Why the rate limiter on 2026 US datacenter capacity is high voltage gear made mostly in China on multi year industrial cycles, a physical clock that press releases cannot accelerate.
The disputed figure is "about half of 2026 US datacenter capacity is delayed or canceled," a number that ricocheted around financial and tech Twitter within days of a Bloomberg article on April 1, 2026, titled "America's AI Build-Out Hinges on Chinese Electrical Parts" (SemiAnalysis, June 18 2026). The SemiAnalysis rebuttal argues the 50% figure is wrong, and on the methodology it is hard to argue: their analysts say they review every hyperscaler site dozens of times a year, cross-checking interconnection queues and SEC filings, and their revised year-end 2026 hyperscaler self-build number does not show a 50% cut. But the rebuttal leaves the durable mechanism exactly where Bloomberg put it in April, and that mechanism is the actual story.
The mechanism is electrical, not digital. A 2026 datacenter needs gigawatts of in-feed power, and gigawatts do not arrive through fiber. They arrive through large power transformers, the high-voltage switchgear that surrounds them, and the grain-oriented electrical steel laminated inside their cores. Most of that equipment class is made in China. A new large power transformer takes years to design, wind, and qualify. A new switchgear bay follows a comparable cadence. The steel mill that supplies the core steel runs on its own multi-year ramp. The orders that ship into US substations in 2026 were largely placed in 2023 and 2024. No analyst forecast, no hyperscaler capex announcement, and no GPU allocation decision in the next twelve months can shorten that clock.
This is the part of the AI datacenter story that is almost never told. Press releases announce gigawatts. Earnings calls confirm gigawatts. Equity research writes the gigawatts into models. But gigawatts are a paper unit until they clear a utility interconnection study, accept a transformer delivery, and pass a commissioning test. The classes of equipment that gate the last two steps run on industrial cycles measured in years, not quarters. The 2023 to 2024 order book for large power transformers is the binding constraint on what can plausibly come online in 2026 and most of 2027. Any forecast that ignores that order book and reads announced capacity as effective capacity is structurally wrong before the quarter ends.
So the "half canceled" meme was wrong on the magnitude, and probably wrong on the mechanism too. Some projects are slipping. SemiAnalysis names STACK Infrastructure among them. The actual rate of delay and cancellation is well below 50% on the filings, and almost entirely concentrated in projects whose transformer or switchgear delivery slot slipped, not in projects that lost their funding or their tenant. The corrective is "not half," not "zero." But the more durable lesson is about the next wrong number, not this one. As long as AI-assisted forecasting treats press-release gigawatts as grid-delivered gigawatts, the consensus will keep oscillating between "everything is canceled" and "nothing is canceled," and the industrial clock underneath both readings will keep ticking on the same multi-year cadence set in Chinese factories the analyst desks rarely call.
Watch the actual order book, not the forecast revisions. Watch the EIA transformer manufacturing reports, the PJM and ERCOT interconnection queues, and the 8-K filings from public datacenter operators when a named project misses a delivery window. Those are the sources a careful 2026 capacity number has to clear. Anyone selling a cleaner answer than that is reading the wrong documents.