We spent a recent post correcting the most-quoted number in the AI power story: the 2,061 GW in Lawrence Berkeley Lab's interconnection queues is a generator queue, not the data-centre queue, and it leaks about 98% of its capacity rather than delivering it. That correction mattered because the queue is the part of the story everyone quotes. This post is about the part nobody quotes — what happens after a project wins its grid position. Between the interconnection agreement and energized silicon sits a purchase order for very large steel objects, and the manufacturing queue for those objects has quietly become as slow as the regulatory one.

The lead times, measured

The reference dataset here is Wood Mackenzie's transformer supply-chain work, most recently its Q2 2025 survey, and the numbers are unusually consistent across outlets that independently reported them. Average lead time for standard power transformers: 128 weeks — about two and a half years. Generator step-up transformers, the units that connect generation to the transmission system: 144 weeks, with some specialized orders running to four years. NERC's own tracking shows transformer lead times first crossed the 120-week mark in 2024 and kept climbing through 2025. Before the pandemic, the same equipment shipped in 12 to 16 weeks. That is not a market tightening; that is a different universe of delivery risk attached to the same nouns.

Transformers are not even the worst of it. High-voltage circuit breakers averaged around 125 weeks in recent surveys, up from 77 weeks in 2023. Contractor-side reporting on data-centre-class electrical gear runs from medium-voltage switchgear near 44 weeks to power circuit breaker switchboards past 84 weeks. Nexans, the cable manufacturer, notes that ABB and Siemens Energy — two of the world's largest transformer makers — have reported high-voltage lead times of 48 to 60 months, with new large orders effectively quoting 2030–2031 delivery. And Wood Mackenzie assessed a supply deficit of about 30% in power transformers against current demand: roughly one unit in three that the market wants this year does not exist.

Why the factories cannot respond

Three facts explain most of it. First, materials: the core of every transformer is grain-oriented electrical steel (GOES), a specialty product whose US supply runs through a single domestic producer, Cleveland-Cliffs, with plants in Pennsylvania and Ohio. GOES prices have nearly doubled since 2020, and copper — the other body-and-blood input — is up more than 50% over the same period. Second, imports: roughly 80% of the large power transformers the US deploys are built abroad, mainly in Mexico and South Korea, which converts every global freight and trade-policy event into domestic grid risk. Third, demand shape: since 2019, demand for GSUs is up 274% and for power transformers up 119%, driven by three waves at once — AI data centres, industrial electrification, and replacement of an aging installed base — none of which share a peak, all of which are currently active. Domestic capacity expansion is underway but Wood Mackenzie does not project meaningful relief before 2028.

The price record corroborates the queue, because shortages of manufactured goods show up in invoices before they show up in surveys. Since 2019: power transformer prices +77%, distribution transformers +78% to 95%, GSUs +45%. Morgan Stanley's chief fixed income strategist, reviewing these dynamics, put the conclusion plainly: power supply constraints are "not peripheral frictions but central to the entire AI infrastructure buildout." When the bond desk is writing about transformers, the constraint has stopped being an engineering footnote.

What it does to a data-centre schedule

Bloomberg's 2026 read of the US build-out made the bottleneck visible at market scale: of roughly 12 GW of data-centre capacity slated to come online in 2026, only about a third was under active construction, with shortages of transformers, switchgear and batteries cited as a primary reason — and industry analyses warned that equipment shortages could delay or cancel a large share of planned developments. The arithmetic that produces this is simple and unforgiving. A large campus build takes 18–36 months. The substation transformers it needs take two and a half years or more if ordered on day one. A transformer ordered when construction starts is a transformer that arrives a year after the servers do — and the servers cannot run at all without it, because a hyperscale load connects at transmission voltage and steps down through exactly the equipment that is backlogged.

The market's response is a leading indicator by itself. "Slot reservation" — signing framework agreements with two or three manufacturers, deposits attached, to hold production capacity before financing closes — has become standard practice for large loads, because production slots are allocated first-come, first-served and cannot be bought back later at any price. Procurement, which used to sit after financial close, has moved to the front of the plan: Wood Mackenzie's analysts now frame transformer planning as something that must precede project approval, or the project inherits an extra two to four years it has no lever to remove.

What this means if you are buying compute

Ask compute providers where their transformers are, not just where their GPUs are. A capacity roadmap is only as good as the energization date behind it, and energization is set by steel on a factory floor in Ohio, Monterrey or Ulsan long before it is set by GPU allocations. The questions that separate real pipelines from slide decks: which substation serves the site, when were the transformers ordered, and from whom. A provider that can answer those with dates and PO numbers has a schedule; one that cannot has a rendering.

Small is still a schedule strategy — now twice over. The interconnection post made the queue-side case for distributed capacity: below the large-load thresholds, process is measured in months. The equipment side compounds the same advantage, because distribution-class transformers and commercial switchgear, while stretched, are not in the four-year tier that transmission-class hardware occupies. A network of small sites on existing service buys its way around both queues at once. This is the concrete, unromantic version of the argument for decentralized compute: not ideology, but the observation that the binding constraints — queue position, transformer slots, energization dates — are all cheaper to avoid than to wait out.

Distrust any capacity announcement without an equipment lead time. The sector already learned to ask whether gigawatt figures are requests or agreements. The next discipline is asking when the steel was ordered. Every announced buildout silently assumes transformers can be procured on a 2021 timetable, and every one of those assumptions is now wrong by a factor of eight.

The 2026 reality check

The caveats: lead-time figures are manufacturer and distributor survey data, self-reported and skewed toward large orders — a standard dry-type unit for a mid-size facility still ships in weeks, and quoting "128 weeks" at a 500 kW load would be wrong. The 12–16-week pre-pandemic baseline refers to standard units; large 100 MVA+ equipment took longer even in 2020. And the numbers here are a snapshot from the Q2 2025 survey round, with 2026 reporting suggesting conditions stayed at or above those levels rather than easing; re-check current surveys before pricing anything. None of the caveats changes the direction: the post-queue hardware wall is real, measured in years, and it is the line item missing from nearly every public capacity model.