A 60MW data-centre project that slips costs its owner $14.2M a month (iRecruit). A buyer with that clock running and no journeyman electricians available is the best customer a labour broker can have, and it is a customer that currently buys from conventional construction staffing firms at conventional markups.
The shortage is quantified and severe: a 439,000-worker construction shortfall in 2025, with 349,000–499,000 more needed in 2026 alone, 41% of the workforce retiring by 2031, and 68% turnover — 73% in specialised trades (iRecruit; Birm Group). Projects that historically peaked at around 750 workers now need 4,000 to 5,000.
Here is why this page sits next to Compute and capacity brokerage rather than inside it. The GPU-hour those buildings will house is a fungible unit with a public index and, from October 2026, listed futures — which is why brokering it earns almost nothing. A licensed electrician in Raleigh-Durham with data-centre experience and an available start date is none of those things. There is no index, and that is where the spread lives.
Whose budget
Deep, urgent and capital-project shaped. budget: 5.
The pay tells you how tight it is. Journeyman electricians on data-centre work earn $120k–$200k, with state averages of $207k in North Carolina, $184k in California and $183k in Texas — a roughly 30% wage premium over standard construction, alongside 9–11% local wage inflation in Raleigh-Durham (iRecruit). Replacement cost runs 16–20% of an hourly worker's salary and up to 213% for highly skilled roles.
A 30% wage premium is a market telling you it cannot clear at the old price. Wage inflation of 9–11% in one metro is the same message, louder.
The budget behind it is the same capital wave funding Compute and capacity brokerage: Anthropic's compute agreements include a $50B Fluidstack US buildout and $100B+ over ten years with AWS (Sacra); OpenAI's include Stargate at $500B over four years (Sacra). Those are buildings before they are GPUs.
Critically, this is not opex a startup would otherwise hire for. It is capital-project spend with a schedule-driven cost of delay — the strongest possible position for a vendor, because the comparison is not to a salary but to $14.2M a month.
Can you get the supply
Genuinely hard, in the way this atlas rates highest. supply: 4.
Four barriers stack. Licensing: a journeyman electrician holds a credential with a multi-year apprenticeship pipeline that no amount of capital shortens. Geography: the worker must be physically at the site, which rules out the offshore substitution that caps prices in Design, video and content production and Outbound and GTM-as-a-service. Demographics: 41% of the workforce retires by 2031, so the pool shrinks while demand multiplies. Retention: 68% turnover means the supply is not a roster you assemble once, it is a flow you must continuously refill.
The comparison worth making is to quality teleoperation, which scores similarly for a related reason — physical presence, trained hands and equipment defeat the laptop-and-a-marketplace playbook. Supply that cannot be assembled remotely is supply that a well-funded competitor cannot buy overnight.
The second, less organised supply is adjacent and arguably better: power interconnection queues, land and permitted sites. Build (build.inc) raised $8.5M to automate data-centre due diligence and sells site-selection and development workflow (TNW; build.inc). A permitted site with power is about as far from a fungible unit as a supply can get.
What the spread looks like
Conventional today, and the interesting question is whether it stays that way.
The intermediaries currently serving this market are ordinary construction staffing firms taking conventional 30–40% markups. That is the honest baseline, and it is a good number — three to four times what a compute broker can hope for, comparable to the 33–36% take that engineer marketplaces achieve on the one disclosed figure in the sweep (Forward-deployed engineering), and well above the low single digits available on an indexed commodity.
Whether an operator can beat 30–40% turns on one thing: does the buyer pay for speed rather than for hours? A staffing markup is a per-head margin, capped by what a rival staffing firm charges for the same head. A vendor who sells "crew on site in three weeks against a $14.2M monthly clock" is selling schedule insurance, and schedule insurance prices against the loss avoided, not against the wage. Nobody in the record has tried this and published the result.
The 30–40% figure describes incumbent construction staffing, not a specialised data-centre labour broker, because no such broker has published economics. Everything on this page about a better spread is a structural argument from the absence of a price index and the presence of a $14.2M/month cost of delay. It is an argument, not a measurement.
spread: 2, revised down from 3. This page previously carried the same six digits as Robotics teleoperation and physical-world data — 5, 4, 3, 3, 4, 2 — and the two vectors were not earned on comparable evidence. Robotics-data has observed price points on both sides of its trade: buyers pay $50–$200 an hour, operators are paid $25–$50, and a 2,000-demonstration programme costs about $62,400. This page has neither side. Nobody has published what a data-centre developer pays a staffing firm per placed electrician, and nobody has published what any intermediary keeps.
Two things then push the digit down rather than merely flagging it. First, a markup is not a margin. A 30–40% markup on cost is a gross margin of roughly 23–29%, which sits at the 2/3 boundary on the rubric before any discount for evidence. Second, that band describes conventional construction staffing serving conventional construction — a different buyer, a different schedule and a different cost of delay — so it is a comparable borrowed from an adjacent market, not a figure from anyone selling into this one. A borrowed comparable converted from the wrong denominator does not support a mid-table score.
