Speakers at a Cleveland energy forum argued the real bottleneck is no longer generation. It is the transmission lines, substations, and transformers that take years to permit and install before a large data center campus can be served.
For nearly two decades, electric utilities built their long-range plans around the assumption that demand would barely grow. Now a single AI data center campus can ask for hundreds of megawatts at one site, and the physical infrastructure needed to deliver that power, from transmission corridors to substation transformers to switching equipment, takes years to permit, manufacture, and install. That gap, between the speed of AI's appetite for electricity and the pace of the physical grid build-out, is the real story of the AI power boom, and it is forcing utilities and grid-reliability bodies to redesign the planning, interconnection, and reliability processes they have used since the flat-demand era began.
A Cleveland forum laid the trade's anxieties in plain terms. At Thompson Hine's "Powering the Future: Energy in the AI Age" event, speakers from a regional reliability organization, a large investor-owned utility, an engineering firm, and the host law firm argued that data-center load growth is no longer a generation problem. It is a deliverability problem, and the grid was not sized for it.
Brian Thiry, director of strategic engagement and reliability at ReliabilityFirst, the regional entity that oversees grid reliability across the Mid-Atlantic and parts of the Midwest, said single-site AI loads can range from hundreds of megawatts to a gigawatt or more, well beyond anything utilities planned for when flat demand was the baseline. His point was not that utilities have suddenly run out of power plants. It is that the wires, substations, and transformers that move electricity from a generator to a data-center fence require years of advance work, and the queue is now stuffed with requests that the existing system was never sized to handle. "It's not just about the generation and the load," Thiry said at the forum. "It's about the deliverability and the infrastructure."
That reframe matters because the easy media story about AI and power is that there will not be enough generation. The harder story is that even when generation is available, getting it to a large data center campus requires physical assets that operate on a multi-year cycle. Permitting a new transmission line, ordering a large power transformer, and building the substation that ties them together is measured in years, not quarters, and that timeline is now colliding with the construction schedules of the largest AI operators.
Rachel Lindesmith, director of national accounts at FirstEnergy, one of the largest investor-owned utilities in the Midwest and Mid-Atlantic, said utilities are pushing data-center developers to begin interconnection conversations earlier, before projects are fully scoped, because the queue of large-load requests has grown faster than the engineering capacity to study them. The practical effect is that the planning clock now starts before the customer's clock does, a reversal of the order utilities used to follow when demand growth was modest and predictable. Lindesmith's framing, reported at the same forum, treats this as the customer-facing edge of the same redesign Thiry described from the reliability side.
For decades, the dominant assumption across the U.S. utility sector was that demand growth would be flat or near zero, a regime driven by energy efficiency, slower growth in industrial electricity use, and the gradual retirement of energy-intensive manufacturing. Thiry noted that ReliabilityFirst had already been raising resource-adequacy concerns tied to plant retirements and electrification before hyperscale AI entered the interconnection queue. AI demand did not invent that pressure, but it has accelerated it sharply. PJM, the regional grid operator that covers much of the Mid-Atlantic and Midwest, has reportedly revised its long-range load forecasts upward to reflect new AI campuses, electrification, and manufacturing. Those specific forecast figures should be confirmed against PJM's own load-forecast report before being cited as fact, because the forum reporting that surfaced them is one source, not a corroborated trend.
What is corroborated, in the sense of being on the public record from a named institutional speaker, is that the planning process itself is being redesigned. Thiry framed the changes as a redesign of transmission planning, substation design, transformer specifications, interconnection studies, and reliability standards. Lindesmith framed them as a customer-engagement change: developers and utilities now have to start talking before a deal is finished, because the engineering work that follows cannot be compressed.
The watch items that follow from the forum are concrete. Will PJM and other regional grid operators publish updated interconnection study practices that account for large single-site loads? Will FERC, the federal regulator that oversees interstate electricity markets, address large-load customers in its interconnection rulemaking? Will investor-owned utilities begin filing new large-load tariffs or rate-case testimony that explicitly addresses data-center service? Each of those moves would be evidence that the flat-demand playbook is being retired in writing, not just in conversation at industry forums.
The AI power boom is not, strictly speaking, rewriting the utility playbook from a blank page. The pressure on resource adequacy, transmission planning, and transformer supply was already visible before the largest AI operators arrived. What the boom has done is compress years of anticipated change into a much shorter window, and force utilities, reliability organizations, and regulators to act on planning problems they had been able to defer. The new bottleneck is not whether enough electrons can be generated. It is whether the wires, transformers, and substations needed to deliver them can be built in time.