AI News

From Capital-Intensive to Labor-Intensive: The Global AI Infrastructure Bottleneck Is Stuck at the Electrician Shortage

HuxiuUpdated Nov 16 08:12

AI summary

The Huxiu article points out that the AI infrastructure bottleneck has shifted from GPUs and power equipment to electricians and construction crews; scheduling for large gas turbines has already been pushed to 2031, while skilled electricians require 4-5 years of training. The United States has about 818,700 electricians, with about 81,000 job openings each year, of which about 20,000 come from retirements; over the next decade, about 300,000 new electricians will be needed, or about 500,000 including replacements. Google.org, Meta, and BlackRock have invested a combined total of more than $265 million in training skilled workers, and electrical work accounts for 45%-70% of data center construction costs.

Why read this

AI score 54

The article moves the AI infrastructure bottleneck from equipment scheduling to the electrician training cycle, using a comparison between delivery lead times and apprenticeship durations to illustrate the part that capital cannot compress.

Topics

Tags

Body · AI translation

[[**The AI infrastructure bottleneck has shifted from GPUs and power equipment to electricians and construction crews; scheduling for large gas turbines now extends to 2031, while training a skilled electrician takes 4-5 years—the labor shortage is becoming a long-term constraint that capital cannot accelerate.**

**Key points**

**1. The equipment-side ceiling now stretches to 2031**

The three gas turbine giants' capacity is booked through 2031, with GE Vernova's slots reaching 116 GW in Q2 2026; U.S. gas projects that have been paused or delayed have more than quadrupled to 159 GW, transformer lead times run 100-150 weeks, and switchgear 2-3 years.

**2. The electrician gap is mainly retirement replacement**

The U.S. has about 818,700 electricians, with roughly 81,000 openings a year, of which 20,000 come from retirements; over the next decade, 300,000 new electricians will be needed, or about 500,000 including replacements. This gap is structural and will not disappear even if AI investment slows.

**3. The training cycle cannot be compressed by capital**

NECA commercial apprenticeship applications rose more than 70% from 2022 to 2024 to about 120,000, but expansion is limited by the number of master electricians; a skilled electrician takes 4-5 years, while data center roles take only 18-24 months. High pay exists, but stability depends on the project pipeline.

**4. Capital building its own training exposes the long-term constraint**

Google.org, Meta, and BlackRock have invested more than $265 million combined in training skilled workers; electrical work accounts for 45%-70% of data center costs, and Gartner expects 40% of existing AI data centers to be power-constrained by 2027. Workarounds still require engineers. 2026-11-02 09:04 In-depth | From capital-intensive to labor-intensive: the global AI bottleneck is stuck on the lack of "electricians" Sikai Has Something to Say Quick overview

Carroll County, Ohio. A gas-fired power plant called Chestnut Run.

The $2 billion investment was in place, the permits had been obtained, and the project had even squeezed into PJM's (the U.S. Midwest power market operator) "fast-track reliability process"—in the context of U.S. power projects, that was almost equivalent to clearing all hurdles.

Then it stopped in place.

The GE Vernova HA.03 model originally planned for use was delayed by at least two years because of excessive global demand. The developer, Advanced Power, had no choice but to apply to switch to the smaller HA.02—changing models meant re-approval, and FERC (the U.S. Federal Energy Regulatory Commission) ultimately did not grant the exemption, so the project's commercial operation date could be delayed by more than two years.

The land was secured, the capital was secured, and the grid connection queue was secured. In the end, it got stuck on the delivery schedule for a single machine.

And even if the machine arrives, there is a further hurdle behind it. In an interview with the energy research firm Rystad, NRG President and CEO Robert Gaudette put it bluntly: turbines are the constraint, switchgear is a problem, EPC (engineering, procurement, and construction) companies are fully booked, and electricians and welders are in acute shortage everywhere.

