We have been documenting the slow-motion collision between the AI capex fantasy and the laws of physics since late 2025, when we noted that some of Oracle’s OpenAI data centers were being pushed back to 2028 (Dec 12, 2025). Since then the delays have gone from anecdote to trend: in April we reported that half of US data centers scheduled to start in 2026 will be canceled or delayed (Apr 13), in August that more than two-thirds of the power sought for US data centers will never materialize (Aug 14), and just last week the two flagship Stargate campuses hit the wall within 24 hours of each other, with SB Energy delaying the IPO meant to fund the world’s largest data center (Sept 23) and Oracle’s Project Jupiter declaring force majeure (Sept 24).
Now we can add the US electrical grid itself to the list of things that won’t arrive on time.
On Wednesday, PJM Interconnection, operator of the largest US grid stretching from Illinois to DC serving 67 million customers across 13 states, and home to Virginia’s Data Center Alley, suspended the one-time “backstop” auction it was about to launch to plug a 6.8 gigawatt hole in its supply stack, less than a day after FERC ordered it back to the drawing board. The procurement, which was supposed to open for offers on Sept 30 and run through Oct 21 with selections by early December, now has, in the words of a PJM spokesman, a launch date that is “yet to be determined.”
For those keeping score, that is a power auction for data centers… delayed. You can’t make this up.
Recall that in July, PJM’s base capacity auction for 2028/29 came up 6.8GW short of its reliability requirement, the first time in PJM’s history that the entire RTO fell short, and only the price cap kept the clearing price at roughly $325/MW-day. As we tweeted that night, without the ceiling, PJM capacity prices would have been 70% higher at $554.72. A few hours later we added that “PJM is already below the critical reliability threshold.”
