The GPU shortage is over. The electrician shortage just started.

What happened

The New York Times reported on July 29 that the largest AI infrastructure operators — Microsoft, Meta, Google, Amazon, and a growing tier of neocloud builders like CoreWeave and Crusoe — are running recruiting programs aimed at electricians, pipefitters, carpenters, welders, and HVAC technicians at a scale the industry hasn't seen since the postwar interstate build.

The numbers in the piece are the story. One Microsoft-linked campus in Wisconsin is staffed by roughly 2,400 tradespeople at peak. A single Meta site in Louisiana is projecting 1,500 electricians for 30 months. Amazon's build pipeline across Virginia, Ohio, and Indiana is being sized against a national pool that, per the Bureau of Labor Statistics numbers cited in the piece, is short about 80,000 electricians against an annual graduation rate near 7,000. Union locals in Phoenix, Columbus, and Northern Virginia are reporting journeyman wages 30-50% above 2023 levels, with signing bonuses for anyone who can pass a medium-voltage switchgear cert.

The bottleneck on the next generation of AI capacity is no longer H100s, HBM, or even grid interconnect queue position — it's the human beings who terminate cable and torque busbar. That is a genuinely new sentence to be able to write, and it reorders a lot of assumptions.

Why it matters

For two years the shared mental model of AI infrastructure has been: chips are scarce, power is scarce, everything downstream is a solved problem. The scarce-chip story broke first — Blackwell allocations loosened through Q2, and the neoclouds started publicly complaining about GPU depreciation curves. The scarce-power story is real but at least legible: interconnect queues, PPAs, SMR pilots, gas peakers. It's an engineering and permitting problem.

The labor story is different in kind. You cannot spin up an electrician in six months the way you can spin up a substation in eighteen. An IBEW inside-wireman apprenticeship is five years. A medium-voltage switchgear specialist is closer to seven or eight years of on-the-job before they're the person you want commissioning a 400 MW hall. Meta and Microsoft can pay whatever they want; they cannot compress that timeline. What they *can* do — and what the NYT piece documents — is fund apprenticeship programs, poach from commercial construction, and quietly cannibalize the trades workforce that was supposed to build everything else: chip fabs, EV plants, transmission upgrades, housing.

The community reaction on Hacker News (247 points, ~600 comments at time of writing) skewed toward two positions worth taking seriously. The first, from people actually in the trades: this is the best labor market of their careers, and it will end badly the moment the AI capex cycle inflects, because there is no fallback demand at these wage levels. The second, from infrastructure engineers: the real risk isn't the headline shortage, it's the quality distribution. When you're hiring 1,500 electricians for a 30-month project, the marginal hire is not the person you'd have chosen in 2019. Commissioning defects — a mis-torqued lug, a mis-phased bus, a missed ground bond — become the actual production-availability story for these sites in 2027-2028.

If Meta's Hyperion campus slips two quarters because there aren't enough qualified switchgear commissioners, that shows up in your Bedrock or Azure OpenAI capacity, not in a NYT headline. The abstraction leaks all the way up to your rate limits.

What this means for your stack

Three concrete implications for anyone whose 2027 planning assumes AI infrastructure keeps compounding on trend.

First, treat announced capacity dates as fiction until proven otherwise. The hyperscalers have historically hit their MW-online commitments within a quarter or two. That pattern is breaking. The signal to watch isn't the press release; it's the local IBEW hall reports and the county-level building permit data for the specific campus you care about. If you're negotiating a multi-year committed-use contract on the assumption that a new region comes online in H2 2027, get contractual remedies for slippage or don't sign.

Second, your inference latency map is about to get weirder, not smoother. The mental model of "more regions, closer to users, lower p50" assumed roughly uniform buildout. What's actually happening is that campuses near strong union locals (Phoenix, Columbus, Dallas) are on schedule and campuses in labor-thin geographies (rural Louisiana, west Texas, parts of the Midwest) are slipping. Expect a bimodal capacity map through 2028 and plan failover and region-affinity accordingly.

Third, the cost curve on inference has a floor that isn't silicon. Everyone's discounted-cash-flow model for foundation-model economics assumes tokens-per-dollar keeps improving on the Blackwell→Rubin→post-Rubin march. It will, at the chip level. But if the fully-loaded cost of a MW of AI capacity is rising 15-25% year over year because tradespeople now cost what they cost, the token price curve flattens sooner than the chip roadmap suggests. That has implications for anyone building a product whose unit economics depend on inference costs halving on a predictable cadence.

Looking ahead

The interesting second-order question is what happens on the other side of the capex peak. The AI industry is currently subsidizing the American electrical trades at a scale the federal government hasn't managed in fifty years — and it's doing so entirely by accident, as a side effect of racing Google to the next model release. When the buildout cycle inflects, tens of thousands of newly-trained electricians will be looking for the next megaproject. The optimistic read is that they pivot to grid modernization, transmission, and the electrification backlog that everyone agrees the country needs and no one has been able to staff. The pessimistic read is a 2027-2028 bust in the trades that makes the current shortage look like a rounding error. Either way, the story of AI infrastructure in 2026 turned out to be a labor story, and the people who noticed first were the ones who could actually read a one-line diagram.