AI datacentres hitting delays

The AI boom was supposed to be a story about chips. Instead, it is turning into a story about the electrical gear nobody talks about until it is missing. In the US, 30% to 50% of the data-center capacity slated for 2026 now looks unlikely to arrive on time, according to Sightline Climate’s February outlook. Sightline says at least 16 gigawatts across roughly 140 projects are supposed to come online this year, but only about 5 gigawatts are actually under construction. The rest is still mostly a promise on paper (sightlineclimate.com). That gap matters because money is not the constraint. Bloomberg reported in February that Alphabet, Amazon, Meta, and Microsoft together are expected to spend about $650 billion in 2026 on data centers and equipment. The surprise is that this flood of cash is running into shortages of transformers, switchgear, and batteries rather than a shortage of GPUs. Bloomberg reported on April 1 that nearly half of planned US data centers for this year are expected to be delayed or canceled for exactly that reason (bloomberg.com 1) (bloomberg.com 2). These are not exotic parts. They are the plain metal boxes and coils that step voltage up and down, route power through a site, and keep a server hall from going dark. A modern AI campus needs them twice over: once inside the facility, and again out on the grid that has to feed it. Bloomberg’s reporting makes the problem stark. The same equipment is also being pulled by utilities, EV charging, factories, and building electrification. That means a data center developer is competing not just with other AI projects, but with the rest of the economy (bloomberg.com). The wait times have become absurd. Finance Yahoo, citing Sightline and Bloomberg, reported that high-power transformers that once took roughly 24 to 30 months before 2020 can now take as long as five years. POWER reported that switchgear averaged 44 weeks in the second quarter of 2025, while large-transformer lead times had already stretched into multi-year territory. When a project depends on every link in that chain arriving in sequence, one missing component can freeze the whole build (finance.yahoo.com) (powermag.com). That is why the supply chain keeps bending back toward China. Bloomberg found that US manufacturing still cannot produce enough of this equipment, forcing developers into import markets even after years of onshoring rhetoric. The bottleneck is not just domestic factory capacity. It is also the messy global route these parts take through raw materials, fabrication, shipping, and trade barriers. A country that leads the AI race in software and accelerators is still getting tripped by the hardware behind the wall (bloomberg.com). The projects that move fastest are the ones that secured gear early or found ways around the grid. Sightline says on-site and hybrid power setups still account for less than 10% of projects by count, but nearly half of announced capacity by megawatts. Bloomberg pointed to the giant Crusoe campus in Abilene, Texas, where more than 6,000 workers have been building an eight-building site that is expected to draw 1.2 gigawatts when complete. Speed there came from locking in electrical equipment early, before the queue got even worse (sightlineclimate.com) (bloomberg.com).