Economist flags 2–5 year delays

The bottleneck in AI infrastructure is not just chips. It is electrical gear. Transformers, switchgear, batteries, cooling loops, and utility hookups are now slowing data-center buildouts hard enough that almost half of the U.S. projects planned for 2026 may be delayed or canceled. That is the actual news here — not that demand is huge, everyone knew that, but that the constraint has shifted from silicon to the stuff that turns a building into usable compute. (bloomberg.com) ### What exactly is running short? The unglamorous parts. Transformers step voltage up and down. Switchgear routes and protects power inside the site. Batteries and backup systems keep the place stable when the grid wobbles. Bloomberg’s April 1 report says shortages of those component(bloomberg.com)26. (bloomberg.com) ### Why are transformers such a big deal? Because you cannot improvise around them. A missing server can be swapped later. A missing transformer means the building does not turn on. Recent industry reporting says delivery times that used to be measured in months are now often measured(bloomberg.com)ied to the April 2026 data-center crunch says some deliveries can stretch to five years. (utilitydive.com) ### Why did this get worse now? AI arrived on top of an already stressed grid-supply chain. The same factories making transformer cores, switchgear, and cable are also serving grid upgrades, renewable projects, industrial loads, electric vehicles, and heat pumps. The IEA said in February 2025 that procurement times and(utilitydive.com) scratch — it piled a giant new load onto a system that was already tight. (iea.org) ### Where does Nvidia fit into this? Nvidia is the reason each new hall is so power-dense. The latest rack-scale systems are not just “more GPUs.” They are giant integrated power-and-cooling objects. Nvidia’s GB200 NVL72 links 72 Blackwell GPUs and 36 Grace CPUs in o(iea.org)eparing a facility that can feed and cool them safely. (nvidia.com) ### Why can’t companies just install the GPUs later? Because installation is part of the bottleneck now. Nvidia’s rack-scale systems need liquid manifolds, bus bars, power shelves, and cooling distribution equipment inside the rack. Nvidia’s own DGX GB rack documentation lists liquid-cooling manifolds and power shelves as core system components. Basi(nvidia.com)is not “liquid-ready,” the compute is stranded. (docs.nvidia.com) ### Does this create pricing power for existing capacity? Yes — and that may be the most important market effect. If new capacity slips by quarters or years, already energized data centers become more valuable. Owners with power in hand can charge a premium, while customers without long-term supply agreements get pushed toward whichever ven(docs.nvidia.com)rn into vendor lock-in. This part is an inference, but it follows directly from the supply crunch and delay risk. (bloomberg.com) ### Is this just a U.S. story? No. The U.S. is where the April 1 Bloomberg reporting focused, but the underlying constraint is global. The IEA’s warning was about grid infrastructure broadly, not one country, and transformer supply has been tight across major power markets. AI data centers just happen to be the newest, richest buyer trying to jump the queue. (iea.org) ### Bottom line The AI buildout is no longer limited by who can buy the most GPUs. It is limited by who can secure power gear, utility access, and liquid-ready floor space first. In this phase of the race, the scarce asset is not intelligence. It is electrified, cooled, commissioned capacity. (bloomberg.com)