There are 780 gigawatts of announced data center projects in the U.S., which is more than the country’s current total peak load of 759 GW. “It’s an additional U.S. of power by 2030,” said Oliver Kerr, managing director of North America at Aurora Energy Research, onstage at Latitude Media’s Transition-AI 2026 conference this month. “I think that’s just a wild number.”
However, the reality of how many projects actually come online may be just a fraction of that, given infrastructural and logistical hurdles. The key question that the energy industry has to answer is precisely how much of this potential load growth is real, and how much generation will be required to meet it.
Estimates are all over the place. Grid Strategies, which in 2023 declared the end of “the era of flat power” with its estimate that peak demand could grow to 38 GW through 2028, quadrupled its earlier prediction to 166 GW over the next five years, with about 90 GW of that tied to data center development specifically — though the researchers themselves said the number is likely too high, given that it’s based on the forecasts of utilities and regional planners, rather than the realities on the ground.
Meanwhile, in December BloombergNEF projected 106 GW of data center power demand by 2035, though it notes that many of those projects remain speculative. That’s a 36% increase since its last forecast.
This discrepancy is because the number of data centers that companies say they need on paper doesn’t necessarily account for the generation, grid infrastructure, and capital required to build them. It’s not just about money, but rather about the physical reality of getting steel in the ground.
Generation and transmission constraints
Kerr emphasized that even though net capacity additions on the grid in the last few years have been increasing, they’re “nowhere near” to meeting the needs of the supposed data center pipeline.
“The implied annual build rate for new generation needed to power ambitious data center forecasts by 2030 is historically unprecedented,” he said in his presentation.
Transmission further complicates things. Transmission infrastructure, especially for interregional transmission, is notoriously challenging to build in the U.S. According to a February report from Grid Strategies and Americans for a Clean Energy Grid, regional planning reforms have made some incremental progress in many parts of the U.S., but significant work remains to be done. And as load growth continues to accelerate, the looming reliability gap is still getting bigger.
Aurora modeled what adding 10 GW of data center load in ERCOT without adding new transmission lines would do to the grid, and the consequences were stark: quintupling congestion costs and spiking power prices by 34%. Relying on grid-enhancing technologies or other ways of making the most of the existing grid isn’t enough to prevent this impact. Kerr emphasized that “there is no way” that this scenario would come true, but that they modeled it purely to emphasize that transmission can’t be ignored.
“If you want to build the transmission system that you need for 2030 you really needed to have started yesterday,” he said, “and this will impose a hard constraint on the pace of build out of data centers.”
The barriers to certainty
Meanwhile, barriers to building the data center projects themselves are mounting as well. Large load tariffs are springing up all over the country; interconnection queue reforms are making it harder to push speculative projects; EPC contractors, who handle actually getting a project built, are fielding so much demand that they can be selective in their choice of who to work with; and all the while the labor and materials to build these projects is increasingly scarce.
Nat Bullard, co-founder of Halcyon, highlighted Fermi America’s six-GW Project Matador in Amarillo, Texas as an example of the challenges in his own presentation at Transition-AI. That single project, he pointed out, will require 90 Siemens industrial gas turbines on a single plot of land, built in a relatively short period of time — which Bullard sees as a long shot.
One major reason is that demand for this equipment, and many of the other materials needed to build and power a data center, is outstripping supply, prompting long waits and high prices. Data center companies are increasingly investing in gas as a fast and essential way to power their ambitious projects. But as the competition mounts for the equipment needed, fossil gas is becoming “not necessarily the quickest, and definitely not the cheapest” option, Bullard said.
Just a week after the conference, Fermi CEO Toby Neugebauer stepped down after publicly struggling to find an anchor tenant for Matador. Before his departure, Neugebauer told Axios that he may have been naive about the complexity of building these massive projects, citing especially the cooling systems for AI chips and “where the supply chain is” for that equipment.
Brett Foster, VP of renewables at the EPC firm McCarthy Building Companies, emphasized the labor constraint. A year ago it took six months to break ground on a data center project, he said, and now it’s more like 18 months — and that’s largely because of the challenges with securing the labor needed. As a result, he’s skeptical of most of the queue.
“You look at the queue and maybe one of four [projects] is real and actually going to move forward,” he said on a Transition-AI panel on delivering clean capacity for data centers. “And so for us, what we prioritize is certainty.”
That certainty is hard to guarantee, however. As tech company priorities change, AI gets more efficient, and the grid itself evolves, the size and shape of the data center pipeline is sure to change as well.
As for the risk that these estimates could all be so overblown that the U.S. risks overbuilding? Well, it went almost unmentioned at Transition-AI; there’s an assumption that even if data center load is a small fraction of what the pipeline would imply, anything the country is able to build will likely be used, either by other types of industrial load or to meet the needs of an electrifying economy. Ultimately the concern is not that demand projections are too high; it’s that deliverability simply isn’t possible at the scale that the tech industry is currently calling for.
