Power remains the first gate for an AI data centre. It is no longer the whole underwriting case.
Bernstein argues that mechanical, electrical and plumbing construction, equipment availability and skilled labour are becoming increasingly important constraints as projects secure generation or grid pathways. Its August work estimated that a 100MW AI data centre requires about 90 MEP craft workers and that, at recent peak recruitment rates, US annual data-centre construction could rise from roughly 12GW in 2026 to 35GW by 2030.
Supplier capital is moving toward the same bottleneck. Vertiv said in March that two South Carolina facilities would raise its regional capacity for integrated infrastructure solutions by roughly sevenfold. The company markets prefabricated power and cooling systems as a way to move integration and testing into factories before equipment reaches a site. Eaton, meanwhile, announced more than $242 million of investment in Arkansas in September to double US capacity for Fibrebond modular electrical enclosures used in data centres and other critical infrastructure.
The economic unit moves one step downstream. A megawatt reservation is not billable compute. Revenue begins after substations, transformers, switchgear, busways, backup systems, liquid cooling, controls and racks have been integrated, tested and energised.
A megawatt reservation becomes valuable only after conversion
Vertiv's MegaMod HDX is designed as a prefabricated liquid-cooling system for high-density deployments and supports blocks of infrastructure up to 10MW. Eaton says its Flexnode modular data-centre product can reduce deployment schedules by about 35% on average. Those are vendor estimates, not independent measures of project returns, but they show what the suppliers are selling: less site work and greater schedule certainty, not merely more equipment.
That matters because AI hardware raises the cost of delay. A completed shell with a power path does not earn on accelerators that cannot yet be installed and cooled. Moving electrical rooms, cooling modules and other repeatable systems into controlled manufacturing environments can reduce contractor interfaces and allow developers to add commissioned blocks as power arrives.
Schneider Electric describes the same pressure from a broader portfolio position, citing long lead times for critical electrical equipment and shortages of electricians, plumbers and HVAC technicians. The point is no longer attributable to one vendor. Vertiv is adding integrated-infrastructure capacity; Eaton is doubling a modular-enclosure footprint; Schneider is pushing prefabricated AI power and cooling architectures.
For shareholders in those suppliers, the opportunity is not simply a rising equipment bill per data centre. It is a larger share of the schedule-risk budget. Customers that value an earlier revenue date can justify paying for qualified, pre-integrated systems and manufacturing capacity that reduces on-site complexity.
Power rights remain scarce and strategic. Once they are secured, however, factory-built electrical and cooling capacity increasingly determines which projects turn those rights into revenue first.
Sources
- Investing.com - Bernstein on US data-centre construction and MEP labour
- Vertiv - Manufacturing expansion for integrated infrastructure, power and rack systems
- Vertiv - MegaMod HDX modular liquid-cooling infrastructure
- Eaton - Flexnode modular data-centre offering
- Eaton - More than $242 million modular-enclosure manufacturing expansion
- Schneider Electric - AI infrastructure growth depends on execution, not demand