The GPU supply chain has the industry's attention. But the constraint that increasingly decides when an AI factory energizes is no longer the chip. It is power delivery, and the licensed electricians who commission it.
On June 1, executives from Oracle, OpenAI, Related Digital and Blackstone gathered with Michigan Governor Gretchen Whitmer in Saline Township, southwest of Ann Arbor, to mark construction of "The Barn," a Stargate campus reported at $16 billion and billed as the largest economic investment in the state's history. The number that mattered most to the schedule wasn't the headline capital figure, and it wasn't the roughly gigawatt-scale power capacity the three buildings will eventually draw. It was 300,000: the union trade hours that more than 700 union tradespeople had already logged on the 250-acre site, working under the first data center labor agreement signed between OpenAI and North America's Building Trades Unions.
That detail is the part of the AI buildout that the chip headlines keep walking past. For two years the industry has read its own constraints off the silicon supply chain: GPU allocation one quarter, HBM yield the next. SCN has argued before that the real bottleneck is rarely where the cameras are pointed. The binding constraint on bringing AI factories and large-scale supercomputing capacity online on schedule has moved again. The wafer can be contracted and the megawatt can be contracted, and the cluster still does not run until the people and the long-lead iron that energize the room are in place.
This is the Power Question, the standing mandate this publication has committed to tracking: can the infrastructure for the next generation of supercomputing actually get built fast enough, and who pays for it? Most coverage of that question stops at generation and grid. The harder edge of it now sits one layer down, in commissioning: the infrastructure work of turning a financed building into an accepted, powered machine. A GPU can be allocated and a substation contracted, but neither produces a running cluster until someone qualified terminates the medium-voltage gear, energizes the room, and signs off the commissioning sequence.
Start with where the money goes. Data center construction costs have climbed from $7.7 million per megawatt in 2020 to $10.7 million in 2025, roughly a 7 percent compound annual rate, with another 6 percent rise forecast for 2026; that same industry analysis singles out electricians, mechanical contractors, controls and commissioning specialists as the labor in shortest supply. Within that cost base, electrical systems are not one line item among many. Estimates circulating through trade and union analysis put electrical at 45 to 70 percent of total build cost, an IBEW-associated figure worth treating as directional rather than audited. Even at the low end, it reframes the problem: the trade that dominates the budget is the trade in shortest supply.
The supply gap is structural, not a hiring-season blip. One widely cited industry estimate, which is not a government figure, puts roughly 340,000 US data center positions unfilled across construction and operations through 2026. Firmer ground sits in the federal data: the Bureau of Labor Statistics projects about 81,000 electrician openings a year through 2034, driven as much by who is leaving as by who is needed. In the union electrical segment, NECA data shows about 7,000 entrants against 10,000-plus retirements a year, which Riverside's analysis projects as a 110,000-electrician shortfall by 2034 before any aggressive AI scenario is layered on. Stack the AI demand on top and the numbers turn vertical. Reporting in Fortune and elsewhere puts the decade's AI- and data-center-specific need in the hundreds of thousands of new electricians; Microsoft's own estimate runs to roughly 500,000 over ten years.
The people running these builds have started saying so plainly. Microsoft's Brad Smith has called the national electrician shortage the single biggest challenge for data center expansion in the United States. The clause is reproduced from Fox Business commentary, and the company has reportedly backed it by flying crews in from beyond 75 miles to keep projects moving. Goldman Sachs Research reached the same point in capital terms, estimating the power industry may need more than 750,000 additional workers by 2030, including roughly 207,000 in transmission and interconnection. As Fortune read the May 2026 note, the human side of this buildout may prove harder to solve than the hardware side, with a projected 45 GW US data-center power shortfall by 2028. It is a gap no amount of model progress closes on its own; software cannot terminate a feeder.
Even fully staffed, the equipment these crews install runs on its own calendar. Power transformer lead times have stretched from about 50 weeks in 2021 to roughly 120 weeks, well over two years, on average by 2024, with large units ranging as wide as 80 to 210 weeks, and stretching to four or five years for some high-voltage classes in Western markets; manufacturers now name generative-AI data centers explicitly as a demand driver. Grid interconnection runs longer still: in PJM, projects reaching service in 2025 averaged more than seven years from request to operation, and more than 2,000 GW of capacity sat in US interconnection queues at the end of 2024. The demand behind those queues is not hypothetical. The IEA's Energy and AI report projects global data center electricity demand roughly doubling to about 945 TWh by 2030 and reaching about 1,200 TWh by 2035, up from roughly 415 TWh in 2024.
That collision of demand and queue length is what SCN has tracked as the 100-gigawatt supercycle, and it is pushing operators toward bringing their own power. xAI's Memphis build is the clearest case study, a bet that on-site generation could outrun both the grid and the regulators. The appeal is timeline: on-site generation can, in principle, shave years off an interconnection wait that already runs well past half a decade in the busiest markets. But it routes around the grid, not around the trades. Someone still has to build the generation plant, pull the cable and commission the switchgear, and gas turbines carry multi-year backlogs of their own.
The trades pool is finite, and AI factories are not the only claimants on it. National labs and research-computing centers building exascale systems draw electricians, controls technicians and commissioning specialists from the same regional halls a hyperscaler campus can. Data-center electrical work commands 25 to 30 percent wage premiums over standard commercial rates, a market-clearing signal that pulls labor toward the highest bidder. There is not yet clear public evidence that research builds are being delayed as a result, but the wage mechanics only point one way. That is the Access and Science question in physical form, and it is part of why Oak Ridge stood up a Next-Generation Data Center Institute aimed at the engineering-and-talent gap the commercial surge has opened. The buildout's labor crunch is a sovereignty fact, too, in the most literal sense: domestic compute capacity is bounded by domestic wiring capacity, and you cannot reshore an AI factory you cannot staff to energize.
The constraint is also uneven, and worth naming as such. It bites hardest in dense metros like Northern Virginia and during the electrical-rough-in and commissioning phases; it is looser in regions with slack trade capacity, or on projects still early in site and shell work. None of this means it is permanent, either. Prefabrication is the industry's main lever, and modular electrical skids and factory-integrated power rooms genuinely cut on-site hours, though they shift that labor to the factory rather than erasing it, and integrated-skid testing can raise demand for senior, license-carrying electricians. POWER magazine's reporting adds a sharper caveat: part of the transformer crunch is self-inflicted, the product of non-standard specifications and late ordering, and standardized designs can still land in 12-to-14-month windows. The point holds regardless of how much prefab absorbs. Saline Township's first phase is targeted to come online as early as 2027. Whether it hits that date will turn less on Oracle's GPU allocation than on whether 700 tradespeople become the several thousand the full campus needs, on a labor market that was already short before the AI surge arrived.
