Delays widen AI supply crunch

The shortage is no longer just about GPUs. The harder thing to buy now is the stuff around them — transformers, switchgear, cooling equipment, grid access, and in some cases the gas turbines needed to make power on-site. That is the real reason a lot of AI data-center timelines are slipping. On April 1, Bloomberg laid it out bluntly: almost half of the US data centers planned for 2026 are expected to be delayed or canceled because the electrical gear needed to energize them is in short supply. (bloomberg.com) ### Why are data centers suddenly waiting on boring industrial gear? A modern AI campus is basically a power plant attached to a warehouse full of chips. The chips get headlines, but the campus cannot run until electricity can be stepped down, routed, protected, cooled, and backed up. (bloomberg.com)s were already tight from grid upgrades and electrification. AI demand piled on top. (bloomberg.com) ### What changed this year? The scale changed. Bloomberg tied the crunch to a huge 2026 build wave, with Alphabet, Amazon, Meta, and Microsoft collectively planning to spend more than $650 billion this year. That kind of capex does not magically create more factory capacity for transf(bloomberg.com)a site turn on. (bloomberg.com) ### Why is power the gating item now? Because utilities and equipment vendors are both backed up. Siemens Energy describes cases where data-center operators were told new utility power would take five years, including roughly two years of equipment lead time. It also says many relevan(bloomberg.com)an accept them on schedule. (siemens-energy.com) ### Can’t builders just make power on-site? Sometimes — but that queue is jammed too. Large natural-gas turbines now take more than five years from order to delivery, and EPRI says a unit ordered today would on average not begin operating until 2031. Smaller turbines are faster, but still typically take 18 to 36 (siemens-energy.com) workaround is getting scarce too. (utilitydive.com) ### Why does this make GPU capacity more valuable? Because a usable AI system is not “chips in theory.” It is chips installed in a live building with enough power density and cooling to run them. When the non-chip pieces are late, every energized rack becomes more valuable. The practi(utilitydive.com)starting to treat power-and-cooling capacity the way the industry used to treat chip supply. (datacenterfrontier.com) ### How are companies responding? They are pre-booking infrastructure years ahead. Schneider Electric signed multi-year supply-capacity agreements worth about $2.27 billion with Switch and Digital Realty, covering things l(datacenterfrontier.com)roject. (datacenterfrontier.com) ### Does this ease soon? Probably not. Gas-turbine demand is broad, not cyclical, and 2025 global orders more than doubled from 2024, reaching 846 units and 100.3 GW. On the electrical side, Siemens is still talking about multi-year equipment delays, and Bloomberg’s April snapshot already showed the damage landing in 2026 schedules. That looks less like a temporary snag and more like a new planning reality. (utilitydive.com) ### What’s the bottom line? AI infrastructure is now constrained by heavy industry. The scarce thing is no longer just compute silicon. It is the whole chain that turns silicon into working compute — power, cooling, interconnection, and generation. Until those bottlenecks ease, “more GPUs” will keep translating into “longer waits for actual AI capacity.”