The American power grid was not built for what is coming. Across the country, developers have announced plans for enormous AI data center campuses, facilities that can each consume as much electricity as a small city. But between the blueprint and the breaker panel sits a piece of equipment most people have never thought about: the large power transformer. These multi-story, custom-built machines step voltage up or down between high-voltage transmission lines and the facilities they serve. Without them, no data center connects to the grid. And right now, the wait time for a new one stretches to five years.
That bottleneck is the central finding of a recent analysis by Sightline Climate, a research organization that tracks clean energy and infrastructure projects. According to its data, roughly 140 AI data center projects representing about 12 gigawatts of new capacity are in various stages of planning across the United States. Sightline’s assessment: approximately half of that pipeline is at serious risk of delay or cancellation because the transformers needed to energize those sites simply cannot be manufactured fast enough.
The collision is stark. Demand for data center power is accelerating at a pace the utility industry has not seen in decades, while the supply chain for the most critical piece of grid hardware remains stubbornly slow to scale.
The demand wave is well documented
Federal data leaves little room for doubt about the trajectory. A 2024 report from the U.S. Department of Energy, drawing on research by Lawrence Berkeley National Laboratory, found that U.S. data center electricity consumption rose from roughly 58 terawatt-hours in 2014 to about 176 TWh by 2023. Under different growth scenarios, the DOE projects that figure could reach 325 to 580 TWh by 2028. Even the low end of that range represents a near-doubling from 2023 levels in just five years.
Globally, the picture is similar. The International Energy Agency reported in its special analysis of energy and AI that data centers worldwide consumed approximately 415 TWh of electricity in 2024, accounting for about 1.5 percent of all electricity used on the planet. That share is climbing as AI training runs grow larger and inference workloads multiply.
What makes AI facilities different from traditional data centers is power density. A conventional enterprise server room might draw a few megawatts. An AI training cluster packed with thousands of GPUs can demand 50 to 100 megawatts per building, running at near-full utilization around the clock. That kind of load does not just require more electricity. It requires heavier electrical infrastructure at every point between the power plant and the server rack, starting with the transformer.
Why transformers are the chokepoint
Large power transformers are not off-the-shelf products. Each one is engineered to specific voltage ratings, cooling requirements, and site conditions. The core is wound from grain-oriented electrical steel, a specialty material produced by a handful of mills worldwide. Assembly, insulation, and testing take months. A single unit can weigh 200 to 400 tons and require a specialized railcar or barge to transport.
Before the current AI boom, transformer lead times were already stretching. Utilities across the country had been deferring replacements of aging units for years, and the push to interconnect large-scale wind and solar farms added new orders to an already crowded queue. Then AI data center developers arrived with requests for substations that might need multiple large transformers per campus, each rated at hundreds of megavolt-amperes.
The result, according to utility industry groups and regional grid operators, is a backlog that now runs three to five years for many transformer classes. Some manufacturers, including major players like Hitachi Energy and Siemens Energy, have announced factory expansions, but new production lines take years to build and staff. The specialized workforce required to wind transformer coils and perform high-voltage testing cannot be trained overnight.
Even when new manufacturing capacity comes online, it must serve competing demands. Aging grid infrastructure needs replacement. Renewable energy projects need interconnection equipment. Industrial electrification and EV charging networks add further pressure. AI data centers are not the only customers in line, and in many cases, they are not at the front.
What ‘killed’ actually means
Sightline Climate’s finding that roughly half of the 140 tracked projects face serious jeopardy deserves careful reading. “Killed” can mean different things depending on the project’s stage. Some developments in early feasibility may be permanently shelved. Others in advanced permitting might be downsized, with a planned 500-megawatt campus scaled back to 100 or 200 megawatts to match available grid capacity. Still others could be paused for years, waiting for transformer deliveries before construction resumes.
The distinction matters. A project that shrinks from its original scope still adds capacity to the grid. A project that relocates to a region with shorter interconnection timelines still gets built, just not where it was first announced. The 12-gigawatt pipeline is not a fixed number that either materializes in full or vanishes. It is a moving target shaped by equipment availability, utility negotiations, and developer patience.
That said, the pattern Sightline describes is consistent with what grid operators have been signaling. PJM Interconnection, the regional transmission organization that manages the grid across 13 states from Virginia to Illinois, has publicly discussed its swelling interconnection queue, which exceeded 250 gigawatts of pending requests as of late 2024. Many of those requests come from data center developers concentrated in Northern Virginia, central Ohio, and other PJM territories. The organization has warned that processing times for new interconnection studies have lengthened significantly.
The regions feeling it most
Northern Virginia’s “Data Center Alley” in Loudoun County remains the densest concentration of data center capacity on Earth, and it is also where grid constraints are most acute. Dominion Energy, the region’s utility, has flagged transmission limitations and has been working to build new substations and transmission lines, but permitting and construction timelines for that infrastructure stretch years into the future.
Central Ohio, where companies including Google, Amazon, and Meta have announced major campuses, faces similar pressures. West Texas and the Phoenix metro area have attracted developers seeking cheaper land and power, but even those regions are discovering that available transmission capacity is thinner than initial site assessments suggested.
For the communities that courted these projects with tax incentives and rezoning approvals, the transformer bottleneck introduces an uncomfortable variable. Promised jobs, tax revenue, and economic development may arrive years later than projected, or at a fraction of the originally announced scale. Local officials who approved water and land-use concessions based on aggressive construction timelines may find themselves waiting alongside the developers.
Policy has been slow to catch up
The federal government has taken some steps. In 2022, the Biden administration invoked the Defense Production Act to support domestic transformer manufacturing, citing grid reliability concerns. But the DPA authority has been used sparingly, and the primary constraint is not regulatory approval but physical production capacity and workforce availability.
Congress has also discussed reforms to the interconnection process, and the Federal Energy Regulatory Commission has issued orders aimed at clearing queue backlogs. These measures address procedural delays but do not solve the hardware shortage. A faster interconnection study does not help if the transformer it specifies cannot be delivered for four years.
Some developers are exploring workarounds. On-site generation using natural gas turbines or fuel cells can provide interim power while grid connections are built out. Modular data center designs allow facilities to start small and scale up as grid capacity becomes available. A few companies have even begun stockpiling transformers, purchasing units speculatively before specific projects are fully permitted.
None of these approaches fully substitutes for grid-connected utility power at the scale AI training demands. On-site generation raises emissions and permitting questions. Modular builds sacrifice the economies of scale that make hyperscale campuses financially attractive. Stockpiling transformers shifts the shortage to other customers without expanding total supply.
Where the buildout actually lands
The tension at the center of this story is not whether AI data centers will be built. The demand signal, backed by DOE and IEA data, is too strong and too well-capitalized to simply evaporate. Microsoft, Google, Amazon, and Meta have collectively committed hundreds of billions of dollars to AI infrastructure. That money will find outlets.
The real question is how much of the planned capacity arrives on schedule, at the intended scale, and in the locations developers originally chose. Sightline Climate’s tracking suggests the answer, for now, is significantly less than half. The transformer bottleneck acts as a physical filter on ambition, converting gigawatt-scale announcements into megawatt-scale realities.
For the power grid, this may be an accidental safety valve. Utilities that were scrambling to figure out how to serve 12 gigawatts of new load in a compressed timeframe may instead face a more gradual ramp, buying time to build generation, transmission, and distribution infrastructure. For AI companies racing to train the next generation of models, it is a constraint that no amount of software optimization can route around. The grid moves at the speed of steel and copper, not silicon.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
