In 2025, 57% of data-center projects tracked by JLL were delayed by at least three months. The average construction period for a 50MW facility was about 18 months, yet some critical materials were being ordered as much as 24 months before they were needed on site.
That mismatch captures an increasingly important feature of the AI infrastructure buildout. The constraint is no longer simply whether another data-center building can be erected. It is whether enough power equipment, cooling capacity, specialist labor and commissioning capability can be assembled at the same location, on the same schedule, to turn the building into usable compute capacity.
- AI data centers increase construction intensity faster than building footprint because rack density pushes more project value into electrical, mechanical and cooling systems.
- Critical equipment procurement can now run on timelines as long as, or longer than, the physical construction of the facility itself.
- The scarce construction resource is increasingly qualified mission-critical execution: MEP integration, modular fabrication, controls and commissioning rather than generic building capacity.
The mismatch: an 18-month building with a 24-month supply chain
Data-center demand is expanding rapidly enough to expose constraints that were less visible during the conventional cloud buildout. The International Energy Agency estimates that global data-center electricity consumption could rise from roughly 485TWh in 2025 to about 950TWh by 2030. AI-focused facilities are expected to account for a disproportionate share of that increase.
The change is occurring unusually quickly for a physical infrastructure cycle. Software demand can scale in months. Semiconductor roadmaps can change annually. High-voltage electrical infrastructure, utility interconnections and specialist construction capacity operate on much slower timelines.
JLL estimates that 97GW of new data-center capacity could be added globally between 2026 and 2030 under its base outlook, with substantially more possible if the most aggressive AI plans proceed. The difficulty is converting announced megawatts into energized megawatts. Power availability has already overtaken traditional variables such as land cost in many site-selection decisions because access to the grid can require multi-year waits.
This changes the role of construction. During the earlier cloud cycle, construction was largely an enabler of demand. In the AI cycle, construction and its upstream equipment supply chain can determine how quickly that demand becomes commercially usable infrastructure.
AI does not scale a conventional data center linearly
A common simplification is to treat an AI data center as a conventional facility with more servers. The physical differences are larger than that description suggests.
Traditional enterprise and cloud racks commonly operated in a range of roughly 5–15kW per rack. AI training clusters already operate at several multiples of those densities. JLL places current AI training configurations broadly around 40–100kW or more per rack, while newer reference architectures have moved beyond that range. Schneider Electric reference designs for Nvidia GB200 and GB300 systems reach approximately 132kW and 142kW per rack, respectively. A 2026 Schneider Electric and AMD reference architecture for Helios-class systems supports racks at roughly 246kW.
Power density is also heat density. Concentrating more compute into each rack increases the electrical current that must reach the IT equipment and the heat that must leave it. Raising rack density therefore changes several construction systems at the same time rather than simply requiring a larger electrical connection.
| Construction issue | Conventional data center | AI-optimized data center | Construction implication |
|---|---|---|---|
| Rack power density | Typically low- to mid-teens kW per rack in traditional configurations | 40–100kW+ increasingly common for AI training; leading designs exceed 100kW | More electrical capacity must be delivered through less physical space |
| Cooling | Predominantly air-based cooling | Hybrid and direct-to-chip liquid cooling increasingly required | Additional CDUs, piping, pumps, heat rejection and water-management systems |
| Electrical distribution | Lower rack-level current density | Higher-capacity switchgear, UPS, PDUs, busways and cabling | Greater MEP content and more complex coordination |
| Procurement | Equipment procurement more closely aligned with construction sequence | Critical equipment may need commitment well before installation | Design, procurement and construction overlap more aggressively |
| Site selection | Connectivity, land, latency and operating cost | Speed to power increasingly dominates | Construction feasibility becomes dependent on utility infrastructure |
| Delivery model | Large share assembled conventionally on site | Growing use of prefabricated power, cooling and IT modules | Part of construction capacity migrates into manufacturing facilities |
| Handover | Building and IT commissioning | Integrated electrical, cooling, controls and load commissioning | Completion of the shell is increasingly distant from operational readiness |
Source: Sector Foundry Research synthesis based on IEA, JLL and Schneider Electric. Ranges vary by facility design and workload.
The construction value pool is moving into MEP
Data centers were already unusually MEP-intensive buildings before the AI cycle. The shell consists of familiar construction work: earthworks, foundations, structural steel, concrete, roofing and interior space. Much of the differentiation sits behind the walls and above the racks.
