The electrician shortage slowing the AI boom

Kann sets this up as a question about rate limiters on AI infrastructure. The candidates are demand, chips, power and labor, and labor gets named constantly and then skipped, probably because it is the least tractable: money can be thrown fairly directly at chips and at power, but not at people. Asked where labor ranks for his own buildout, Hawkins puts power first, says community opposition has climbed the list startlingly fast, and places labor after both. Then he makes the distinction the rest of the episode rests on. Labor is not stopping anything. It costs much more, it strains the schedule, and it forces constant management of which people are slotted where and when. If he had all the labor in the world, he would go faster. That is a throttle on pace and cost rather than a ceiling on what gets built, and it is a weaker claim than the episode’s title.

Why electricians in particular is really two questions, and Hawkins answers both with “and” rather than “or.” A data center is unusually electrical: the product is sold by the kilowatt because power is the limiting factor in the building, so the electrical work is enormous in volume and far above residential grade, while the heavy equipment operators, plumbers and drywall crews look much like any other large construction project. He needs hundreds of electricians and he needs them senior. Asked whether the fix is therefore training more electricians or training more specialized ones, he says the industry could absorb both today, and compares it to the utilities: the same sudden pace that overwhelmed the people who build generation and wires has overwhelmed the people who wire the other end.

The market response is visible and slow. Kann presses the efficient-market logic, that the chance to earn triple wages should pull people into the trade even without special programs, and Hawkins agrees it is happening in essentially every market where Compass works. The more interesting part is why a developer would fund training itself, which is not the historical pattern, and the reason is duration. A lease signed today means a campus under construction for four or five years, so someone starting a two-year apprenticeship now is still useful on that same site for another three or four. That arithmetic makes a training pipeline rational for a party that does not even employ the electricians, since the general contractor or electrical contractor is the actual employer and Compass is only the party that needs them to exist. Hence a twelve-week program with a community college near their Dallas campus, built alongside equipment partners whose gear those electricians will install. In the meantime the gap is filled by moving people: recruiting from other regions, housing them in camps near sites in smaller markets, and rotating crews through in blocks of weeks, a pattern he compares to both military deployments and travel nurses.

The other half is reducing how much labor a building needs, and here the honest answer complicates the good news. Compass has treated data centers as an assembly process rather than stick-built construction from the start, but Hawkins is careful that reduced site labor was a consequence rather than the motive. The original reasons were speed, quality, safety and a more diverse workforce, since the traveling workforce is overwhelmingly men and factory work that ends at five is open to people who cannot relocate. Roughly 85% of a building now arrives on trucks, including electrical distribution as preassembled units. But the skilled labor does not disappear; it relocates to partners’ manufacturing facilities, so the shortage moves rather than lifts. And Kann’s closing observation is the one to keep: even with that degree of prefabrication, a single building still takes about nine months, a campus still takes five years, and the projects are still short of people. The last third of the conversation turns to community license, where Hawkins argues the industry’s secrecy was inherited from security-conscious early customers rather than chosen, and that local training is one of the few moves that serves the build and the community relationship at once.