Why transformers can delay an AI data center

Why custom transformers, specialized materials and difficult delivery routes can become the hidden clock behind an AI campus.

7 min read

An AI data center can have land, permits, financing and a utility agreement, yet still wait on a machine that does no computing at all.

That machine is a transformer. It changes electricity from one voltage to another so power can move efficiently across the grid and then enter equipment safely. For a large campus, transformers may sit at several points between the transmission network and the server racks. If the required units are not available, a promised supply of electricity cannot simply jump around them.

This makes transformer availability a useful test of an AI project’s schedule. A power contract can be real while the equipment needed to deliver that power remains months or years away.

A giant electrical adapter with a very specific job

Electricity travels long distances at high voltage because that reduces losses. Buildings and machines use lower voltages. Transformers handle the handoffs. The U.S. Energy Information Administration describes substations and transformers as essential parts of the journey from power plants, through high-voltage transmission, to the distribution lines serving customers. EIA: How electricity reaches consumers

Think of a transformer as an enormous electrical adapter. The analogy is useful because it explains the job: the electricity must arrive in a form the next part of the system can use. It stops being useful if it makes the machine sound interchangeable. A large grid transformer is designed around a particular voltage, capacity, electrical behavior, protection scheme and installation.

A data center may also use smaller transformers inside its own electrical system. “The transformer” is therefore not necessarily one object at the gate. The exact arrangement varies by utility connection and facility design. Losing one critical unit can matter even when other transformers elsewhere in the project are ready.

Why the biggest units are not sitting on a shelf

The Department of Energy says large power transformers are expensive, difficult to move and usually custom-built. Procurement can involve qualifying a manufacturer, designing the unit, producing and testing it, and arranging special transportation. In its July 2024 resilience report, DOE said conversations with domestic producers indicated lead times of up to and beyond 36 months. That was a report of market conditions, not a universal promise that every order now takes exactly three years. DOE: Large Power Transformer Resilience

Another DOE assessment published in January 2025 reported average lead times across transmission and distribution equipment ranging from 51 weeks for distribution transformers to 137 weeks for power transformers. It also reported price increases of 37% to 80%. DOE traced the pressure to limited manufacturing capacity, labor shortages and constrained supplies of grain-oriented electrical steel and copper. DOE: Supply Chain Readiness Levels

Those figures describe broad equipment markets, not a quote for a particular campus. A project’s actual schedule depends on the required design, manufacturer, testing, transport route and order position. The responsible conclusion is that transformers can be a long-lead item, not that a public opening date is automatically wrong.

The last miles can be harder than the first thousand

Manufacturing is only part of the problem. Large transformers can weigh hundreds of tons. DOE notes that the final five to ten miles of delivery can be especially difficult and expensive. Bridges, turns, rail access, road strength, clearances and crane plans can all shape the route.

That helps explain why “equipment ordered” and “site energized” are separate milestones. A finished transformer still has to reach its foundation, be installed, connected, filled or prepared as its design requires, tested and accepted. A delay at any step can move the date when the campus receives full service.

The same distinction applies inside a data center. Completing the building shell does not prove that its electrical path is commissioned. The racks cannot use grid power until the whole chain works under the conditions promised by the design.

Why not keep a spare for every project?

Utilities do keep spares and share equipment through some resilience programs. But a spare is most useful when it matches the failed or delayed unit closely enough to do the job. DOE’s Transformer Resilience and Advanced Components program says large transmission transformers are often tailored to customer specifications, are not readily interchangeable and cost enough to make extensive spare inventories impractical. DOE: Transformer Resilience and Advanced Components

This is closer to replacing a custom bridge section than buying another household adapter. Standardization could improve interchangeability, but utilities also need equipment that fits existing voltage levels, protection systems, space and operating requirements.

DOE is supporting more flexible, modular and adaptable transformer designs. Those efforts are intended to improve reliability and make equipment useful in more locations. They are development strategies, not evidence that today’s custom equipment has already become a universal plug-in product.

AI growth makes an old supply-chain problem more visible

Transformer shortages did not begin with generative AI. Utilities also need equipment to replace aging assets, connect new generation, rebuild after disasters and serve ordinary growth. Data centers join that competition because some request very large amounts of power on aggressive schedules.

Lawrence Berkeley National Laboratory estimated that U.S. data centers used about 176 terawatt-hours of electricity in 2023. Its 2024 report modeled a wide range of 325 to 580 terawatt-hours in 2028. That is a forecast range built from assumptions about equipment shipments, computing demand and efficiency. It should not be treated as a meter reading from the future. Berkeley Lab: 2024 United States Data Center Energy Usage Report

The connection between that growth and transformers is physical. More electricity demand can require new substations, upgraded lines and voltage-changing equipment. Yet the exact upgrade is site-specific. A campus beside strong infrastructure may need different work from one in an area with limited capacity.

What can shorten the wait?

Earlier planning can let a utility order equipment before every other project detail is finished. Standard designs may allow larger production runs. Staged service can bring a smaller first phase online while later infrastructure is built. In some arrangements, large customers may construct or fund parts of the connection themselves.

Each option moves risk rather than erasing it. Ordering early can strand equipment if the customer never arrives. Standardization can conflict with site requirements. Partial service supports less computing than the final campus target. Customer-funded construction still needs engineering coordination and approval.

Berkeley Lab’s 2026 report on faster large-load connections identifies advance utility procurement, clearer interconnection processes and customer-built facilities among potential approaches. It also stresses uncertainty: utilities can overbuild for speculative requests or underbuild for real demand. Berkeley Lab: Speed to Power

How to read the next campus announcement

When a developer says a project has secured power, ask what has been secured. Is there a supply contract, an approved service agreement, completed network studies, ordered transformers, a construction schedule or an energized connection? Each is meaningful. None automatically proves the others.

Also ask whether the announced megawatts describe the first operating phase or the eventual campus. A utility may be able to serve part of the load before the final substation and transformer work is complete. Reporting the ultimate number as present-day capacity can hide that sequence.

The public effect is not automatic. A transformer delay does not by itself establish higher household bills, local outages or a canceled project. Those outcomes depend on utility planning, contracts, regulation and who pays. Our explainer on data-center grid-upgrade costs follows that separate question.

The central lesson is simpler: electricity has to be transformed before it can become computing. AI companies can buy chips faster than the grid can always manufacture, move and install the equipment serving them. The least glamorous machine in the project may control when the most expensive machines can finally turn on.