AI data centers are pulling the grid's 145-year-old transformer into the semiconductor era

The largest power transformers are still custom-wound by hand from copper coils wrapped around a steel core, using a design that dates to the 1880s, and customers now wait years for delivery, Ars Technica reports (Ars Technica). That wait is starting to dictate the pace of the AI build-out.

A demonstration completed in August 2026 by NC State University, the New York Power Authority, and EPRI points to one possible way out. The team ran a solid-state transformer, or SST, under real-world grid conditions for the first time (NC State). An SST does the same job, stepping voltage up or down, but uses high-frequency semiconductor switches, typically silicon carbide, instead of hand-wound copper coils.

Silicon-carbide switches operate at frequencies far above the 50 or 60 hertz of a conventional iron-core transformer. That lets an SST shrink the magnetic components to a small fraction of the size, which is why the unit can fit in a data-center electrical room and be moved with a forklift rather than a crane. The same property lets an SST convert AC grid power directly to the DC that AI server racks need, in a single box, without a separate rectifier stage.

Modern AI training clusters draw so much power per rack that operators are willing to change the entire building-side power architecture. NVIDIA has publicly described 800-volt DC architectures for "AI factories" (NVIDIA developer blog), and an SST is the device that makes those architectures practical at the building service entrance.

Conventional large transformers are built one at a time, often in dedicated plants, with multi-year lead times now routine. An SST is an assembly of standard semiconductor modules, built on a fabrication-style production line, swapped in pieces rather than replaced whole, and shipped in a fraction of the weight and volume. For a buyer planning a multi-gigawatt campus on a 12-to-24-month build schedule, the difference between a custom metal box and a modular semiconductor stack is the difference between a project and a plan. Silicon-carbide SSTs are still more expensive per megavolt-ampere than hand-wound iron-core units, and cost parity at scale has not been demonstrated outside the lab.

Electricity consumption from AI-focused data centers surged roughly 50% in 2025, the IEA reported, with about half of the growth in data-center demand globally met by new renewables (IEA, Energy and AI). That is the curve that conventional transformer suppliers cannot keep up with, and it is the first time a 145-year-old component has been forced to scale to hyperscaler build cycles.

AI data centers are the first customers with a workload dense enough to justify the SST swap, but they are not the only buyers feeling the multi-year lead-time problem. Utilities ordering replacement transformers for aging substations are competing with hyperscalers for the same limited manufacturing capacity, and SSTs do not address that market. The queue does not get shorter just because one customer class switched to a different device.

An SST is not a substitute for a 100-megawatt substation transformer, and the broader grid-strain problem, routine expansion and the replacement of aging units, is not what this technology solves. AI demand is pulling the semiconductor-built design into the distribution layer and the building service entrance, where power ratings are smaller and the buyer has the most to gain. It is not pulling it onto the transmission backbone, where the multi-year lead-time problem is most acute.

Three things to watch: the next round of NC State / NYPA / EPRI follow-up tests, the first hyperscaler to order an SST at multi-megawatt scale, and whether silicon-carbide supply, itself a constrained commodity, can keep up with a new customer class on top of electric vehicles and grid-tied solar inverters.