When talking about bottlenecks in AI data centers, the previous pieces covered GPUs, HBM, CoWoS, optical modules, liquid cooling, and power supplies. But what is really causing a batch of already approved AI data center projects to delay construction until 2026 is not chips, nor cooling, but something almost nobody has ever cared about: transformers and switchgear.
First, the conclusion: multiple reports in the U.S. now show that AI data centers have already shifted the focus from “not enough GPUs” to “not enough electrical equipment.”
Industry estimates say that in the U.S. alone, about 7GW of AI data center capacity has been delayed or canceled because transformers and switchgear could not be obtained. The delivery lead time for generator step-up transformers has already stretched to 144 weeks, nearly three years. The overall lead time for large power transformers has also lengthened from more than a year before the pandemic to two to three years today. In order to get equipment, utilities have had to place orders five years in advance. The severity of this bottleneck is already on the same level as CoWoS and HBM.
Why have transformers suddenly become scarce? The reason is almost exactly the same as with HBM: demand exploded, and capacity is fixed.
Transformers are a heavy-asset manufacturing business. A large power transformer weighs over 100 tons, factory expansion takes years, and the key raw material—grain-oriented electrical steel—can only be made by a handful of suppliers worldwide.
Switchgear, circuit breakers, and busbars are in a similar situation. For decades, these products grew at only single-digit annual rates in stable industries, and then AI suddenly doubled order volumes. Capacity simply cannot keep up.
Transformers and electrical steel are now the HBM of power equipment.
The deeper change is that the electrical architecture of data centers itself is being rewritten.
The traditional power path in a data center looks like this: medium-voltage AC comes in from the grid, is stepped down to low-voltage AC through a transformer, then converted by the UPS into DC to charge batteries, then inverted back into AC, and finally rectified again into DC inside the server power supply. Along the way, there are five or six AC-DC-AC conversions, each with losses, and each taking up space.
Once GPU power consumption reaches the kilowatt level and racks reach the hundred-kilowatt level, this architecture is no longer efficient in either space or power.
The industry’s new direction is “DC-ification,” and the core device is called SST—a Solid-State Transformer. Using silicon carbide power semiconductors, it directly converts medium-voltage AC in one step to an 800V DC bus, supplies power straight to the racks, and eliminates all the intermediate conversions.
The consensus in academia and industry is that this medium-voltage AC to 800V DC architecture will become standard in next-generation AI data centers.
Companies making SiC devices such as Microchip, Infineon, and onsemi, as well as companies handling data center power distribution such as Vertiv, Eaton, and Schneider, are all pushing their own SST solutions. You can even see in Microchip’s promotional materials that they directly tie SST to GPU token generation per second—power efficiency determines how much inference you can run within the same power budget.
SST is not an optional choice, but an inevitability forced by GPU power density.
A GB200 NVL72 rack consumes 120 kilowatts. The next-generation Rubin rack is expected to climb to several hundred kilowatts. Beyond that, traditional AC power simply cannot fit that much copper and so many transformers. Only a DC architecture can free up the space.
This means that over the next few years, a large share of new data center purchases will shift from traditional low-voltage AC equipment to medium-voltage AC plus solid-state transformers plus low-voltage DC busbars. This is a quiet but structural reconfiguration.
So when looking at AI infrastructure, GPUs and HBM are the visible main line, while transformers, switchgear, and SST are the other main line on the hidden side. GPUs can be accelerated quarter by quarter; transformers can only expand capacity year by year. What will really decide when an AI data center can come online today is not Nvidia’s production schedule, but the delivery slips from power equipment manufacturers.
If you look at this chain without looking at the electrical equipment, you are not seeing the whole picture.