The real AI infrastructure bottleneck isn't chips. It's transformers

Everyone's tracking the chip shortage. Almost nobody's tracking the transformer queue, and that's the one actually setting the AI buildout timeline.

In 2021 it took about 50 weeks to build a large power transformer. By 2024, 120 weeks. Today it's 128 weeks for a standard unit and 144 for the generator step-up transformers data centers actually need, nearly three years. The largest custom units are reportedly closing in on four.

Omdia's chief analyst said it plainly this spring: AI infrastructure isn't constrained by semiconductor innovation. It's constrained by grid interconnection queues running three to seven years and transformer lead times nearly as long. His estimate: 30–50% of planned 2026 data center capacity slips to 2028.

Here's the opportunity inside the problem: the U.S. imports about 80% of its power transformer supply. That's not a chip shortage. That's a manufacturing gap closeable by real industrial capability, structured the right way, fast enough to matter.

Altakon maps that gap for manufacturers deciding whether, and how, to build U.S. capacity, before the bottleneck closes itself the slow way, without them in it.