A version of this story was published in the AI-Energy Nexus newsletter on September 23, 2026. Subscribe to get pieces like this — plus expert analysis, original reporting, and curated resources — in your inbox every Wednesday.
The AI infrastructure buildout needs a lot of everything, from transformers to gas turbines, and from grid capacity to electricians. As readers of this newsletter well know, the huge influx of demand, on top of the electrification that was already underway, is putting a strain on the power system and its supply chain.
But when I met with John Zuleger over the phone last month, he told me the strain is also complicating a less visible layer of the electrical system: maintenance.
Zuleger is president and CEO of Integrated Power Services, which is focused on the power sector’s so-called “aftermarket,” meaning everything that happens to equipment “after the switch is on.” IPS specifically provides life-cycle services for electrical assets, such as electric motors, generators, transformers, and switchgear, and its customers include utilities, data centers, and industrial plants.
Zuleger says years of underinvestment, both at the utility and industrial level, mean some transformers and other electrical equipment are already three-quarters of the way through their asset lives — meaning that load growth is straining infrastructure that is already worn out.
The following is an excerpt from our conversation, edited for brevity and clarity.
How’s the load growth affecting your customers?
John Zuleger: We work across many industries, and we can see that demand is spilling over: into steel mills and copper mines, for example. I was visiting six copper mines, and they’re all lifting their output against a base of mining equipment that has been under-maintained.
And the needs of our customers are changing. Because of the supply constraints, the demand to keep equipment running is increasing, and life extension is becoming a bigger interest. Our customers can’t afford to spend $50 million on a brand-new switchgear lineup, especially when hyperscalers are buying forward capacity two years in advance. In the visit to the copper mines I mentioned, I was in a meeting with a senior procurement person, and they shared that they were attempting to place an $800-million switchgear order with the OEM they’d been buying from for decades. The OEM told them they couldn’t take their order because of the data center build-out.
The constraint on the OEMs is having a cascading effect and creating a need for other strategies. Those can be as simple as getting a switchgear lineup, deenergizing it, and running an infrared energy analysis to identify the components reaching their end of life inside the cabinet.
Will there be long-term consequences of how load growth is affecting the maintenance process?
John Zuleger: Some OEMs are delivering equipment to those building the data centers at 70% completion, because they’re so behind in their schedules. So the OEMs are hiring our company to do last-mile manufacturing. This equipment is designed to be manufactured in a controlled factory, in a clean environment, and in a semi-automated way. I’m concerned about the maintenance and uptime challenges that may come as this equipment gets used over a longer period of time.
The other thing that is going to drive up maintenance complexity is mixed fleets. The hyperscalers have guaranteed orders for some of the big brands, but there are tier-two and tier-three data centers and other industrial customers who are scrambling to find equipment because they can’t get it from the main OEMs. Some of these data centers are being built with a mixed fleet. Or in the case of industrial customers, some have maybe been buying from one OEM for 20 years, and all their spares, parts, and equipment are on that brand. Now, what are they going to have to do? Are they going to have to introduce a different brand and a different equipment design? From a long-term maintenance standpoint, I think the plan is just starting to get written.
Is providing electrical and maintenance services for data center customers different?
John Zuleger: Data centers have to be up 99.999% of the time, which affects how you provide maintenance. Our direct business with aftermarket emergency response for data centers is up 300% this year; we have agreements with hyperscalers where we have to acknowledge their call on a failure code within four hours, and then deploy our team within four to 24 hours.
Because of that five-nines reliability, data centers have two circuits rather than one, and complete redundancy. So when you respond to an emergency, you have to deal with two different layers, as you figure out, “If I take this down, am I going to trip off standby power generators? Am I going to trip off UPSs? Can I isolate this area that’s in fault without triggering more faults?” And the cost of downtime in a data center is high, with some coming with penalties of $20 million-plus per day.
What’s the worst case scenario?
John Zuleger: Some of the customers we interface with the most are the reliability engineers at steel mills, copper mines, and so forth. They know where the risks of power failure are, and they go into capital planning meetings and say to the operator, “This is getting close to failure, and it could be catastrophic.” The operator has to make a decision, and they say, “Well, I can’t take downtime right now; the crack spreads are wide, and copper prices are at an all-time high.” 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. It’s the same with data centers. It’s all about speed to compute.
Relative to the cost of performing predictive maintenance, the cost of a failure can be 100 to 1000 times more expensive. So the worst-case scenario for our customers is more downtime that’s more costly.
But the thing that I’m really concerned about is longer term, as we become more dependent on AI and on its life-saving uses. What happens if downtime occurs then? This could move from being just a financial cost to having a human cost. These new AI data centers have not been operating long enough for us to really understand if we have the right maintenance plan for what happens when something goes wrong. That’s being learned right now.
