Nearly half of the data centers planned to open in the US by 2026 face delays or outright cancellation, according to Sightline Climate analysts. Of the 12 gigawatts worth of computing infrastructure announced for this year, only a third is actually under construction. The gap widens further out: just 6.3 gigawatts of the 21.5 gigawatts planned for 2027 have broken ground, with most projects between 2028-2032 existing only on paper.
This isn't just a construction hiccup—it's a fundamental constraint on AI scaling. While everyone obsesses over GPU availability and model parameters, the unglamorous electrical infrastructure has become the real chokepoint. Batteries, transformers, and circuit breakers represent less than 10% of data center costs but can kill entire projects when delayed. As Andrew Likens from Crusoe Energy noted, "If one piece of your supply chain is delayed, then your whole project can't deliver."
The root cause traces to manufacturing dependencies on Canada, Mexico, South Korea, and China for critical electrical components. These aren't simple parts you can rush-order from McMaster-Carr—they're complex assemblies requiring months of lead time and ocean shipping. The industry built its expansion plans assuming these supply chains would scale seamlessly, a bet that's clearly not paying off.
For AI developers, this means the compute capacity you're planning around may not materialize when promised. The hyperscaler promises of infinite scale hit real-world physics: someone still has to manufacture transformers and ship them across oceans. Plan accordingly.