The average cost of a new data center surged to $475 million in the trailing twelve months through April 2026, up from $177.9 million during the same period a year earlier: a 167 percent increase that reflects something more fundamental than inflation. (Source: ConstructConnect, June 2026.) The facilities being built now require a different category of infrastructure: higher power density, liquid cooling systems, and fiber cabling at a scale that didn’t exist in mainstream data center design two years ago.
What that means in practice, on the floor, during installation, is the subject most industry coverage skips. This is a breakdown of the four constraint areas that define AI infrastructure installation in 2026, with the data behind each one.
Power distribution: the first thing to solve
Traditional data center racks run at 10 to 20 kilowatts. AI server configurations are pushing 100 to 240 kilowatts per rack, with some next-generation systems spec’d beyond 370 kilowatts. That’s not a higher-density version of the same infrastructure. It requires a different distribution architecture from the transformer to the rack.
At 240 kW on a 208V three-phase circuit, the current draw approaches 660 amperes. The conductor sizing, conduit routing, termination hardware, and circuit protection that serve that load differ fundamentally from what standard PDU infrastructure requires. Overhead busway distribution, installed before racks arrive and not after, is becoming the standard approach for AI deployments at this density.
The installation implication is sequencing: the electrical infrastructure has to be in place and verified before the compute hardware is positioned. Projects that treat power infrastructure as a parallel workstream, installed alongside the racks, discover the conflict when the schedule has no room to absorb it.
Cooling: from air to liquid, in a facility that’s still running
According to the AFCOM State of the Data Center Report 2026, 36% of data centers have already deployed liquid cooling, and another 28% plan to adopt it within the next 12 to 24 months. Most of those deployments aren’t happening in new facilities; they’re retrofits of infrastructure that was designed for air cooling.
The installation sequence for a liquid cooling retrofit is fundamentally different from a greenfield deployment. In a greenfield facility, cooling distribution units are specified and positioned from day one, coolant loop routing is part of the structural design, and servers arrive from the factory with cold plates pre-installed. In a retrofit, none of those conditions apply.
In practice, a liquid cooling retrofit proceeds in stages: the cooling distribution unit gets installed and connected to water supply and drain infrastructure before servers are touched. The coolant loop is pressure tested before any connections are made to compute hardware. Servers are connected rack by rack with flow verification at each step. Air cooling stays operational throughout; both systems run in parallel until DLC has proven stable under operating load, then air cooling is decommissioned incrementally.
The period when both systems run simultaneously is where the risk concentrates. Leak detection, flow monitoring, and temperature verification require active attention during this phase, not just at commissioning.
Structured cabling: the most underestimated constraint
AI racks require up to eight times more fiber connectivity than traditional server infrastructure. That figure comes from the combination of port count, bandwidth per port, and the scale of GPU interconnect fabric that AI clusters depend on.
It’s worth distinguishing two separate fabric layers. The scale-up interconnect, NVLink in NVIDIA’s architecture, handles GPU-to-GPU communication within a rack or rack-scale unit. A single NVL72 rack unit contains 72 GPUs communicating across a full mesh at high speed; the fiber density serving that unit is a direct function of that internal fabric. The scale-out interconnect, 400G or 800G InfiniBand or Ethernet, handles communication between racks and across the cluster. Both layers contribute to the total fiber count, and both have to be planned and installed correctly. A traditional server rack running 10G connectivity might have a few dozen active fiber connections. A GPU rack running both NVLink and 400G/800G scale-out fabric can require hundreds of connections per rack, with no tolerance for latency variance across either.
This changes the physical infrastructure in three ways. High-count MPO/MTP trunk cables replace individual patch cord runs, reducing pathway congestion by approximately 75 percent while delivering the fiber density AI racks need. Overhead cable tray distribution replaces under-floor routing, which can’t accommodate the volume or bend radius requirements of high-density fiber trunks. And rack-level fiber cassettes and breakout hardware have to be installed and tested before compute arrives, not during installation.
Fiber type matters as well. OM4 and OM5 multimode handle shorter GPU-to-GPU runs within and between adjacent racks. OS2 single-mode covers longer inter-rack distances across the cluster. A cabling design that uses the wrong fiber type doesn’t fail obviously; it degrades performance in ways that are difficult to trace back to the root cause.
Commissioning: where the schedule risk lives
Commissioning AI infrastructure accounts for 10 to 20 percent of the total project schedule. The reason isn’t documentation or process overhead; it’s that power distribution, cooling systems, structured cabling, and fire suppression systems that were installed as separate workstreams have to be tested together, under simulated full load, for the first time.
That’s the moment when installation errors that weren’t visible during individual system testing become visible. A PDU that held at 40 percent load but trips at 85. A liquid cooling loop that passed pressure testing empty but weeps at operating temperature. A fiber run that cleared visual inspection but fails OTDR testing under full-length loss measurement.
None of these surface during installation. All of them surface during commissioning, at the point when the schedule has already consumed its buffer.
The planning implication is straightforward: commissioning needs to be resourced and scheduled as a primary phase, not treated as the activity that gets whatever time remains after the build. Projects that plan for commissioning the way they plan for installation have contingency where the actual risk is.
Note: The 10 to 20 percent commissioning schedule estimate reflects field experience across AI infrastructure deployments; this range is consistent with Uptime Institute commissioning guidance and industry practitioner reporting, though published benchmarks vary by project type and scope.
Equipment lead times and what they mean for scheduling
The constraint that sits upstream of everything above is procurement. High-capacity transformers and PDUs are currently running 12 to 14 months from order to delivery. Medium-voltage switchgear can reach 18 months in some markets. Cooling distribution units for DLC systems run 6 to 10 months depending on capacity. (Source: industry practitioner reporting; lead times vary by manufacturer and market conditions as of mid-2026.)
These lead times exist because data center construction spending reached $49.5 billion through April 2026, up from $13.6 billion in the same period a year earlier. (Source: ConstructConnect, June 2026.) Every project in that pipeline is drawing from the same supply pool.
For a project with a Q1 2027 commissioning target, long-lead equipment orders need to be placed by Q3 2026 at the latest, which means the spec has to be defined by Q2 2026. Running procurement and engineering in parallel, issuing preliminary purchase orders based on preliminary specs and then finalizing as design progresses, is the practical workaround. It requires a different kind of coordination between procurement and engineering than most project structures assume.
What this adds up to
AI infrastructure installation in 2026 involves four interacting constraint areas: power distribution architecture being redesigned for load profiles that didn’t exist two years ago; liquid cooling systems retrofitted into facilities built for air; structured cabling at significantly higher fiber density than traditional infrastructure; and commissioning that tests all of it together under conditions no individual workstream has validated.
The projects that navigate this successfully treat installation as a sequencing problem, not just a logistics problem. The order in which systems go in, and the dependencies between them, matters as much as the quality of each individual installation.
We Install AI deploys AI and HPC infrastructure across the US: rack & stack, structured cabling, power distribution, liquid cooling, and commissioning.
The hardest part of an AI infrastructure deployment isn’t installing the equipment; it’s sequencing power, cooling, cabling, and commissioning before schedule risk builds. If you’re planning a build or retrofit, reach out early.