What is even more telling is that even Wall Street has begun to list "people" as the number-one risk in this chain. Jefferies analysts noted in a research report that the labor shortage is "undermining the integrity of the power supercycle itself." When the investment bank most bullish on power infrastructure begins to worry that the machines cannot be installed, it shows that it has realized: money is no longer the hard currency in this chain.

The bottleneck has changed position four times, and this time it has moved to "people"

From 2023 to 2024, the bottleneck was GPU supply. From 2024 to 2025, the bottleneck was power and grid connection queues. From 2025 to 2026, the bottleneck is electrical equipment such as transformers and switchgear, with lead times stretching to 2028-2032.

Starting in 2026, the bottleneck has shifted to electricians and EPC construction crews.

The first three times could all be solved in the same way: place orders, pay, and queue. In those three layers, capital was omnipotent.

The fourth time, it is not. Training a skilled electrician takes four to five years, while lead times for large transformers are about two to three years. People are slower than equipment.

The ceiling on the equipment side has already been pushed to 2031

Before discussing people, it is necessary to see clearly the ceiling pressing down from above.

The global large gas turbine market is controlled by three companies: GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries.

GE Vernova's gas turbine backlog and slot reservations stood at 83 GW at the end of 2025 and had reached about 100 GW by Q1 2026—an increase of about 17 GW in a single quarter—and further rose to 116 GW in Q2 2026, with the company forecasting at least 125 GW by the end of 2026. Production scheduling has extended to 2031—what the company is accepting now are delivery slots for 2031. Global annual gas turbine orders have reached the 100 GW scale, about three times the pre-pandemic average. According to Global Energy Monitor data, U.S. utility gas projects that have been paused or delayed have grown more than fourfold (+420%) since early 2025, to more than 159 GW.

In its April 2026 report "U.S. Gas Turbine Market: Navigating Manufacturing Shortages and Demand Growth," Wood Mackenzie was quite forceful in its characterization: "The supply chain crisis in gas turbine manufacturing is one of the sharpest bottlenecks in meeting data center demand."

And "scheduling slots" themselves have already become an asset that can be hoarded. GE Vernova uses a slot reservation mechanism. The three-party cooperation among NRG, GE Vernova, and engineering firm Kiewit is a typical case: two 7HA units are reserved, with the first 1.2 GW to come online in 2029, then filled in batches through 2032, for a total of 5.4 GW—through this cooperation, NRG has locked in both turbine and EPC capacity through 2032.

Scheduling slots themselves are a scarce asset. The market has seen this logic before with GPUs. The difference is that GPU production can be expanded, while the capacity expansion cycle for large gas turbines is on the order of five years—an expansion announced only in 2026 does nothing for the immediate shortfall.

Equipment lead times are likewise out of control. Heavy-duty gas turbines are scheduled into 2031 to 2032, gas engines to late 2028 to 2029, generator step-up transformers to 144 weeks, large power transformers to more than 100 to 150 weeks, averaging 128 weeks, with special orders taking up to 4 years, medium-voltage switchgear close to two to three years, and cooling systems 30 to 70 weeks. A transformer ordered in 2026 will not arrive until late 2028 at the earliest. Prices are also out of control: Wood Mackenzie estimates that the average price of gas turbines has risen 195% cumulatively since 2019, and tariffs and country-of-origin restrictions are making the situation worse.

The people gap is essentially a "replacement problem"

The BLS (U.S. Bureau of Labor Statistics) baseline is: 818,700 electricians nationwide, with employment growth of 9% from 2024 to 2034, far above the average for all occupations, and about 81,000 job openings each year. The key point is that the vast majority of openings are "replacing those who leave," not net new additions.

The retirement wave is the main cause: about 20,000 electricians retire each year, roughly 200,000 over the next decade. For the ten-year total gap, the industry estimates that more than 300,000 new electricians are needed; including retirement replacement, the total is about 500,000. Several studies published by McKinsey between 2023 and 2026 offer a more focused framing: for AI infrastructure alone, between 2023 and 2030 the United States needs to train an additional 130,000 electricians, 240,000 construction workers, and 150,000 site supervisors.

