Somewhere in central Ohio, a half-built data center campus sits on 200 acres of freshly graded land, its concrete pads poured and its fiber backbone lit. What it does not have is electricity. The developer, like dozens of others racing to stand up AI training clusters across the United States, filed an interconnection request with the regional grid operator and was told to get in line. That line, according to federal data published in early 2025, now stretches roughly five years from request to commercial operation for the typical large power project. As recently as 2015, the same process took closer to two.
The bottleneck has turned grid access into the single largest constraint on AI infrastructure expansion, and it is one that no amount of capital spending can fix on its own. Land is available. Construction crews are available. Billions in financing are committed. But the physical equipment needed to move bulk power from transmission lines into a new facility, above all the high-voltage transformers that step electricity down to usable levels, is backordered on a scale the U.S. power sector has not seen in decades.
Five years in the queue
The clearest measure of the problem comes from Lawrence Berkeley National Laboratory, whose Queued Up: 2024 Edition tracks every power plant and large electrical load seeking to connect to the U.S. transmission system. The dataset covers all seven independent system operators and regional transmission organizations, plus dozens of individual utilities, through the end of 2023.
Its central finding: the median duration from interconnection request to commercial operation for projects that actually reached service in 2023 was approximately five years. That is a measured outcome, not a forecast. It reflects real projects passing through real engineering studies, equipment procurement, and construction milestones. A decade earlier, the same median hovered around two to three years.
The queue itself has also swollen. Berkeley Lab counted more than 2,600 gigawatts of generation and storage capacity waiting for interconnection at the end of 2023, roughly double the installed capacity of the entire U.S. grid. Most of those projects will never be built. But the sheer volume clogs the study process, because grid operators must evaluate how each new entrant affects system reliability before granting approval. Every speculative solar farm or battery project in the queue slows the path for the data center behind it.
The transformer shortage underneath it all
Even if the queue moved faster on paper, the hardware would not keep up. The U.S. Department of Energy’s Office of Electricity, which maintains a supply chain monitoring program for critical grid equipment, documented distribution transformer lead times climbing into the 12-to-30-month range by 2023. Utilities that once ordered a standard distribution unit and received it in a few months found themselves waiting more than two years.
For the largest units, the picture is worse. Large power transformers, the 100-MVA-and-above class that hyperscale data centers typically require, are custom-engineered products weighing hundreds of tons. Only a handful of factories worldwide produce them. A 2024 DOE assessment of large power transformers noted that domestic manufacturing capacity remains limited and that lead times for these units can stretch well beyond two years, with some industry participants reporting waits of three years or more. The United States imports roughly 80 percent of its large power transformers, primarily from South Korea, Mexico, and Europe, leaving procurement timelines exposed to trade disruptions and competing global demand.
The shortage is not new, but AI has intensified it. When a single training campus can draw 300 to 500 megawatts, and companies like Microsoft, Google, Amazon, and Meta are each planning multiple campuses simultaneously, the aggregate transformer demand lands on a supply chain that was already strained by routine utility replacement cycles and the buildout of renewable generation.
What the hyperscalers are doing about it
The largest cloud and AI companies are not waiting passively. Several have moved to secure power outside the traditional interconnection queue entirely.
Microsoft’s 2023 agreement to purchase power from the restarted Unit 1 reactor at Three Mile Island, operated by Constellation Energy, was structured in part to access an existing grid connection at a site that already had transmission infrastructure in place. Amazon’s power purchase agreement with Talen Energy’s Susquehanna nuclear plant in Pennsylvania followed a similar logic: buy electricity from a generator that is already interconnected, avoiding years of queue delay. Google has signed agreements with Kairos Power for small modular reactor capacity, though those units are not expected to deliver power until the early 2030s.
Behind-the-meter generation is another workaround gaining traction. Some developers are installing on-site natural gas turbines to supply initial power while they wait for full grid interconnection. Others are exploring fuel cells or co-locating with existing industrial power users who have spare capacity on their grid connection. Each of these strategies adds cost and complexity, but the calculus has shifted: paying a premium for power today is cheaper than letting a $3 billion data center sit idle for three years.
Site selection has also changed. Developers increasingly target locations where retired coal or gas plants left behind substations and transmission rights that can be repurposed. Those brownfield sites command steep premiums precisely because they shave years off the interconnection timeline. But the supply of such locations is finite, and each one claimed tightens the market for every competitor behind it.
Policy is moving, but not fast enough
Federal regulators have recognized the problem. In July 2023, the Federal Energy Regulatory Commission issued Order 2023, the most significant overhaul of interconnection rules in two decades. The order requires grid operators to process queue applications in clusters rather than one at a time, imposes stricter financial deposits to weed out speculative filings, and sets firmer study deadlines. FERC estimated the reforms could cut years off the process for projects that survive the new screening.
On the supply side, President Biden invoked the Defense Production Act in June 2022 specifically to accelerate domestic production of transformers and other grid components. The Inflation Reduction Act and the Bipartisan Infrastructure Law together directed billions toward grid modernization, including incentives for domestic transformer manufacturing. Several new or expanded factory projects have been announced since then, but building a transformer factory takes years of its own, and none of the new capacity is expected to meaningfully ease backlogs before 2027 at the earliest.
State-level action varies widely. Virginia, which hosts the largest concentration of data centers in the world in Loudoun County, has seen Dominion Energy struggle to keep pace with connection requests. Texas, where ERCOT’s relatively streamlined interconnection process once attracted developers, is now seeing its own queue swell as AI campuses compete with solar, wind, and battery projects for grid access. PJM Interconnection, the grid operator covering 13 states from New Jersey to Illinois, had the longest queue backlog in the country as of late 2023, with more than 260 gigawatts of capacity awaiting study.
The gap between ambition and electrons
Wall Street analysts and tech executives routinely cite figures suggesting the U.S. will need 30 to 50 additional gigawatts of data center capacity by 2030 to meet projected AI demand. Building that capacity is, in principle, a solvable engineering and financing problem. Connecting it to the grid is not, at least not on the timelines the industry is advertising.
The Berkeley Lab data makes the math uncomfortable. If the median interconnection timeline remains near five years, a project that files its request in mid-2026 would not reach full commercial operation until 2031 under typical conditions. Transformer procurement adds a hard floor beneath that timeline: you cannot energize a substation with equipment that has not been manufactured yet. And the queue is getting longer, not shorter, as AI demand layers on top of the already record volume of renewable energy and storage projects seeking connection.
None of this means AI data center growth will stop. The financial incentives are too large and the strategic stakes too high for companies to walk away. But it does mean the growth trajectory will be shaped as much by grid logistics as by chip supply or capital availability. The companies that secure power fastest will have a durable competitive advantage, and the ones that assumed grid access was a formality are learning, expensively, that the hardest part of building an AI data center may not be the data center at all.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
