AI's $335 Billion Buildout Hits the Apprenticeship Ceiling

OpenAI is hiring electricians on 10-hour days, seven days a week at its Saline Township, Michigan site, calling the buildout the biggest single investment in Michigan's history, according to the New York Times. One of the workers on the project was wiring houses in the same county a year ago. This is the AI data-center buildout, the physical layer beneath the models: the wire, conduit, switchgear, and fiber that connect GPU clusters to the grid. The IBEW local handling the project has an apprenticeship queue that is, by the labor-market signals in the primary reporting and a Hacker News discussion of the IBEW apprenticeship bottleneck, more than 200 applicants deep for seats that graduate a small number of journeyworkers a year.

The industry's numbers around the same buildout do not line up. Alphabet and Meta are projected to spend more than $335 billion on AI data centers in 2026. The training commitments from the same companies and their peers, including OpenAI and BlackRock, total about $265 million, according to figures in the NYT and carried in Quartz and Yahoo Finance. The gap between the two numbers is more than three orders of magnitude, and it reframes the training-funding press cycle as a down payment rather than a fix.

The individual commitments, read on their own, look substantial. Meta's $115 million America's Workforce Academy is training recruits in Louisiana for the data-center trades, and the program is one of the largest single-company trades-training pledges on record. Google has committed $10 million to the Electrical Training Alliance, a joint program of the IBEW and the National Electrical Contractors Association that runs the country's main electrical apprenticeship pipeline. A Business Model Analyst read of the same numbers puts the ratio at roughly 0.08 percent of combined capex.

The bottleneck the money cannot move is institutional. The International Brotherhood of Electrical Workers, the largest U.S. electricians' union, runs apprenticeships that last four to five years and graduate a fixed number of journeyworkers per class. Seats are allocated by local unions and NECA chapters, not by corporate checks, and the queues reflect that. In Saline, the queue is roughly 200 deep even for applicants with prior high-voltage experience, per a Hacker News thread that mirrors the labor-market signal in the primary reporting. The capex is elastic; the seats are not.

The scale of the gap sits on top of a labor market that was already short. The Associated Builders and Contractors estimates the U.S. construction industry needs about 349,000 net new workers in 2026 to meet demand. The Bureau of Labor Statistics projects a shortfall of roughly 81,000 electricians per year through 2034, a number carried in the AI Weekly digest of the NYT reporting. AI demand is concentrated in a small number of metros, but it competes for the same apprentices the rest of construction is trying to hire.

The same cyclicality that has always shaped construction now follows the data-center boom. A 2026 capex revision, a power-allocation delay, or a slowdown in model training would land on the same apprenticeship seats the training programs are filling. The IBEW pipeline cannot compress a four-year program into a 12-month spike, and a residential electrician pulled into a data-center build today is a residential electrician not on a hospital or school job tomorrow. The training-funding press releases are real commitments, and the labor shortage is real. The next test is whether the seat allocation can absorb either at the pace the capex is moving.

For a tradesperson weighing the call, the demand is concentrated and the pay is, by the NYT's account, the highest the trades have seen. For a reader weighing the AI buildout, the next test is whether the training commitments scale, or whether the capex hits the apprenticeship ceiling first.