AI data centers are burning out aging power grid equipment and blocking industrial replacements, straining grid reliability.
The machines powering the AI boom are eating through electrical equipment at rates engineers never planned for—and the resulting shortage has grown severe enough that a single data center buildout can now strand an entire mining company's $800 million equipment order.
That figure comes from John Zuleger, president and CEO of Integrated Power Services (IPS), in an analysis published by energy publication Latitude Media. Zuleger's company—one of North America's leading electrical aftermarket service firms, with more than 115 locations and 2,000-plus skilled technicians—is the company hyperscalers call when something fails. Its emergency response work for data center clients surged 300% in 2026 alone.
What Zuleger describes is a maintenance crisis that extends far beyond silicon and server racks into the substations, switchgear cabinets, and transformer vaults that make AI infrastructure possible—and it is metastasizing into the steel mills, copper mines, and chemical plants that need the same equipment and can no longer get it.
Traditional industrial load curves follow predictable rhythms: seasonal air conditioning peaks, morning ramp-ups, nighttime valleys. That pattern allows electrical infrastructure to thermally recover between demand cycles. A substation transformer running conventional loads might last 30 to 40 years. AI training is fundamentally different. When a large GPU cluster executes a training run, tens of thousands of processors execute identical computational steps in lockstep—alternating between a high-draw compute phase and a lower-draw synchronization phase every few seconds. The result is a power oscillation that swings demand by tens of megawatts within minutes, with the fastest ramp rates documented at 1.9 times base power in under 250 milliseconds. That is not a slow seasonal ramp; it is a mechanical impact.
The consequence for transformers is governed by the Arrhenius relationship: every 6 to 8 degrees Celsius of sustained temperature increase above a transformer's rated thermal class approximately halves the insulation's remaining service life. AI training's rapid load oscillations create repeated localized hot-spot temperature spikes in transformer windings—not the stable elevated heat of a seasonal load peak, but a pattern of thermal cycling that inflicts qualitatively different and more damaging stress. Industry data published in June 2026 puts the accelerated degradation rate at two to three times faster than conventional seasonal load patterns. Equipment that might have lasted another decade under its designed load profiles is burning through its remaining service life at double or triple speed—with no external warning sign until something fails.
High-voltage substation transformers had lead times of roughly 50 weeks in 2021. By 2026, lead times have stretched to 140 to 160 weeks, with generator step-up transformers exceeding 160 weeks by Q1 2026, according to Wood Mackenzie data. High-voltage circuit breakers climbed to 125 weeks in the second half of 2025, up from 77 weeks in 2023. A project manager who places a transformer order today will not receive equipment until late 2028 at the earliest. For hyperscalers that have publicly committed hundreds of billions of dollars in AI infrastructure capital expenditure, this is no longer a planning note but a hard physical ceiling on expansion timelines.
Wood Mackenzie projects the US data center electrical equipment market will expand from roughly $20 billion today to $65 billion by 2030 as hyperscale AI construction accelerates. That demand competes with utilities and industrial customers for the output of a manufacturing base that has not meaningfully expanded capacity despite a 116% rise in transformer demand and a 274% increase in demand for generator step-up transformers since 2019. Roughly 80% of large US power transformers are imported, primarily from Mexico, South Korea, and European and Asian suppliers whose own order books are saturated.
Zuleger visited six copper mines where operators were lifting output against a base of under-maintained equipment. In one meeting, a senior procurement officer disclosed that the company had tried to place an $800 million switchgear order with an OEM it had worked with for decades. The OEM declined the order entirely because of data center build commitments. Hyperscalers are monopolizing the output of the manufacturers that every other capital-intensive industry depends on. Steel mills, copper mines, chemical plants, and utilities that need to replace worn-out transformers and switchgear now find themselves competing against Google, Microsoft, Amazon, and Meta—and losing.
The consequence is involuntary deferred maintenance at industrial scale. Reliability engineers flag at-risk equipment in capital planning meetings, then are told they cannot schedule a maintenance window because copper prices are at an all-time high, the blast furnace cannot stop, or the mine cannot interrupt production while the order backlog clears. "Reliability engineers may know what equipment is going to fail next, but they're unable to get the support to address it because of how valuable the product is," Zuleger told Latitude Media. "It's the same with data centers. It's all about speed to compute."
The arithmetic of deferral is unforgiving. Industry data from multiple infrastructure analysts puts the cost of an unplanned failure at roughly 10 times the cost of preventive maintenance; Zuleger's field experience at IPS places the ratio for severe incidents at 100 to 1,000 times. An $800 million switchgear order deferred for three years does not become an $800 million problem when something fails; it becomes a potentially catastrophic operational disruption on top of replacement costs.
A June 2026 survey by OxMaint found that 26% of utilities had not updated maintenance intervals since large AI load interconnection requests began flooding in—meaning the maintenance procedures governing substantial portions of the country's grid infrastructure were written for load patterns that no longer exist.
Some OEM manufacturers, behind on delivery schedules and under intense pressure from data center developers, are now shipping equipment at 70% completion to construction sites and then contracting with service firms like IPS to finish assembly in the field. Electrical infrastructure—switchgear lineups, transformer assemblies—is engineered to be manufactured in controlled factory environments: clean rooms, climate-controlled workspaces, semi-automated assembly that minimizes human error and dimensional variation. Field manufacturing cannot replicate those conditions. Equipment may meet short-term functional specifications while harboring subtle quality gaps that only manifest years later, when components begin to fatigue or when maintenance access reveals workmanship issues that would not have passed factory inspection.
"I'm concerned about the maintenance and uptime challenges that may come as this equipment gets used over a longer period of time," Zuleger said.
Hyperscalers with long-standing OEM relationships have largely locked in guaranteed orders. Tier-two and tier-three data center operators, and industrial customers who could not place orders with their preferred manufacturers, are sourcing from secondary markets and alternative suppliers—compounding the mixed-fleet problem as equipment from diverse manufacturers with varying quality standards and technical compatibility becomes integrated into the same facility.