The electrical system includes transformers, switchgear, uninterruptible power supplies, power distribution units, busways, cables, backup generation and protection systems. Mechanical infrastructure includes chillers or other heat-rejection equipment, pumps, air-handling systems and controls. AI adds another layer of thermal equipment as direct-to-chip liquid cooling introduces cold plates, coolant distribution units, secondary cooling loops, additional piping and fluid-management systems.
Korean industry research reviewed for this article estimates that MEP can account for roughly 50–60% of conventional data-center construction work and could move toward 60–70% in high-density AI facilities. Those percentages should be treated as indicative rather than universal: project scope, land treatment, utility interconnection and owner-supplied equipment differ substantially. The direction of travel is more important than the exact ratio.
Higher density pushes more dollars and engineering hours toward systems that require specialist design and installation. More expensive switchgear does not only raise equipment cost. It affects electrical-room layouts, protection schemes, cable routing, testing and commissioning. Liquid cooling does not only add cooling hardware. It creates interfaces between facility water systems, coolant distribution, rack architecture, leak detection and control software.
The resulting construction problem is less modular at the system level even when individual components become more modular. Every subsystem has to operate together under a load profile that may be far denser than the facility designs contractors were executing only a few years earlier.
A finished shell can still be an unfinished data center
The most consequential construction constraint may sit outside the building. A data center without an energized electrical connection is effectively unfinished regardless of how much structural work has been completed.
JLL reported an average global lead time of roughly 33 weeks for major data-center equipment in its 2026 outlook, rising to around 42 weeks in the United States. Selected items are being secured much earlier. Large transformers are an extreme case: Reuters reported in July 2026 that some U.S. transformer lead times had moved beyond 160 weeks as utilities, industrial projects and data centers competed for limited manufacturing capacity.
These timelines change the critical path. Equipment procurement can precede detailed construction progress and, in some cases, parts of final design. Developers increasingly reserve manufacturing slots, order long-lead electrical equipment early and phase campuses so that the first usable capacity can be energized before an entire site is complete.
That introduces a different type of execution risk. Ordering equipment earlier protects the schedule but reduces design flexibility. Waiting for certainty preserves flexibility but can push energization back by quarters or years. AI hardware roadmaps are moving in the opposite direction: rack architectures and power densities are changing rapidly enough that a facility designed around one accelerator generation may need to accommodate materially different equipment by the time it opens.
The construction schedule has therefore become a negotiation between two clocks. Electrical infrastructure rewards early commitment. AI hardware rewards late commitment.
Construction is moving off site because the schedule has nowhere else to go
One response is to move more work from the project site into controlled manufacturing environments.
Schneider Electric’s current AI data-center architecture increasingly uses prefabricated medium-voltage equipment, electrical houses, low-voltage power trains, power skids, modular IT pods, packaged cooling equipment and selected piping assemblies. The objective is not simply lower labor cost. Factory integration allows multiple workstreams to proceed in parallel and reduces the amount of coordination that has to occur on a congested construction site.
This begins to blur the line between construction and industrial manufacturing. A conventional project moves material to a site and assembles much of the system there. A modular AI project can move a larger portion of fabrication, wiring, piping and testing upstream, then transport integrated assemblies to the site for final connection.
The shift is already visible in the contractor ecosystem. Comfort Systems USA, one of the largest U.S. mechanical and electrical contractors, reported that technology markets represented 58% of year-to-date 2026 revenue in its August investor presentation. Modular activity accounted for 17%. Its total backlog reached approximately $14.1 billion at the end of the second quarter, compared with roughly $8.1 billion a year earlier.
Those figures are not a clean measure of AI data-center economics; Comfort serves multiple technology and industrial markets. They are useful evidence of the scale at which specialist mechanical, electrical and modular capacity is being absorbed.
Prefabrication can relieve the field-labor constraint, but it does not make the physical bottleneck disappear. It moves part of it upstream into factory floor space, engineering capacity and component supply.
The scarce resource is qualified execution, not generic labor
Construction labor is often discussed as if one craft hour were interchangeable with another. Mission-critical infrastructure does not work that way.
A hyperscale data center requires electrical crews familiar with redundant power architecture, mechanical contractors capable of high-density thermal systems, controls specialists, commissioning engineers and project managers able to coordinate owner-furnished IT equipment with facility infrastructure. The relevant capacity is therefore narrower than total employment in the construction industry.
