YouTube flags data‑center limits

The bottleneck in artificial intelligence is no longer the chip alone. In early 2026, Sightline Climate estimated that 30% to 50% of announced United States data-center capacity slated for 2026 is unlikely to come online by year-end. (sightlineclimate.com) That sounds strange until you remember what a data center is: a warehouse full of computers that needs the electrical equivalent of its own small city. Deloitte says United States power demand from artificial-intelligence data centers could rise from 4 gigawatts in 2024 to 123 gigawatts by 2035. (deloitte.com) The new artificial-intelligence buildings are also much denser than older server farms. Deloitte says a five-acre site that used to draw 5 megawatts with central processing units can jump to 50 megawatts when it adds specialized graphics processors. (deloitte.com) So the race is not just for land. It is for a block of power big enough to run thousands of graphics chips at once, plus substations, transmission lines, batteries, and cooling equipment to keep the machines from overheating. (cbre.com) That is why real-estate firms now talk about electricity like beachfront property. Jones Lang LaSalle said in August 2025 that North American colocation vacancy had fallen to 2.3%, while its report described the market with the phrase “power is the new real estate.” (jll.com) Even where developers have money, they still need a place in line with the utility. Dominion Energy told Virginia regulators in February 2026 that it had about 70 gigawatts of large-load requests in its queue, almost three times its all-time system peak of 24.7 gigawatts from January 2025. (datacenterdynamics.com) Virginia matters because Northern Virginia is the biggest data-center market on the continent. Jones Lang LaSalle put that market at 5.6 gigawatts in mid-2025, more than triple Dallas-Fort Worth at 1.5 gigawatts. (jll.com) The queue problem is spilling beyond one utility. PJM Interconnection, the grid operator across 13 states and the District of Columbia, said in its 2025 load-forecast materials that it has special rules for adding “large load” requests because these projects can materially change the forecast. (pjm.com) Then there is the hardware behind the hardware. Bloomberg reported on April 1, 2026 that shortages of transformers, switchgear, and batteries were a major reason nearly half of planned United States data centers for this year were expected to be delayed or canceled. (bloomberg.com) Transformers are the heavy metal boxes that step voltage up or down so power can move from the grid into a building safely. Reporting summarizing the Bloomberg piece said high-power transformer lead times that were 24 to 30 months before 2020 can now stretch to as long as five years. (finance.yahoo.com) That is why the story in those YouTube videos is less “artificial intelligence demand is cooling” and more “the plumbing cannot keep up.” CBRE said 74.3% of all under-construction capacity in major North American markets was already preleased in the first half of 2025, which means much of tomorrow’s supply is spoken for before the doors even open. (cbre.com) Companies are already reacting by hunting for sites with existing power, moving to secondary markets, and exploring on-site generation instead of waiting for the grid. Sightline said only about 5 gigawatts of the 16 gigawatts slated for 2026 was actually under construction as of February 24, 2026, which is why so many headline-grabbing projects may slip. (sightlineclimate.com) So the next limit on artificial intelligence may look less like a software problem and more like a utility map. If the permits, substations, transmission lines, and transformers do not show up on time, enterprise artificial-intelligence rollouts can stall even when the demand for models and chips stays high. (deloitte.com)