AI data centers, the bottleneck migrates

The transformer, the bottleneck no one saw

Once the chip is fabricated, the constraint that bounds compute becomes a precise, unglamorous machine: the large power transformer. Its lead times exploded to about 128 weeks (2.5 years) in the second quarter of 2025, and up to 144 weeks, four years, for a power-plant step-up transformer. The cause lies upstream: grain-oriented electrical steel (GOES), a highly specialized core material, has only a handful of qualified producers, and the United States imports most of its large transformers.

A data center installs its GPUs in a few months but waits years for the transformer that connects it. Copper, switchgear and circuit breakers suffer parallel delays. The compute frontier is now bounded by heavy electrical equipment whose supply chain is as concentrated and slow as that of the chips themselves, except that no CHIPS Act has yet taken aim at it.

The flight off-grid

Because grid connection is now counted in years, at PJM, the delay between application and commissioning went from under 2 years in 2008 to over 8 years in 2025, and more than 2,000 GW wait in the US queues, about double the existing fleet, hyperscalers stop waiting for the grid and generate their own power "behind the meter."

Two escape routes. Gas first: GE Vernova's turbine order book has swollen to ≈80-100 GW, sold out through 2029-2031, of which roughly 20 % for data centers. Nuclear next: Microsoft signed a 20-year power purchase agreement to reopen Unit 1 of Three Mile Island (835 MW, restart targeted for 2028), Amazon leaned on Talen, Meta launched its own tenders. The bottleneck has migrated a second time: after "silicon → megawatt," it becomes "megawatt → who can build their own power plant." The dependency no longer points to TSMC, but to turbine makers and reactor restarts, a chain even slower to scale.

Dossier sources