The electricity math behind AI infrastructure is stark. Global data center power consumption ran roughly 415 terawatt-hours in 2024, according to the International Energy Agency. By 2030, the IEA projects that figure reaches approximately 945 TWh—a doubling in six years, growing at about 15 percent annually. That rate is more than four times faster than the rest of the economy.
In the United States alone, data centers consumed about 4.4 percent of total electricity in 2023. That share is on track to rise sharply by 2028, with hyperscale facilities pushing aggregate U.S. data center demand from roughly 50 gigawatts today toward more than 134 GW by 2030.
The driving force is AI compute. Unlike conventional servers, AI training and inference workloads run GPU clusters at maximum sustained load rather than cycling through peaks and troughs. A traditional server rack draws 5 to 15 kilowatts. A modern AI-optimized rack pulls more than 100 kW. Some hyperscale campuses are now designed around gigawatt-scale load targets—single sites that would rank among the largest power consumers in any country they occupy.
That density puts extraordinary pressure on a piece of equipment that most analysts overlook entirely: the power transformer. Transformers are the devices that step voltage up or down between generation, transmission and end-use systems. Every data center depends on them to connect to the grid and regulate incoming voltage. In an AI facility running continuously at or near full capacity, transformers operate under conditions they were never engineered to sustain.
Large-scale power transformers rated above 100 megavolt-amperes require custom engineering, specialized materials and precision manufacturing. Lead times for these units have stretched to two to four years—meaning a hyperscale developer who breaks ground today cannot receive the transformers needed to energize the facility until 2027 or 2028 at the earliest. Utilities facing grid upgrade requirements to serve new AI campuses run into the same queue. Regulatory approvals and substation construction add further delay on top of equipment procurement.
The constraint is structural, not a temporary supply-chain issue. Existing grid infrastructure was designed around dispersed, moderate loads—not the concentrated, always-on power draw of a gigawatt-scale AI campus. Upgrading that infrastructure requires new substation transformer installations, which carry their own permitting cycles and construction timelines. The equipment backlog and the permitting backlog compound each other.
The industry's response is to redesign around the constraint. AI data center developers are pursuing solid-state transformers—electronic devices that switch currents between AC and DC and handle higher voltages more efficiently than conventional copper-and-iron units. Unlike traditional transformers, solid-state versions can be manufactured with fewer custom components, offer faster production cycles and support full-DC system architectures that maximize efficiency at high power density.
At the grid-interconnection level, some facilities are already deploying 185 MVA super-grid transformers fed by 380 kV lines. That scale of equipment—purpose-built for a single data center campus—illustrates how far removed modern AI infrastructure is from the power systems that utilities built for general commercial and industrial use.
The transformer bottleneck has a direct effect on hyperscaler competitive positioning. When physical power infrastructure limits how fast new AI compute capacity can come online, it reinforces the advantage of operators who secured land, grid connections and equipment years in advance. Hyperscalers who locked in transformer orders in 2022 and 2023 are now commissioning facilities. Those ordering today are staring at 2028 delivery schedules. The lead time is not an engineering curiosity—it is a competitive moat, and it compounds with every quarter the queue lengthens.
