The thing most readers get wrong about AI's power problem is that it is a power problem. On July 22, 2026, a transmission line fault in Ashburn, Virginia — the heart of the world's largest data center cluster — knocked more than 3 gigawatts of load off the grid in seconds. Nothing was wrong with supply. Nothing had been generated and lost. The load simply left, all at once, because the equipment meant to protect it was built for a different kind of customer.
Two years earlier, a single failed surge arrester dropped roughly 60 Virginia facilities and 1,500 megawatts simultaneously. No one had anticipated that so much uniform load would respond to grid faults the same way at the same time.
That is the contradiction running underneath every headline about data center electricity. The public argument is about electrons — more turbines, more solar, more transmission. The actual failures are about architecture, and the actual delays are about the wiring between available power and the point where a customer can take service.
The actual failures are about architecture, and the actual delays are about the wiring between available power and the point where a customer can take service.
Start with the queue, because the queue is where the schedule actually lives. Google's Marsden Hanna said last January that transmission barriers are "the number one challenge we're seeing on the grid," with utilities in many markets quoting four to ten years to interconnect. One utility told Google it would need 12 years just to study the request.
Berkeley Lab's numbers explain why. Interconnection wait times have doubled, from under two years for projects built between 2000 and 2007 to over four years for those built between 2018 and 2024. Only 13% of capacity that requested interconnection from 2000 to 2019 had reached commercial operation by the end of 2024.
Nearly 2,300 gigawatts of generation and storage now sit in U.S. interconnection queues. Total installed capacity in the country is about 1,400 gigawatts. The queue is larger than the grid it wants to join.
Here is the part that gets misread. Berkeley Lab's 2026 queue dataset covers generation requests seeking transmission connections — and explicitly excludes load interconnection requests, distribution-connected projects, and behind-the-meter projects. Generation queues and demand queues are not the same thing and are not measured the same way. Treating them as one number obscures the technical differences between adding a power plant and connecting a gigawatt-scale consumer.
That distinction matters because a utility can have plenty of generation in its territory and still be unable to serve you. "Power is available" is not a schedule. The connection study has to become an actionable design, the network upgrades have to be approved and funded, the transformers have to arrive, the protection systems have to be tested, and the site has to be energized.
ERCOT's Batch Zero filing shows what large-load demand actually looks like when it shows up in one pile. Base case: 66.4 gigawatts across 204 projects. Studied: 125.4 gigawatts across 158 projects. The studied figure is nearly twice the base figure — far more generation proposed than load projected, which invites the obvious question of whether developers are overbuilding now for expansions they haven't announced yet.
Within that, a category called Withdrawal Limited Private Use Network — essentially bring-your-own-generation for large consumers — totals 15.4 gigawatts across 11 projects. Of the 15,444 megawatts of generation in it, 12,824 are gas. Solar accounts for 1,654 megawatts and batteries 330. Building solar in ERCOT right now, bring-your-own-generation mostly means bring-your-own-gas.
Then look at what a single project's connection actually costs and takes. The Department of Energy, Brookfield, and NextEra plan for the Paducah Site: 1,825 megawatts of load, $436 million in transmission, full in-service date of June 2031. That is roughly $239 per kilowatt just for the transmission piece, and a lead time worth watching closely. The obvious unanswered question is what generation is contracted behind it.
Multiply that across a pipeline and you get the grid-spending math. Goldman Sachs Research estimates about $720 billion in grid spending through 2030 may be needed to support data center expansion. Investor-owned utilities plan $1.1 trillion in grid investments through 2029.
Transmission lines take six to seven years to build, and long-distance transmission can take up to 15. Regional transmission organizations are pushing 765-kilovolt backbones, but the timelines don't compress. Gas turbines for the most efficient models are backordered five to six years. The equipment and the civil works are not waiting on each other politely; they are all on the critical path at once.
So developers are routing around the problem. Reuters Events reports that data center developers were planning roughly 56 gigawatts of on-site capacity as of early 2026 — about 30% of all planned data center builds nationwide. Texas leads with over 20.6 gigawatts of planned behind-the-meter capacity, followed by New Mexico at 9.2, Pennsylvania at 7.5, and Utah at 6.0.
Grid operators are adapting the rules to reward that. Southwest Power Pool's High Impact Large Load Generation Assessment framework offers connection agreements in under 90 days to data centers that bring their own generation to support their load. That is the queue being replaced by a shortcut, for the customers who can afford to buy their way past it.
Now the harder problem, the one that no amount of generation solves. An AI campus can swing 70% of its load in milliseconds during a training run, then trip offline just as fast at the first sign of trouble upstream. Each facility is behaving rationally — protecting billions of dollars of compute. Together, at gigawatt scale, they are a grid behavior nobody engineered for.
The standard power stack has not changed in decades: medium-voltage power arrives, transformers step it down, low-voltage UPS units condition it, and it reaches the racks. Push that design to AI scale and it cracks in three places. The batteries are a spare tire sized for minutes, not for absorbing volatile swings around the clock. The UPS spends most of its life in bypass, so compute swings go out raw and grid transients come in unfiltered. And the protection logic was written when "large load" meant 50 megawatts — in the 2024 Virginia event, most of the lost load traced to schemes designed to disconnect on the third voltage dip.
The proposed fix is three moves made together: move power protection up from 480 volts to medium voltage, move it out of the data hall to modular enclosures near the substation, and move it into the path so every electron runs through it rather than around it. In early 2026, a full-scale system was tested at the National Laboratory of the Rockies, hitting real AI load profiles and real grid faults, including a full zero-voltage event, in the same loop. It cleared ERCOT's large-load voltage ride-through requirements with room to spare. The utility then certifies one medium-voltage box instead of untangling every transformer, chiller, and switchgear lineup behind it.
Which brings us to the strangest result in the evidence. Verrus is building data centers designed to answer every objection the public has raised: batteries instead of diesel generators, making them grid assets that help hold down power prices; closed-loop cooling using a fraction of the water; quieter equipment; more servers squeezed into the same grid connection. Both of its proposed sites — in Michigan and Oregon — are mired in fierce local opposition. It has not yet built a data center.
The engineering is largely solvable, and in some cases already solved. The interconnection queue, the transmission math, and the permitting politics are not. As one analyst put it, the bottleneck isn't silicon or software — it's electricity, and specifically the last mile of it. Bernie Sanders and Alexandria Ocasio-Cortez have introduced a bill calling for a federal moratorium on new AI data center construction until environmental and cost rules are written. More data center projects were delayed or blocked in a single quarter last year than in the previous two years combined.
That is the real collision. Not supply versus demand, but a grid optimized for two decades of flat consumption meeting an industry that needs gigawatts delivered in a specific place, on a specific date, behaving in a way the grid was never designed to tolerate.