The case for more is still on the page and it is still untested: schedule insurance priced against $14.2M a month rather than against a wage would be a different business with a different number. The moment any operator here discloses a real take, this score should move — in either direction.
Can you hold it
Moderately, and more strongly on the buyer side than is usual in this atlas.
Data-centre construction is a repeat-purchase business with a small number of large developers and hyperscalers running multi-year programmes. A vendor who delivers a crew on schedule for one campus is on the list for the next one, and safety record, prequalification and master service agreements create real procedural switching costs — the sort of friction that Expert networks get from compliance indemnity rather than from anything technological.
The supply leaks in the ordinary trades way: a worker placed on a long project gets hired directly by the contractor. That is Getting cut out, and 68% turnover means the roster is churning regardless.
The genuine risk is One customer is a binary event. A handful of developers and hyperscalers commission most of this capacity, so a broker's customer list is short by construction — the same exposure that made Appen a $4.3B company with 80% of revenue in five clients before Google left — and whose top five were still 74.3% of revenue in FY2025, up from 67.3%, on the way down (Appen Annual Report 2025). hold: 3.
What AI does to it
ai: 4. Models create this demand and cannot perform the work.
The entire shortage is downstream of AI capital expenditure. Every dollar committed to a Stargate or a Fluidstack buildout converts into electricians, pipefitters and controls technicians who do not exist yet. And unlike every function-section vertical in this atlas, the deliverable is physically embodied: nothing a model does installs busbar or terminates a cable.
That immunity is the point. The recurring failure mode across this sweep is that when the unit of work becomes model-tractable, the price collapses 70–95% within eighteen months — legal document review from a $460,000 quote to $36,000 on the same matter (Decover), commodity egocentric video to free. Physical trades are structurally exempt from that mechanism. See What better models do to each layer.
The half-point deducted: models compress the brokerage work — credential verification, scheduling, compliance paperwork, site matching — which is exactly what build.inc is doing to due diligence. The margin on coordination is at risk even when the margin on labour is not.
What would kill it
The capex cycle turns. This vertical has one demand driver and it is the AI buildout. Venture and infrastructure capital is at a record at the top while LP fundraising sits at a seven-year low and first-time fund formation at a ten-year low (PitchBook). A labour broker with a two-year ramp and a one-cycle demand curve can arrive precisely as the cycle ends.
Two others. Speed to first invoice is genuinely bad — prequalification, safety records, insurance, bonding, licensing and union relationships take quarters, not weeks, which is why this scores speed: 2, the lowest in the capacity section. The incumbents are competent — construction staffing firms already hold the relationships and the compliance apparatus, and "we do it with better software" is a claim buyers in this industry have heard before.
Who is already there
| Who | What they do | Money | Evidence quality |
|---|---|---|---|
| Conventional construction staffing firms | Trade placement at 30–40% markups | Not disclosed in this record | Markup band is the only figure |
| build.inc | Automates data-centre due diligence; site selection and development workflow | $8.5M raised | Funding only; no revenue |
| Data-centre security guarding vendors | Guarding at a lower skill tier | Nothing found | No pricing located at all |
Three rows, one funding figure, and no revenue anywhere. That is what an unoccupied vertical looks like, and it is also what an unverifiable one looks like — the two are indistinguishable from outside, which is the whole problem with this page.
An empty table is not evidence that a business works; it is the absence of evidence either way. The atlas gives build to Robotics teleoperation and physical-world data and Adversarial evals and red-team crowds because each has something observed on the operating side — real rates on both sides of the teleop trade in one case, a funnel whose pricing behaviour can be checked from outside in the other. This vertical has a superbly documented buyer and not one fact about any seller, which is the same condition that puts Voice, speech and low-resource language data and Community, campus and events on watch: a plausible business rather than an observed one. The structural argument is unchanged and it is good — no index, no offshore substitution, a $14.2M-a-month clock — but an argument built entirely from one side of a market is a reason to go and get the missing side, not a reason to start. One week of calls to two construction staffing firms and one hyperscaler build programme would settle it, and would move this page further than anything else in the record.
Where the record is thin
Every number on this page describes the buyer's problem — shortage counts, wage premiums, delay costs, turnover. Not one describes an intermediary's business. No data-centre labour broker has published revenue, margin, headcount, placement volume or retention. The 30–40% markup is the general construction-staffing band, not a figure from anyone serving this specific demand.
Named holes, in order of value. Security guarding for data centres is the same shape at a lower skill tier and no pricing was found for it at all — a whole adjacent segment with zero record. Nobody has published what a data-centre developer actually pays a staffing firm per placed electrician, which is the single number that would confirm or destroy the spread argument. And the cost of delay is one source's figure for one project size — $14.2M/month at 60MW — with no independent confirmation and no sense of how it scales.