Meeting the deadlines on the data center projects already proposed would take roughly 500,000 more electricians, 300,000 more welders, and 550,000 more plumbers in the United States, according to NBC News. The hardest roles to fill are electricians, HVAC technicians, and mechanical pipefitters. Contractors are turning away work or poaching crews from smaller firms to keep sites staffed. Reported wages for AI data center electricians under 30 now run in the $240,000 to $280,000 range, with year-over-year wage growth in the mid single digits for journeypersons and near 10 percent for commercial electricians. Those wage figures come from industry reporting rather than a federal series, so treat the exact bands as directional.
This is the part that gets read as a general construction story and is not one. A direct-to-chip deployment is pipefitting work end to end: secondary loop headers, rack manifolds, CDU tie-ins, hydrostatic testing, flushing to a particulate spec, glycol charging, and commissioning the flow balance across a row. It is the same craft and the same certifications as the chilled water plant it hangs off. HVAC technicians own the plant. Electricians own everything feeding it. A shortage concentrated in those three trades is a shortage aimed directly at the mechanical scope, which is the scope that determines whether a hall accepts GPUs on schedule. The workforce gap has been the most cited barrier to liquid cooling adoption for two years running.
Crews get pulled to the electrical critical path first because switchgear has the longest lead time and the most visible milestone. Mechanical trim, balancing, and loop commissioning are the tail of the job, so that is where a thin crew shows up as slip. Commissioning a 100MW hall is where cooling risk concentrates, and a flow balance done fast by a crew that has not done many is the failure mode that surfaces two weeks into production load. Flow balancing across parallel racks is unforgiving: add racks and per-rack flow drops on the ones already installed.
The vendors who solve this over the next twenty-four months beat the vendors shipping a better cold plate. EdgeConneX built a global training center for precisely this reason. Quick-disconnect designs that tolerate a less experienced installer, factory-terminated manifold assemblies, pre-charged skids, and commissioning software that catches a starved rack before the GPUs do are all worth more right now than another two percent of thermal resistance. Ask any vendor pitching you a liquid cooling platform how many labor hours their system takes per rack against a competitor's, and how many of those hours need a journeyperson card. Most of them have never been asked. The ones with an answer are the ones to shortlist.