, everywhere, to preserve optionality, with no cost for walking away.
But utilities use these same queue numbers to justify transmission capex and rate cases. Regulators approve it. Capital gets deployed against it. Gas plants, batteries, substations.
Let me give an example. GridStor, a battery storage developer, just closed $220M for a 400MWh Arizona battery, de-risked by a 20-year tolling agreement where the utility eats the price risk and ratepayers ultimately backstop it. Is anybody pricing in what happens if the queue is 70% air?
The Toxic Asset in Finance
That’s the toxic-asset parallel. Subprime mortgages got bundled and re-rated as safe because everyone downstream trusted the number one step upstream instead of underwriting the actual asset. Nobody re-checked the loans. They checked the rating.
In case you want to read more about Toxic Assets, here the Wikipedia entry.
The Toxic Asset in the AI Buildout
In the data center game, nobody’s really re-checking the interconnection queue. They’re checking the utility’s forecast, which is checking the queue. The unvalidated queue. The free option queue.
Isn’t it the same failure mode? Risk gets laundered through a chain of parties who each have a reason not to be the one who stress-tests the underlying number. Developers want optionality. Utilities want rate base growth. Regulators want to look responsive to demand. So maybe the ghost load is the toxic asset in the AI buildout.
Learning from Texas
The Texas audit is the first forced mark-to-market. Grid analysts have already flagged that a large chunk of that 474 GW is speculative or duplicated. That means every capex plan and financing structure built on 'AI will need 2x the grid' needs repricing. Not just in Texas.
The AI Buildout is not Fake
I don’t think this means the AI-power buildout is fake. Demand growth is real and under-estimated in many markets. But treating an unverified queue number as a demand signal solid enough to underwrite billion-dollar, multi-decade infrastructure bets is exactly the kind of category error that looks fine right up until an audit forces someone to check.
Disclaimer: This article was originally written for my AI-native Power Grid Newsletter on Substack, but adapted for the Energy Central Community.