Reliability requirements reinforce that distinction. A contractor that has completed multiple hyperscale projects accumulates knowledge about sequencing, tolerances, testing procedures, commissioning failures and owner standards that is difficult to reproduce instantly by adding labor. Data-center reference history can consequently become a practical barrier to entry even when the underlying construction technologies are widely understood.
The backlog of large specialist contractors illustrates the pressure. EMCOR reported remaining performance obligations of approximately $17.1 billion at the end of the second quarter of 2026, up nearly 44% from a year earlier. Comfort Systems reported similarly sharp backlog expansion. Neither backlog should be interpreted as a pure data-center order book, but both indicate that the same mechanical and electrical capacity required by AI projects is already being pulled by a wider infrastructure investment cycle.
This is where AI changes construction economics most meaningfully. The constraint is not a national shortage of people capable of pouring concrete. It is simultaneous demand for the relatively small set of suppliers and contractors able to deliver, integrate and validate high-density infrastructure on hyperscaler schedules.
Korea is a useful stress test for announced capacity
Korea provides a useful example because the prospective AI data-center pipeline is now much larger than the country’s recent data-center construction base.
In June 2026, the Korean government outlined private-sector plans associated with its large-scale AI and semiconductor infrastructure initiatives that included approximately 18.4GW of prospective AI data-center capacity. The broader investment plans cited by the government total roughly KRW 550 trillion across AI, semiconductor and related projects. Subsequent policy materials explicitly identified timely electricity and water provision as prerequisites for execution.
The distinction between planned capacity and commissioned capacity is critical. An announced 1GW campus is not equivalent to 1GW of operating infrastructure. The project must secure a site, permits, grid or on-site generation, transformers and switchgear, cooling equipment, specialized contractors and an anchor tenant before construction can translate into usable compute.
Korea does have an experienced base of large contractors. Broker research compiled from company disclosures shows repeated data-center references across GS E&C, Hyundai E&C, Samsung C&T and DL E&C, including large domestic facilities and projects for global cloud customers. That provides a starting point. AI-scale projects still require a step change in project size, equipment density and electrical integration.
The larger the announced pipeline becomes, the less useful headline gigawatts are as a measure of near-term supply. The relevant question is how many megawatts can be connected, equipped, commissioned and handed over each year.
The bottleneck can move again
Construction should not be treated as a permanent scarcity story. Several mechanisms could reduce the pressure.
Accelerator efficiency can improve the amount of useful computation delivered per watt. More inference workloads may distribute toward smaller regional facilities rather than remain concentrated in giant training campuses. Transformer, switchgear and cooling manufacturers are adding capacity. Reference designs can standardize engineering. Prefabrication can move work away from constrained sites. Contractors can train additional crews.
A slowdown in AI capital expenditure would relieve the constraint more directly. If data-center commitments grow more slowly than current plans imply, equipment queues and contractor backlogs could normalize before all announced campuses are built.
But standardization does not eliminate physical dependencies. It relocates them. A prefabricated power skid still requires transformers, breakers, copper and factory capacity. A liquid-cooled rack still requires heat rejection. On-site generation still requires fuel infrastructure and interconnection. Faster server deployment can make the slower parts of the system more visible rather than less important.
That is why construction is becoming part of the AI supply chain rather than merely the location where the supply chain is installed.
For the next phase of the buildout, announced capital expenditure and planned gigawatts will be less informative than four operational measures: time to power, major-equipment lead times, MEP and modular-construction capacity, and the interval between structural completion and successful commissioning.
The relevant unit for the construction industry is no longer square meters built. It is megawatts that can be energized, cooled, commissioned and handed over on schedule.
Sources and Methodology
This analysis combines current public information from the International Energy Agency, JLL, Schneider Electric, U.S. contractor disclosures, Korean government materials and public reporting on electrical-equipment supply chains. Korean securities research from Kiwoom Securities and Hanwha Investment & Securities was used as secondary background for construction scope, domestic contractor references and indicative MEP intensity.
Broker forecasts and construction-cost assumptions were not treated as official market forecasts. The discussion distinguishes announced or planned data-center capacity from operating capacity. Rack-density figures represent example architectures and industry ranges rather than a single universal design standard. Contractor backlog figures include businesses beyond data centers and are used only as evidence of demand for overlapping specialist construction capacity.