For AI data centers, the grid connection is the bottleneck, not the power

The bottleneck is the connection, not the kilowatt

The popular story about artificial intelligence and electricity is a story about scarcity. The models are hungry, the grid is tapped out, and the United States is running short of power. The numbers tell a stranger story. At the end of 2024, nearly 2,300 gigawatts of generation and storage were sitting in interconnection queues, waiting for permission to plug in, according to Lawrence Berkeley National Laboratory. That is more than double the capacity the country already has installed. The electricity is not missing. It cannot get connected. Building a power plant is now the easy part. Attaching it to the grid is the hard part.

That distinction matters because it changes what you fix. Berkeley Lab found that the median time from an interconnection request to commercial operation has doubled, from under two years for projects built between 2000 and 2007 to more than four years for those built from 2018 to 2024. Many projects never finish the trip. The lab reported that roughly 700 gigawatts withdrew from the queues in 2024, outpacing new requests, as developers walked away rather than pay for the network upgrades that studies kept assigning them.

Where the strain lands

The pressure is not spread evenly. It concentrates where data centers cluster, and nowhere more than Northern Virginia, the region the industry calls Data Center Alley. Dominion Energy said it had about 40 gigawatts of data-center capacity under contract as of December 2024, an 88 percent jump from 21 gigawatts the previous July. To serve it, Dominion raised its five-year capital plan to roughly $50 billion through 2029 and is building new 500 kilovolt transmission lines, because the wires into Loudoun County cannot carry the load. Most of that demand is still early. Dominion reported only about 9 gigawatts had reached signed electrical service agreements, the last step before construction.

The fragility of that interface showed up in July 2024. NERC, the grid reliability regulator, later reported that a fault on a 230 kilovolt line near Fairfax caused about 1,500 megawatts of data-center load to disconnect at once, across dozens of sites and 25 substations. The generation was fine. The transmission path, and the way the data centers reacted to a voltage dip, were not. NERC stood up a Large Loads Task Force the following month.

Texas tells a parallel story. Rather than let data centers crowd the ERCOT grid unchecked, the legislature passed Senate Bill 6 in June 2025, which lets the grid operator remotely curtail large loads during emergencies and makes those customers help pay for the interconnection they trigger. Even in the state that sells itself as the easy place to build, the response to data-center demand was more process, not less.

Copper, steel, and lead times

Dig into why a connection takes years and you reach a mundane answer: the hardware. A transmission line needs transformers, and transformers have become one of the hardest things on the grid to buy. Utility Dive reported in February 2025 that a transformer ordered now can take up to three years to arrive, against four to six weeks about five years earlier, citing the trade group NEMA. Transmission-scale units can run three to six years, Hitachi Energy told the outlet.

The reasons are structural. Roughly 80 percent of large power transformers used in the United States are imported, and the specialty electrical steel they need is scarce. A June 2024 National Infrastructure Advisory Council report counted more than 80,000 distinct transformer types in the American system, which makes standardizing and stockpiling them difficult. You cannot order your way out of this quickly, and the AI buildout is competing for the same units utilities need for ordinary load growth.

This is why generation is the wrong thing to count. A gas plant or a solar farm can be financed and sited in a matter of months. The wires, substations, and transformers that carry its output to a server hall cannot be, and they are what projects wait on.

The workarounds and their ceilings

Faced with multi-year queues, operators are trying to skip the line. The most aggressive version is to bypass the utility grid and generate on site, usually with natural gas, an approach the industry calls behind-the-meter or bring-your-own-generation. It is real, and it is hitting its own wall. GE Vernova said its gas-turbine backlog climbed past 100 gigawatts in 2025 and that it is largely sold out through the end of the decade, with lead times around three years. The queue did not disappear. It moved to the turbine factory.

The other workaround is co-location, parking a data center next to an existing plant and drawing power straight from it. That is what Talen Energy proposed at the Susquehanna nuclear plant in Pennsylvania, selling up to 480 megawatts directly to an Amazon campus. In November 2024, FERC rejected the arrangement, and reaffirmed that in 2025, worried about who pays for the grid services the data center still relies on. The ruling is a reminder that even a plant next door stays tied to the shared grid, and regulators are not ready to let large loads opt out of paying for it.

What to watch

The useful question is not whether America has enough electricity. It is whether a given project can secure a queue position, a transmission path, and the transformers to build it, inside the window an AI operator is willing to wait. On the current evidence, the gating items are process and hardware, not fuel or generation. So when a company announces gigawatts of new AI capacity, the honest follow-up is short. Where is it in the interconnection queue, and who is building the wire.