The slow parts: transformers, turbines, and the queue

A server ships in months. A new accelerator generation turns over on a roughly annual cadence. But none of that hardware runs until the physical plant around it exists — and the handful of parts that plant depends on move on a completely different clock, set not by silicon but by heavy industrial equipment that has to be custom-wound, forged, or slotted into a manufacturing queue years long. Three items set that clock more than any others: the substation transformer, the onsite turbine, and the transmission connection tying the whole campus to the grid. This is a page about queues, not chips.

The transformer queue

Every AI campus above roughly 100 MW needs its own substation, and at the center of it sits a large power transformer — a custom-engineered, oil-filled machine wound and tested to a specific voltage, MVA and impedance spec rather than pulled off a shelf. The taxonomy behind this catalog carries one figure on this part, and it’s the one that matters: lead times of 2–4 years, flagged as the single biggest schedule risk for a gigawatt-class campus. That’s longer than it typically takes to build the building the transformer feeds. The listed supplier base for this equipment skews European, Korean and Japanese — Siemens Energy, Hitachi Energy, GE Vernova, Hyundai Electric, Hyosung Heavy, LS Electric, Mitsubishi Electric — with several of the US names in the taxonomy sitting privately held, which is one reason the public-market conversation about domestic transformer capacity keeps routing through the same few foreign tickers.

The steel behind the transformer

A transformer’s core is a stack of thin electrical steel laminations that carries the magnetic flux linking its windings — get the grade of that steel wrong and the transformer runs less efficiently or takes longer to build. grain-oriented electrical steel is the standard material for large power and distribution transformers, and the taxonomy is specific about where the real constraint sits: it names GOES supply itself, not transformer manufacturing capacity, as the input constraint behind those multi-year lead times, and identifies Cleveland-Cliffs as the sole US producer of it. That’s the pattern worth remembering across this whole layer — the bottleneck on a finished, assembled product often sits one step further upstream, in a raw material most people never think to ask about.

Turbines: the same story, a different industry

Campuses that can’t wait for grid interconnection increasingly turn to gas turbines built or located onsite, behind the meter — either as simple-cycle units or paired with a heat-recovery steam generator for combined-cycle efficiency. The mechanism is straightforward: a turbine burns natural gas to spin a generator directly, producing firm power without waiting on a queue the campus doesn’t control. But that route has its own queue now. The taxonomy’s figure here is directional rather than a specific count: new-build turbine slots are sold out years ahead, and refurbished and aeroderivative units — turbines adapted from jet engines, prized for faster startup and modularity at campus scale — command a premium as buyers compete for whatever capacity a smaller pool of specialists can free up. The heat-recovery steam generators paired with the largest units add a second, separate capacity constraint behind the turbine order itself.

The interconnection queue

Behind both of those sits transmission and interconnection equipment — the HVDC gear, switching equipment, transmission structures and grid-enhancing technology that move bulk power from a generation source to a campus’s point of connection. Before a large new load can connect to the grid, a utility or grid operator studies the impact on the system, and that study queues behind every earlier request still being worked through — which is part of why AI load growth has turned transmission capacity and interconnection queues into one of the most visible bottlenecks in bringing new data-center capacity online. It’s also why grid-enhancing technology and advanced conductors, which squeeze more capacity out of lines that already exist, are being adopted as faster-to-deploy alternatives to building an entirely new transmission line from scratch. Onsite turbines are, in part, a way to sidestep this queue rather than wait in it — at the premium noted above.

Why none of this shows up in a chip roadmap

Compute silicon moves in product generations; this layer compounds in years. A transformer that’s two years late, a turbine slot that never opens up, or an interconnection study still queued behind a dozen earlier applicants all do the same thing to a campus: they cap how much load it can actually energize, no matter how many racks are sitting built and ready to plug in. None of the three parts on this page is exotic technology — a transformer, a turbine and a stretch of transmission line are all things the power industry has built for a century. What’s changed is the volume and timing of orders stacked on top of a manufacturing base that wasn’t sized for a wave of simultaneous, gigawatt-scale requests, and that’s a schedule problem money alone doesn’t solve any faster than the factories can wind the next core.

Where to go next

• Large power transformers — the single biggest schedule risk for a GW-class campus.
• Gas turbines — how a campus buys its way around the grid queue, at a price.
• Grain-oriented electrical steel — the raw-material constraint one step behind the transformer.