Why is this distinction crucial? It determines the policy solution. If the gap comes from a "surge in demand," then once AI investment slows, the gap will naturally disappear—this is a cyclical problem. But if the gap comes from "a retirement wave plus a clogged training pipeline," then even if the AI bubble bursts, electricians will still be in short supply—this is a structural replacement problem.

The former can be solved with capital; the latter cannot.

There is one indicator that proves demand in reverse: NECA (National Electrical Contractors Association) data shows that commercial apprenticeship applications grew by more than 70% between 2022 and 2024, rising from about 70,000 to about 120,000. There is no shortage of people applying; what is lacking is training capacity. The speed at which union apprenticeship programs can expand their slots cannot keep up with the influx of applicants. And the expansion of apprenticeship programs is in turn limited by the number of master electricians—a master can only take on a limited number of apprentices, and the masters themselves are retiring. It is the combination of these two constraints that is the real reason "applications grew 70% yet there is still a shortage."

A cross-sectional comparison: electricians exceed HVAC (8%/40,100) and plumbers (4%/44,000) on both growth rate (9%) and annual openings (81,000).

Pay: the same trade, two prices

The most honest evidence of shortage is written on the paycheck.

A first-year apprentice (with zero student loans) has total annual compensation of about $42,000. The median U.S. electrician (2024, all industries) earns about $62,350. Data center construction workers average about $81,800, 32% higher than on a standard job site. A union journeyman electrician (IBEW Local 26) earns $59.50 an hour, or more than $120,000 a year. A foreman or site supervisor earns $140,000 to $180,000. A senior commissioning engineer earns $200,000 to $260,000. A senior data center electrician (Texas, total compensation) earns $240,000 to $280,000.

Several points of clarification are needed.

First, "$240,000 to $280,000" is not base annual salary. It is the total of base wages plus overtime plus per diem plus project bonuses. On base rates, the common market rate for a licensed electrician in Northern Virginia is $40 an hour plus per diem, and base pay has already risen above $120,000.

Second, the location of IBEW Local 26 is itself a footnote. This union is based in Northern Virginia—the "global data center capital," home to about 13% of global data center capacity. Apprentices start at about $26 an hour; after completing a five-year apprenticeship, journeymen earn about $59.50 an hour, and with overtime or foreman roles, close to $200,000 a year.

Third, high pay does not equal stability. Data center construction jobs usually last only 18 to 24 months per campus. The truly valuable career path appears in states with a continuous project pipeline. This is why "high pay" and "career" must be discussed separately.

Fourth, the overlooked segmentation of trades. Within the IBEW system, there are at least inside wiremen, overhead linemen, and telecommunications VDV installation technicians. Data centers mainly need inside wiremen and commissioning specialists, who are not the same group of people as transmission line workers. This means that grid expansion and data hall construction are not competing for the same group of people; rather, two pressures are stacking at the same time—each is short of people, and they cannot substitute for each other.

Why "people" are the hardest link

Put the equipment cycle and the human cycle side by side, and the conclusion emerges naturally.

Training a skilled electrician takes four to five years, and the IBEW apprenticeship is typically five years. Large power transformers take 100 to 150 weeks or more, roughly two to three years. Medium-voltage switchgear is close to two to three years. Heavy-duty gas turbines are scheduled for 2031 to 2032. A single data center construction job lasts about 18 to 24 months per campus.

Transformers can be queued up with money; electricians cannot. Money can buy capacity expansion—GE Vernova is expanding production in 2026—but money cannot buy time, because an electrician's growth curve cannot be compressed.

