The AI Boom's Bottleneck Isn't Chips. It's a Steel Box With a 3-Year Waiting List.
Hyperscalers can buy all the GPUs money can buy. What they can't buy quickly is the machine that steps grid power down to something a data center can use — and the wait for one just stretched past three years.
Two build timelines for the same AI campus, run side by side
Illustrative, rounded timeline for a hypothetical large AI campus: GPU cluster procurement and installation in ~9 months (a commonly cited hyperscaler build cadence), against a custom high-voltage grid transformer delivered in ~42 months (the middle of Wood Mackenzie's 3-5-year range for large custom units). Real projects vary by vendor, region, and grid queue position — this is one plausible case, not a specific company's numbers. Drag the slider or press Run to move both timelines forward and watch the idle-GPU gap grow.
The plain version
Think about everything an AI data center needs: land, chips, cooling — and, easy to forget, a way to actually get power from the grid into the building. That job belongs to a transformer: a giant steel-and-copper machine, sometimes the size of a house, that takes electricity from high-voltage transmission lines and steps it down to a voltage the building's equipment can use. Every data center, every new factory, every wind farm needs one to connect to the grid.
Here's the problem: these machines used to take about a year or two to order and build. Now the wait is three to five years, sometimes longer for the biggest, most customized units — longer than it takes some companies to plan, build, and fully stock an AI data center with chips. So you get a strange picture: a company with billions of dollars of GPUs sitting in a warehouse or an unpowered building, unable to turn anything on, because the one part that connects them to the grid isn't ready yet.
Why can't manufacturers just build more, faster? Two reasons. First, the steel: transformers use a special magnetic steel, precisely engineered so it doesn't waste energy as heat, and only a handful of mills on Earth — mostly in Japan, South Korea, and a few other countries — know how to make it well. Second, the labor: winding the copper coils inside a transformer's core is still largely a hand-built, highly skilled trade, and there simply aren't many people trained to do it. Neither constraint can be solved by throwing money at a factory overnight.
Manufacturers are racing to catch up — Hitachi Energy alone is putting over $1 billion into new U.S. transformer plants — but new factories take years to come fully online too. Until then, the AI boom's real speed limit isn't how fast Nvidia can ship chips. It's how fast the world can make steel boxes.
The expert version
A power transformer changes AC voltage between circuits via electromagnetic induction, letting electricity travel efficiently over long transmission lines and then step down to usable voltage. The units that matter here — large power transformers (LPTs) above roughly 100 MVA, and generator step-up (GSU) transformers that connect new generation and large loads to the grid — are custom-engineered to a utility's exact voltage, impedance, and cooling spec, not stocked as interchangeable inventory. There is no warehouse of spares to draw down when demand spikes.
Two structural constraints set the ceiling on output. The core requires grain-oriented electrical steel (GOES): silicon steel processed via secondary recrystallization so its magnetic domains align along the rolling direction, minimizing eddy-current and hysteresis losses. That metallurgy is mastered by only a handful of integrated steelmakers worldwide — POSCO, Nippon Steel, Baowu/Baosteel, ArcelorMittal, and JFE Steel account for the large majority of global GOES capacity — and adding capacity means multi-year furnace and rolling-mill investment, not a new assembly line. Separately, the copper windings inside an LPT core are still wound largely by hand by specialized coil-winders, a trade with a thin, aging workforce across the US and Europe.
The result shows up directly in delivery times. Wood Mackenzie's Q2 2025 survey put average lead times at 128 weeks for standard power transformers and 144 weeks for GSU units, with custom high-voltage orders running 3 to 5 years — up from roughly 24-30 months as recently as 2019 — and modeled a 30% national supply deficit for power transformers.
Demand is compounding the shortage: AI data center power draw is pulling GSU and substation-transformer orders up sharply, on top of renewable-interconnection and grid-modernization backlogs already in the queue. In June 2026, FERC issued show-cause orders directing all six major US grid operators — PJM, MISO, SPP, CAISO, ISO-NE, and NYISO — to justify or overhaul how they interconnect large loads like data centers, a tacit admission that interconnection queues, not generation, are now the binding constraint in several regions. Manufacturers are responding with real capital: Hitachi Energy's $1 billion US buildout includes a $457 million large-power-transformer plant in South Boston, Virginia, set to become the country's largest such facility when it opens in 2028 — supply that lands years after this generation of AI campuses needs it.
Why it matters for tech + supply chain: the constraint on the next wave of AI infrastructure isn't semiconductor fabs anymore — it's 100-year-old steel-and-copper hardware with a design-and-build cycle measured in years, not quarters.
Why it matters for tech + supply chain: it reframes AI capex risk — GPU procurement is a chip-supply problem with multiple sourcing paths, but grid interconnection is a single-supplier-per-region, multi-year-lead-time problem that capital alone can't compress, which is why hyperscalers are increasingly building or buying on-site generation rather than waiting in the interconnection queue.