An AI Data Center Can Be Built in 18 Months. The Transformer That Powers It Can Take Five Years.
Hyperscalers have gotten very good at pouring concrete and racking GPUs fast. What they can't speed up is the giant, custom-wound transformer that has to arrive before any of it draws real power — and its waitlist has gotten longer every year since 2020.
Two timelines, one shared ruler
Drag the slider to the year a transformer gets ordered, then press Run to watch both builds race from zero. Transformer weeks are rounded, blended averages of Wood Mackenzie's large-power-transformer and generator-step-up survey data as reported in industry press; the 78-week data-center shell timeline is held constant across years for comparison.
The plain version
Finishing the shell of a hyperscale AI data center — the concrete, the steel, the cooling pipes, the racks — now takes roughly a year and a half. That's fast. The part that actually lets the building draw power from the grid is not.
Every big data center campus needs at least one enormous transformer, often called a generator step-up (GSU) unit, that converts grid or power-plant voltage to exactly what the campus needs. These aren't off-the-shelf parts you can grab from a warehouse. Each one is custom-wound — copper coils wrapped tightly around a core built from a special magnetic steel — to the voltage, size, and cooling spec of one specific site, by one of a small number of factories worldwide that know how to build them.
A few years ago, ordering one took about a year. Today, industry surveys put the average wait at two and a half to three years, and the largest, most custom orders have been quoted out to four or five. Demand simply tripled almost overnight: utilities replacing 40-year-old grid equipment, wind and solar farms needing new substations, and now dozens of AI campuses are all bidding for slots at the same handful of factories.
So picture it: the data center's walls, racks, and chillers are ready to flip on in 18 months. The transformer that has to arrive before any of it can draw real power won't show up for another year or two after that. Analysts now estimate close to half of the AI data center capacity U.S. companies planned to open in 2026 is being delayed or quietly shelved — not because of a chip shortage, but because of a piece of equipment nobody thought to stockpile.
The expert version
The component at the center of this bottleneck is the generator step-up (GSU) transformer and its close cousin, the large power transformer (LPT) — custom-engineered, oil- or air-cooled units rated in the hundreds of megavolt-amperes, built to step voltage between a generation source or transmission tap (often 115–345 kV) and a site's own substation. Unlike distribution transformers, GSUs and LPTs are not stocked; each is wound to order from grain-oriented electrical steel (GOES) laminations and copper windings, a process that alone can run months, and the global GOES supply itself is concentrated among a handful of producers, Nippon Steel, voestalpine, and Baosteel among them.
Wood Mackenzie's lead-time survey shows the scale of the shift: pre-pandemic orders cleared in roughly 30–60 weeks; by the second quarter of 2025 the average had reached 128 weeks for LPTs and 144 weeks for GSUs, with some 2026 quotes for GSUs reported above 160. Unit prices have risen in step — Wood Mackenzie puts the increase at 77% for LPTs and 45% for GSUs since 2019 — while GSU order volumes are up roughly 274% over the same stretch, driven jointly by renewable-interconnection queues, utilities replacing an aging fleet (the average U.S. transformer in service is about 40 years old), and now AI campus demand competing for the same production slots. U.S. imports of high-voltage transformers from China rose from under 1,500 units in 2022 to more than 8,000 in 2025 as buyers chased any available capacity, per Wood Mackenzie data.
Bloomberg reporting citing Sightline Climate put large-transformer delivery at up to five years against an AI data center's typical sub-18-month shell-construction timeline — a structural mismatch analysts estimate puts close to half of the roughly 12–16 gigawatts of U.S. data center capacity slated for 2026 at risk of delay or cancellation. Hyperscalers are responding by pre-paying years ahead, taking equity stakes in transformer makers, and building captive generation (gas turbines, and in a few cases small modular reactors) to route around grid-interconnection equipment entirely.
Why it matters for tech + supply chain: The AI buildout's real bottleneck isn't GPUs anymore — it's a 19th-century grid component now backlogged years behind the buildings it's meant to power.
Why it matters for tech + supply chain: Capacity planning for AI compute now has to be done in GSU-lead-time units, not construction-schedule units — and the handful of factories that wind these units, plus the electrical-steel supply behind them, are becoming as strategically important as the fabs themselves.