There is also a finer layer of intensity change: the power consumption of a single GPU rack has risen from 10 to 14 kilowatts a decade ago to 120 to 140 kilowatts—about tenfold. And electrical systems account for 45% to 70% of the total construction cost of a data center. An AI data center is essentially a giant power distribution cabinet, not a building. The building shell accounts for only 10 to 15% of the total budget, while electrical accounts for nearly half to 70%.

The density of money has not decreased, but once the money is spent, it must be delivered by human hands.

Multiple surveys show that Gen Z's willingness to enter skilled trades is rising significantly. Thumbtack's skilled trades report released in October 2026 shows that nearly half of respondents aged 16 to 30 are learning or seriously considering learning a skilled trade, and a considerable proportion explicitly said that the rise of AI has made them more interested in skilled trades. An earlier Jobber survey also found that high school- and college-age respondents generally believe blue-collar jobs offer more career security than white-collar jobs—precisely because of the spread of AI.

Changes on the cultural side are equally obvious: stigma has decreased, TikTok exposure has increased, and discussions about student loans and wages have become more open; at the same time, many entry-level white-collar positions "feel more like a trap than a ladder."

A 22-year-old who originally would never have considered becoming an electrician now attends classes two evenings a week at the local IBEW chapter, earns about $42,000 in the first year, and expects about $71,000 after graduating as a skilled worker—in his own words, what he does every day is "adult Lego."

Tech giants begin training workers themselves

How severe is the shortage? Customers are starting to build their own supply.

Google, through Google.org, invested $50 million and partnered with the Electrical Training Alliance to raise annual electrical apprenticeship enrollment from 19,500 to 30,000 within three years. Meta invested $115 million to launch America's Workforce Academy, training about 5,000 people in the first year. BlackRock invested $100 million to launch a skilled trades training program. Together, the three total more than $265 million.

But there is a key irony here: BlackRock is not just "doing charity." It is an 80% owner of Meta's $12 billion data center in El Paso, deeply tied to the returns on AI infrastructure assets.

It invests in training and protects its own assets.

When financiers start paying out of their own pockets to train construction crews, it shows that the market has concluded: this is not a cyclical problem, but a long-term constraint on the supply side. If capital only needed to pay interest, it would not set up vocational schools. The fact that it is setting up vocational schools shows that it has discovered that money no longer works at a certain link.

Attempts to bypass the bottleneck ultimately cannot bypass people

Facing the dual chokepoints of equipment and people, the industry has proposed several alternatives: co-locating desert solar power with data centers, using AI to regulate electricity use in reverse, building self-owned off-grid power sources, and innovations in transmission and microgrids. These solutions all try to skip the link of "grid connection + electrician construction," but they all require another group of engineers. The problem has been shifted, not solved.

The demand-side numbers themselves expose the dilemma of planning: forecasts of U.S. data center energy consumption by 2030 range from about 200 TWh to more than 1,000 TWh—the width of this range is itself news: even planners do not know the real scale. A forecast released by Gartner in November 2024 pointed out that by 2027, 40% of existing AI data centers will face operational constraints due to shortages in power availability—the risk has escalated from "construction delays" to "operational cutbacks."

The uncertainty on the demand side, combined with the dual oligopoly on the supply side, is the real structure behind the "electrician shortage."

A bill that cannot be repaid with debt

GPUs can be expanded in production. Transformers can be queued. Gas turbines can have schedules reserved, and can even be stockpiled on the basis of "void if unused."

Only electricians require four years and a young person willing to attend classes two evenings a week.

Now this premise has been broken.

The bottleneck has moved from the capital-intensive end to the labor-intensive end. And people are the only asset in the world that cannot be accelerated with a balance sheet.

Capital can stack computing power, buy up equipment schedules, and overdraft future credit, but it can never buy an electrician's five-year growth cycle. The ultimate involution of the AI industry is no longer in technology or funding, but in the supply speed of industrialized labor. AI industry signal channel: Frontier Technology

Has published 87 articles on Huxiu. This content is authorized by the author for publication; the views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is in the Huxiu APP[ ]]]

Source

Huxiu